The research, led by Zaineb M. Alhakeem and Suroor M. Dawood of the College of Oil and Gas Engineering at the University of Basrah in Iraq, together with Qabeela Q. Thabit and Batool Dahham Al-Ali, addresses a threat that has grown quietly alongside the digitization of healthcare. As hospitals connect electroencephalography equipment to networks, cloud archives, and telemedicine platforms, the signals they carry traverse the same vulnerable infrastructure as any other data traffic. An adversary who intercepts or manipulates that traffic does not need physical access to a patient to cause harm. A distorted EEG waveform can mask the signatures of a seizure, fabricate abnormalities that trigger unnecessary treatment, or, in the case of a brain-computer interface, translate into commands the patient never issued. The stakes, in other words, are not merely informational but clinical and even physical.

What makes the new work distinctive is its dual mandate. Most published approaches to medical signal security stop at detection, flagging a stream as suspicious so that a clinician or an automated system can discard it. Discarding, however, is a luxury that real clinical workflows do not always have. If a neurologist is mid-examination, or if a brain-computer interface is the patient’s only channel of communication, a binary alarm that silences the system creates its own kind of failure. The Basrah team therefore designed a framework built on a Long Short-Term Memory neural network, a deep learning architecture specifically engineered to learn patterns in sequential data, that performs detection and correction simultaneously on the same incoming signal.

The choice of architecture is technically significant. Long Short-Term Memory networks, first introduced by Sepp Hochreiter and Jürgen Schmidhuber in 1997, overcome a fundamental weakness of earlier recurrent networks by using gated memory cells that decide what information to retain, update, or forget as a sequence unfolds. EEG is precisely the kind of data this suits: a stream of voltage fluctuations whose diagnostic meaning depends on temporal context, rhythm, and evolution over hundreds of milliseconds. An LSTM can internalize what a healthy EEG trace looks like as it unfolds in time, which gives it two capabilities at once. It can recognize when an incoming segment deviates from the learned statistical structure of genuine brain activity, and it can use its internal model of that structure to reconstruct what the corrupted segment should have contained.

To stress-test the framework, the researchers simulated six distinct categories of cyber-attack, each distorting the signal in a different mathematical way. Fixed-point attacks freeze the signal at a constant value, as though the transmitter had stopped sending real data. Random-point attacks inject unpredictable noise at scattered samples. Sequence attacks replace genuine segments with artificial or reordered data. Scaled attacks multiply the signal’s amplitude, subtly inflating or deflating every value while preserving the overall shape. Ramp attacks impose a steadily increasing or decreasing offset, a drift that can masquerade as a physiological trend. Finally, denial-of-service attacks aim to interrupt transmission entirely, starving the receiver of data. These attack families mirror classification schemes developed in the broader literature on cyber-physical systems, where similar manipulations have crippled power grids and industrial controllers, and their translation to the medical domain is a reminder that biomedical telemetry shares the attack surface of critical infrastructure.

The results, measured across two separate datasets, were striking. On the first dataset, the model’s overall detection accuracy ranged from 80 to 98 percent depending on the attack type and conditions. On the second dataset, accuracy climbed to between 83 and 99 percent. The datasets themselves lend the findings credibility: one is an openly available EEG dataset of repeated measurements published in Scientific Data, and the other comes from the well-known BCI Competition, a long-standing benchmark series in the brain-computer interface community. Testing on established, publicly documented recordings rather than proprietary data means other research groups can verify, reproduce, and extend the results, a standard that security research in medicine has often struggled to meet.

Detection, however, is only half of the story. The team quantified correction quality using the root mean square error, a standard metric that measures the average magnitude of the difference between the reconstructed signal and the original uncontaminated one. Here the spread across attack types was wide and instructive. For sequence attacks, the model achieved its best performance, with an RMSE of approximately 0.8, meaning the reconstructed signal tracked the true brain activity remarkably closely. At the opposite extreme, ramp attacks proved the hardest to undo, producing an RMSE of roughly 33 in the worst case. That asymmetry makes intuitive sense: a ramp attack superimposes a slowly drifting bias that can mimic genuine low-frequency physiological phenomena, forcing the network to disentangle an adversarial trend from real brain dynamics. Fixed, random, and scaled perturbations, by contrast, violate the expected statistical texture of EEG in ways the LSTM’s temporal memory can more readily identify and subtract.

The implications reach well beyond the neurology ward. The Internet of Medical Things, the sprawling ecosystem of connected monitors, infusion pumps, wearable sensors, and implantable devices, has been repeatedly identified in the security literature as an underserved attack surface, and deep recurrent networks have already been proposed for detecting intrusions in smart medical environments. The Basrah study extends that line of defense from the network layer down to the physiological signal itself, treating the waveform, rather than the packet, as the object to be protected. This matters because sophisticated adversaries do not always attack in ways that network intrusion detection systems can see. A manipulated EEG stream can arrive over a perfectly legitimate connection; only an analysis of the signal’s intrinsic structure can reveal that something is wrong. The authors’ framework effectively adds a semantic layer of security, one that judges data not by where it came from but by whether it looks like something a human brain could have produced.

The study also arrives amid a rapidly growing body of work on EEG-specific threats. Researchers have demonstrated data-poisoning attacks on emotion-recognition systems that read affective states from brain signals, backdoor attacks that exploit the spatial and spectral characteristics of EEG to subvert deep neural networks in brain-computer interfaces, and adversarial perturbations that fool motor-imagery classifiers into misreading a patient’s intent. Countermeasures have included adversarial training, alignment-based robustness techniques, and hierarchical detection architectures. What most of these defenses share, however, is a focus on keeping the classifier’s decision correct in the face of attack. The new work reframes the problem at the level of signal integrity, asking whether the raw data itself can be restored, which is a prerequisite for every downstream application, clinical or otherwise.

Significant challenges remain before such a system could protect patients in production. Real-world deployments face streaming latency constraints, channel dropouts, electrode artifacts, and patient-to-patient variability that laboratory datasets only partially capture, and the wide RMSE spread across attack types shows that some manipulations will resist reconstruction more than others. The researchers report that their framework operated robustly under even severe attacks in real-time transmission conditions, and they received no external funding for the work, but clinical integration will demand rigorous validation, regulatory review, and fail-safe behavior when the model’s confidence is low. Still, the direction of travel is clear. As medicine wires the human nervous system into its digital infrastructure, the security of that infrastructure becomes inseparable from patient safety. A deep learning model that can look at a stream of brain voltages, recognize the fingerprint of an attack, and hand the clinician back a clean signal represents a meaningful step toward a healthcare system in which the integrity of the data is defended as vigilantly as the health of the patient it describes.

Subject of Research: Deep learning detection and correction of cyber-attacks on electroencephalogram signals

Article Title: Detection and correction the cyber-attack of electroencephalogram signals using deep learning model

Article References: Alhakeem, Z. M., Dawood, S. M., Thabit, Q. Q., & Al-Ali, B. D. (2026). Detection and correction the cyber-attack of electroencephalogram signals using deep learning model. Neural Computing and Applications, 38(19), Article 806. https://doi.org/10.1007/s00521-026-12542-w

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12542-w

Keywords: EEG, cybersecurity, deep learning, LSTM neural network, brain-computer interface, healthcare, signal correction, denial of service, medical Internet of Things, anomaly detection, Neural Computing and Applications, root mean square error

The research addresses a long-standing division in the field of dialogue systems. Open-domain dialogue, often called chitchat, encompasses free-ranging conversation about movies, weather, travel plans, or personal interests, with no specific goal in mind. Task-oriented dialogue, by contrast, is engineered to help users complete concrete objectives such as booking a hotel, reserving a taxi, or finding an attraction. Historically, these two modes have been built and evaluated separately, and systems that attempt both often struggle to move smoothly between them. Previous efforts have explored fusing the two styles, including work on adding chitchat to enhance task-oriented exchanges and systems that transition from casual talk to tasks, but the question of how a policy should decide when and how to make that leap has remained difficult to solve with hand-crafted rules.

The core conceptual innovation of the new framework is deceptively simple: chitchat is treated as an action, on equal footing with task-oriented actions such as requesting a slot value or confirming a booking. At every turn of a conversation, the dialogue policy must choose from a shared action space that includes both conventional task actions and the option to engage in casual conversation. When the policy selects the chitchat action, the actual response is generated by prompting a large language model with the ongoing dialogue context, allowing the system to produce natural, fluent small talk tailored to the moment. This design means the decision of whether to chat, probe, or push toward a task is learned from experience rather than dictated by rigid scripts.

Training such a policy requires an environment in which it can practice, and the authors supply one in the form of an LLM-based user simulator. Rather than relying on traditional agenda-based user simulators, which follow predetermined goals and mechanical rules, the simulator is a large language model prompted with a randomly generated user goal converted from structured JSON into natural language, along with task instructions and a set of few-shot examples. Crucially, the simulated user is instructed to hint at its goal through chitchat without directly mentioning the task itself. This creates a genuinely challenging training ground: the system must learn to read between the lines of casual conversation, inferring latent intentions from indirect cues, much as a perceptive human salesperson might notice a customer mentioning an upcoming anniversary and gently steer toward gift suggestions.

To make reinforcement learning feasible in this setting, the researchers introduce a component called the Spawn Evaluator. This module tracks the progress of each dialogue and computes reward signals that guide the policy’s learning. The rewards reflect whether the system successfully manages the chitchat portions of the exchange and whether any tasks it spawns from the conversation turn out to be genuinely relevant to the user’s underlying interests. The policy itself is trained using proximal policy optimization, a widely used reinforcement learning algorithm known for its stability, and the authors build on established practices for training deep neural networks and for ensuring that deep reinforcement learning experiments are conducted rigorously and reproducibly.

The evaluation compares the proposed system against task-oriented dialogue baselines using established multi-domain benchmarks, including datasets in the spirit of MultiWOZ and the schema-guided dialogue corpus, covering domains such as attractions, hotels, trains, taxis, and restaurants. The results show that the system trained under the chitchat-based framework achieves performance comparable to the task-oriented baselines, despite facing a considerably harder problem. Where a conventional baseline receives the user’s request directly, the proposed system must first detect the task latent in casual conversation, decide whether and when to spawn it, and then complete it efficiently. The appendix of the paper includes transcript examples illustrating both successes and failures: in one evaluation dialogue, the system correctly infers a destination hinted at in the user’s small talk and proactively initiates a train-booking task; in another, it mistakenly spawns an attraction-domain task from an incidental mention, which the simulated user then rejects.

These failure cases are as instructive as the successes. They reveal the fundamental tension at the heart of proactive task generation: acting on every passing mention risks annoying users with unwanted suggestions, while acting too cautiously means missing genuine opportunities. The reward structure mediated by the Spawn Evaluator is designed to calibrate this balance, teaching the policy to distinguish between incidental chitchat and meaningful signals of user intent. The authors frame this capability in terms of potential business value, noting that systems able to spawn tasks from conversation could facilitate product sales or restaurant bookings that users never explicitly requested, converting otherwise idle chat into productive outcomes.

The technical architecture reflects a broader trend in which large language models and reinforcement learning are combined rather than competing. The LLM serves two distinct roles in the framework: as the generative engine producing chitchat responses when the policy selects that action, and as the simulated user providing a realistic, linguistically rich training environment. This separation allows the policy network itself to remain a compact, learnable decision-making module, in the tradition of partially observable Markov decision process dialogue management, while delegating the open-ended language generation to the foundation model. The work builds on the authors’ earlier research, including a review of dialogue system trends and a prior study on prompting large language models for user simulation in task-oriented systems, and it connects to toolkits such as ConvLab-3 that standardize dialogue system experimentation.

The implications extend beyond commercial applications. A unified framework in which conversation and task completion are learned jointly points toward assistants that feel less like vending machines and more like capable collaborators, ones that understand context, pick up on hints, and offer help at the right moment. At the same time, the study underscores open challenges: evaluation of dialogue systems remains contentious, with unsupervised metrics known to be unreliable, and the hit rate of spawned tasks, a measure borrowed from marketing of how many attempts convert into accepted outcomes, will need careful measurement in real deployments with real users. The research was supported by the Saudi Data and AI Authority and King Fahd University of Petroleum and Minerals through their Joint Research Center for Artificial Intelligence, and the dataset analyzed in the study is publicly available with the published article, offering other researchers a concrete starting point for building assistants that can genuinely chat their way to getting things done.

Subject of Research: Reinforcement learning-based dialogue policy that spawns task-oriented dialogue from open-domain chitchat

Article Title: Spawning tasks from chitchatting – reinforcement learning-based policy

Article References: Algherairy, A., & Ahmed, M. (2026). Spawning tasks from chitchatting – reinforcement learning-based policy. Neural Computing and Applications, 38(19), Article 803. https://doi.org/10.1007/s00521-026-12533-x

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12533-x

Keywords: dialogue systems, reinforcement learning, large language models, task-oriented dialogue, chitchat, user simulator, dialogue policy, proximal policy optimization, Spawn Evaluator, open-domain dialogue, proactive task generation, Neural Computing and Applications

The core insight behind the new model is that class imbalance in graphs is not merely a counting problem. In standard tabular data, an under-represented class simply has fewer examples. In a graph, the problem is compounded by topology: minority-class nodes frequently sit in sparse or noisy neighbourhoods dominated by majority-class neighbours, so the message-passing machinery of a graph neural network actively drowns out their signal. The authors call this structural marginality, and they formalise it with a difficulty score for every node. The score blends two ingredients: how isolated a node is within its own class neighbourhood, and how distinctive its features are, estimated by a simple classifier trained on a balanced subset. Nodes are then ordered from easiest to hardest, creating a curriculum the model can follow.

That curriculum unfolds in three stages with names drawn from pedagogy: Engage, Enact and Embed. In the Engage stage, the model sees only structurally simple nodes, those with low difficulty scores, and applies uniform attention across every neighbourhood. This deliberate restriction acts as a form of attention regularisation, preventing the network from prematurely latching onto majority-class patterns before it has built robust, class-agnostic foundations. In the Enact stage, the training set expands to moderately difficult nodes and the uniform attention gives way to learnable attention, allowing the model to begin separating class-specific signals. Crucially, this stage also enforces orthogonality between two parallel feature streams, a graph convolutional pathway that captures structural regularity and a graph attention pathway that captures semantic relationships. Finally, the Embed stage admits the entire graph, including the hardest minority nodes, and switches to minority-boosted attention that explicitly upweights neighbours belonging to under-represented classes.

The dual-pathway design deserves particular attention, because it addresses a known failure mode of graph neural networks on heterophilous graphs, where connected nodes often carry different labels. The Wikipedia-derived Chameleon dataset, with a heterophily ratio of 0.83, is a case in point: a purely structural aggregator struggles because a node’s neighbours are poor guides to its class. By keeping structural and semantic information in separate streams and combining them only after controlled integration, CL3AN-GNN can lean on whichever pathway is informative. The authors back this intuition with theory, proving that on heterophilous graphs the disentangled representation achieves a lower generalisation error than an entangled one, with an improvement proportional to the graph’s heterophily and inversely proportional to the imbalance ratio.

The theoretical contribution extends further. The team derives a curriculum convergence guarantee that relaxes the restrictive global smoothness assumptions of earlier work, requiring only stage-specific conditions on the feasible node subsets. They also show that the curriculum reduces the effective Rademacher complexity of the hypothesis class, which translates empirically into a train-test gap of just 0.032, compared with 0.058 to 0.067 for competing methods. Perhaps most striking for practitioners is the memory result: by factorising attention projections into low-rank approximations, the model cuts memory requirements from quadratic in the number of attention heads and hidden dimensions to linear, yielding a reported tenfold to hundredfold reduction. On the large-scale OGBN-Arxiv benchmark, with 169,343 nodes and more than a million edges, this shrinks GPU memory from roughly 2.1 gigabytes to 0.2 gigabytes, opening the door to inference on edge devices.

The empirical evidence is broad. Across eight benchmarks spanning citation networks (Cora, Citeseer, PubMed), co-purchase graphs (Amazon Photo and Computers), a collaboration network (Coauthor CS), the heterophilous Chameleon graph and OGBN-Arxiv, CL3AN-GNN outperformed thirteen state-of-the-art baselines retrained under identical conditions. Accuracy gains ranged from 1.9 percent on OGBN-Arxiv to 15.2 percent on Citeseer, and the advantages were statistically significant under Wilcoxon signed-rank tests on nearly every dataset. A per-class analysis on Cora and CiteSeer under severe imbalance showed the model reaching F1 scores of up to roughly 73 percent on classes comprising less than five percent of the data, absolute gains of up to eleven points over the strongest baselines, while still holding majority-class performance near 96 to 99 percent.

Robustness under escalating imbalance is where the curriculum design shows its teeth. When the researchers swept the imbalance ratio on Amazon Computers from 10 to 150, CL3AN-GNN’s accuracy degraded gracefully from 0.916 to 0.678, a shallower decline than either NodeImport-GCN or GATE-GAT, and regression analysis confirmed its degradation slope was significantly flatter. Convergence followed the predicted logarithmic scaling with imbalance: the model needed between 28 and 68 epochs to reach a target accuracy as imbalance grew, while a prior curriculum method required 41 to 210 epochs, savings that climbed to 68 percent at the most extreme ratios. Gradient stability analysis on OGBN-Arxiv reinforced the picture, with gradient norms and variance shrinking systematically across the three stages, exactly as the convergence theory predicts.

Interpretability, often an afterthought in this literature, is built into the framework. Because the curriculum stages impose distinct attention regimes, the researchers can verify that attention is uniform in Engage, bounded in variance in Enact, and measurably biased toward minority neighbours in Embed, with a learned boost parameter of about 0.08 emerging without manual tuning. t-SNE visualisations on Cora show CL3AN-GNN producing the most compact and well-separated clusters of the eight methods compared, achieving the lowest intra-class dispersion and the highest silhouette score, 0.42 versus 0.11 for a vanilla graph neural network. The authors argue this transparency matters in regulated domains such as finance and healthcare, where practitioners need to know whether a misclassification stems from a node’s awkward position in the network or from genuinely ambiguous features.

The authors are candid about limitations. The theoretical guarantees rest on relaxed but still nontrivial assumptions that may not hold in every deployment; the framework assumes a static initial topology, so dramatic structural changes require retraining; and extremely heterophilous graphs remain hard, since the minority boost cannot fully compensate when neighbours rarely share labels. Future work, they write, will target dynamic graphs, automated difficulty metrics and federated settings. Still, the broader message is compelling: rather than treating imbalance as a sampling nuisance to be patched with oversampling or reweighting, CL3AN-GNN embeds the curriculum directly into the architecture, letting attention, loss weighting and feature disentanglement evolve together. For any application where the rare cases carry the highest stakes, from fraud detection to rare-disease diagnosis, that integrated approach may prove to be the lesson worth learning.

Subject of Research: Curriculum-driven graph neural networks for class-imbalanced node classification

Article Title: CL3AN-GNN: curriculum-driven for incremental feature disentanglement with multi-phase attention for class-imbalanced node classification

Article References: Fofanah, A. J., Chen, D., Wen, L., Zhang, S., Sesay, M., & Dumbuya, I. (2026). CL3AN-GNN: curriculum-driven for incremental feature disentanglement with multi-phase attention for class-imbalanced node classification. Neural Computing and Applications, 38(19), Article 780. https://doi.org/10.1007/s00521-026-12579-x

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12579-x

Keywords: graph neural networks, class imbalance, curriculum learning, node classification, feature disentanglement, attention mechanisms, machine learning, heterophilous graphs, minority classes, deep learning, fraud detection, Neural Computing and Applications

The research team built a patient-specific prediction framework and evaluated it on the OhioT1DM dataset, a widely used benchmark collected from people with Type 1 Diabetes wearing continuous glucose monitors. What sets the study apart is its commitment to minimal preprocessing. Many machine learning pipelines in diabetes research rely on elaborate cleaning, imputation and feature engineering stages that can be fragile in real-world deployment, where sensor dropouts and irregular meal logging are the norm. By keeping data preparation light, the Egyptian team aimed to test how well modern machine learning models perform when fed something much closer to the raw stream of clinical measurements that a real monitoring system would produce.

The experimental design was systematic. The researchers compared several machine learning models using standard regression metrics: Mean Absolute Error, which captures the average size of prediction mistakes; Root Mean Square Error, which penalizes large errors more heavily; and the coefficient of determination, or R², which measures how much of the variability in future glucose values the model actually explains. Crucially, they did not stop at statistical scores. They also applied Clarke Error Grid analysis, a clinical assessment tool developed in 1987 that judges predictions the way clinicians would, by asking whether a forecast error would lead to a harmless decision or a dangerous one. A model can look numerically decent while still making the kind of mistake that would trigger a wrong insulin dose, and the Clarke Error Grid is designed to catch exactly that.

Feature selection was explored through three combinations drawn from the literature. The simplest used blood glucose readings alone, testing how far a model can get on the glucose time series itself. The second added insulin-on-board, an estimate of how much previously injected insulin is still actively working in the body, together with meal information, since carbohydrates are the dominant driver of glucose excursions. The third configuration added physical activity data on top of everything else, acknowledging that exercise can send glucose plummeting or, in some cases, spiking, in ways that neither insulin nor food history fully explains. Each configuration was tested with a sliding input window of five historical time steps sampled at five-minute intervals, meaning the models looked back twenty-five minutes into each patient’s recent metabolic past.

Two prediction horizons were evaluated: thirty minutes and sixty minutes ahead. This choice reflects the practical realities of diabetes management. A thirty-minute forecast gives enough lead time for a pre-emptive snack or a delayed bolus, while a sixty-minute forecast is considerably harder because the effects of meals, insulin and activity compound over time, but it offers a much larger safety margin. The gap in difficulty between the two horizons is one of the central challenges in the field, and any framework that maintains accuracy at the longer horizon is doing genuinely useful work.

The centerpiece of the study is the Tri-Ensemble model, constructed by combining the predictions of the three strongest benchmarked models using simple averaging. Ensembling is a well-established strategy in machine learning: when individually strong models make somewhat different errors, averaging their outputs can cancel out idiosyncratic mistakes and produce a more robust overall prediction. The authors describe this aggregation choice as robustness-oriented, and the results supported that framing. The Tri-Ensemble demonstrated competitive predictive performance across multiple feature configurations and both prediction horizons, with statistically significant improvements observed in several of the experimental settings. In other words, the gains were not just artifacts of chance but held up under statistical scrutiny.

Beyond raw accuracy, the study contributes an interpretability layer. The researchers conducted an exploratory explainable AI analysis using SHAP, or SHapley Additive exPlanations, a technique borrowed from cooperative game theory that assigns each input feature a contribution value for every individual prediction. Rather than treating the model as a black box that spits out a glucose number, SHAP lets researchers open the box and see which variables pushed each forecast up or down, and by how much. The team applied this analysis over the complete available feature space, going beyond the three curated feature combinations used in the benchmarking experiments.

The SHAP analysis was performed patient by patient, and this patient-wise perspective is one of the study’s most interesting aspects. Type 1 Diabetes is famously heterogeneous: two people with the same diagnosis can have very different insulin sensitivities, eating patterns, activity levels and glucose dynamics. A feature that dominates predictions for one patient, such as insulin-on-board, may play a secondary role for another whose glucose is driven more by activity or meal timing. By examining feature importance patterns for each individual, the authors aimed to identify potentially informative variables that could guide feature engineering in future blood glucose prediction studies, effectively using the explainability tool as a discovery instrument rather than merely a compliance checkbox.

The significance of this work lies partly in its restraint. Rather than proposing an ever-larger deep learning architecture trained on heavily curated data, the study asks a more deployable question: how well can we predict glucose with modest models, light preprocessing and honest clinical evaluation? The answer, supported by the OhioT1DM benchmark and Clarke Error Grid analysis, is encouraging. A simple averaged ensemble of the strongest learners, given just twenty-five minutes of history and a handful of clinically meaningful signals, can deliver forecasts that are both statistically and clinically competitive. That combination of simplicity and rigor matters for anyone hoping to move glucose prediction out of the lab and into the pumps, pens and smartphone apps that patients actually use.

The road ahead, as the authors frame it, involves using the SHAP-derived feature relevance patterns to design smarter, personalized feature sets, potentially letting each patient’s model emphasize the signals that matter most for their own physiology. As continuous glucose monitoring spreads globally and the International Diabetes Federation projects continued growth in diabetes prevalence, tools that forecast blood glucose accurately, transparently and with minimal data fuss could become a quiet but transformative part of daily care. This study is a reminder that in medical machine learning, sometimes the most valuable innovation is not a bigger model, but a clearer view of what the model is actually seeing.

Subject of Research: Machine learning prediction of blood glucose in Type 1 Diabetes using the OhioT1DM dataset with SHAP-based explainability

Article Title: Blood glucose prediction in Type 1 Diabetes under minimal preprocessing using machine learning and exploratory SHAP analysis

Article References: Abdelaziz, D. M., Refaat, A. M., & Muhammad, S. T. (2026). Blood glucose prediction in Type 1 Diabetes under minimal preprocessing using machine learning and exploratory SHAP analysis. Neural Computing and Applications, 38(17), Article 706. https://doi.org/10.1007/s00521-026-12330-6

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12330-6

Keywords: Type 1 Diabetes, blood glucose prediction, machine learning, OhioT1DM dataset, Tri-Ensemble, SHAP, explainable AI, Clarke Error Grid, insulin-on-board, continuous glucose monitoring, patient-specific modeling, Neural Computing and Applications

The researchers’ starting point is the ant colony system, a family of algorithms inspired by the way real ants find short paths to food. Individual ants deposit pheromone along the trails they walk, and subsequent ants preferentially follow stronger trails, reinforcing good routes while weaker ones evaporate. Translated into computation, artificial ants construct delivery routes step by step, guided by two signals: pheromone trails that encode collective experience about which customer-to-customer moves have historically produced good routes, and heuristic information that estimates the immediate desirability of each possible move. Over many iterations, the colony converges toward high-quality solutions without ever enumerating the full search space. Since Marco Dorigo introduced ant colony optimization in the early 1990s, the approach has been applied to everything from network routing to scheduling, and it remains a workhorse for routing problems where exact methods run out of steam.

What distinguishes the new algorithm is the way it deploys not one but four heuristic techniques inside the colony’s decision-making machinery. In a conventional ant colony system, each ant typically relies on a single heuristic rule when choosing the next customer to visit. That rigidity can be a liability, because different heuristics excel in different regions of the search: one might be better at deciding which customer to add to a route, another at deciding where to insert a customer into an existing route, and still others at repairing or restructuring routes that have become inefficient. Rather than betting on a single rule, the authors let the algorithm learn which heuristic serves it best at any given moment, drawing on the complementary strengths of all four.

The mechanism that orchestrates this cooperation is a ranking-based selection procedure built on the classic roulette wheel. Each of the four heuristics is assigned a weight reflecting how well it has performed recently, and when the algorithm needs to make a decision, it spins the metaphorical wheel, giving better-ranked heuristics proportionally larger slices. Crucially, the weights are not fixed. As the search progresses, heuristics that consistently contribute to improved solutions gain a larger share of future decisions, while those that fail to help are gradually demoted. This adaptive allocation means the algorithm effectively conducts a continuous, built-in experiment, reallocating its computational effort toward whichever strategy is currently paying off. The result, according to the authors, is the generation of promising solutions that no single heuristic could reliably produce alone.

A second design element addresses one of the most persistent failure modes of swarm intelligence: stagnation. Because pheromone reinforcement favors routes that have already been found good, a colony can settle into a rut, repeatedly exploring the same small set of routes while better alternatives elsewhere in the search space go unnoticed. The new algorithm counters this by incorporating a no-improvement counter into its solution selection process. When the counter indicates that the colony has gone a stretch of iterations without finding anything better, the algorithm adjusts its selection behavior to push the search toward less-explored territory, restoring diversity among the candidate solutions. This balance between intensification, which exploits known good regions, and diversification, which explores new ones, is the central tension of all metaheuristic design, and the counter gives the algorithm an explicit, self-monitoring handle on it.

The third and perhaps most inventive contribution is a new local improvement operator the authors call 2-opt-special, inspired by the venerable 2-opt move that has been a staple of routing heuristics since the 1960s. In its classic form, 2-opt takes a route, removes two edges, and reconnects the remaining segments in the alternative way, undoing any crossings and shortening the tour. The special variant introduced here adapts and extends this idea to bolster both the efficiency and the effectiveness of the ant colony framework, giving the colony a targeted tool for refining the routes its ants construct. Local search operators of this kind matter enormously in practice: constructive methods such as ant colonies are good at assembling plausible routes, but the difference between a plausible route and an excellent one often lies in small, surgical repairs that only a well-designed neighborhood move can deliver.

To find out whether all this machinery actually works, the team ran extensive computational experiments on well-known benchmark instances containing 100 and 400 customers, the standard proving grounds for VRPTW algorithms. They compared their method against other ant colony-based algorithms and against established solutions from the literature. The results, the authors report, affirm that the proposed algorithm is effective and competitive in addressing VRPTW challenges, holding its own against prior approaches across the benchmark suite. The comparisons were supported by the kind of statistical methodology that has become standard practice for evaluating evolutionary and swarm algorithms, following nonparametric testing procedures designed to compare stochastic optimizers fairly rather than relying on single lucky runs.

The significance of the work extends beyond the leaderboard. Vehicle routing with time windows sits at the heart of logistics, e-commerce, and emergency transportation, and even percentage-point improvements in route quality translate into real savings in fuel, fleet size, and emissions. The paper’s emphasis on combining multiple heuristics under an adaptive selection scheme reflects a broader trend in metaheuristics research: rather than designing ever more exotic single algorithms, researchers increasingly build frameworks that can assemble and tune proven components on the fly. The authors’ own research trajectory points in the same direction, including earlier work on hybrid adaptive large neighborhood search for capacitated routing problems and on metaheuristic techniques applied to engineering design problems.

For the field, the study also offers a reminder of how much headroom remains in a technique that is more than three decades old. Ant colony optimization was conceived as a simple model of distributed learning, yet its modern descendants now incorporate ranking mechanisms, adaptive strategy portfolios, stagnation detection, and specialized local search, layering decades of algorithmic insight onto the original biological metaphor. The version of the algorithm described here, published in volume 38 of Neural Computing and Applications as article number 705, was received in October 2024 and accepted in July 2026, and the authors declare no conflicts of interest and no specific funding for the work.

As delivery networks grow denser and customer expectations for precise delivery windows tighten, the computational problems underlying logistics will only intensify. Approaches like the one developed by Kuyu, Fajardo-Calderin, and Onieva suggest that the path forward lies not in any single clever rule but in flexible systems that know when to switch tactics, when to double down on what works, and when to break out of a rut. Virtual ants, it turns out, still have plenty to teach us about finding the best way home.

Subject of Research: A novel ant colony system using multiple heuristic strategies for solving the vehicle routing problem with time windows

Article Title: A novel ant colony system for vehicle routing problem with time windows: leveraging multiple heuristic strategies for improved solutions

Article References: Kuyu, Y. Ç., Fajardo-Calderin, J., & Onieva, E. (2026). A novel ant colony system for vehicle routing problem with time windows: leveraging multiple heuristic strategies for improved solutions. Neural Computing and Applications, 38(17), Article 705. https://doi.org/10.1007/s00521-026-12418-z

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12418-z

Keywords: ant colony optimization, vehicle routing problem, time windows, heuristics, metaheuristics, swarm intelligence, combinatorial optimization, logistics, local search, 2-opt, benchmark instances, Neural Computing and Applications

The problem the researchers tackle, known as examination timetabling, belongs to a family of combinatorial optimization challenges that are computationally intractable at scale. The goal is to assign each exam to a timeslot without creating conflicts, meaning no student should be scheduled for two exams in the same period. Beyond hard constraints, there are soft constraints that translate into penalties: spreading exams out so students are not crammed with back-to-back tests, for example. Because the number of possible timetables grows explosively with the number of exams and students, finding the best possible schedule by brute force is out of the question for real institutions. Instead, researchers rely on metaheuristics, intelligent search strategies that explore the vast space of possible schedules in search of very good, if not provably optimal, solutions.

The core engine of the new approach is the Rat Optimization Algorithm, or ROA, a relatively recent nature-inspired method introduced in 2024 that mimics the swarming and foraging dynamics of rats. In its original form, ROA is designed for continuous optimization, where solutions are points in a smooth numerical landscape. Examination timetabling, however, is a discrete problem: an exam either sits in timeslot three or it does not, and there is no meaningful in-between. The researchers therefore reformulated ROA’s operators so that the algorithm’s search moves operate directly on discrete timeslot assignments, a necessary adaptation that preserves the spirit of the rat-inspired search while making it compatible with the structure of the problem.

Two further ingredients round out the hybrid. The first is the Lévy Flight, a mathematical description of a movement pattern observed across nature, from albatross foraging paths to human hunter-gatherer travel, in which mostly short, local steps are punctuated by occasional long jumps. In optimization terms, Lévy Flights allow an algorithm to escape the gravitational pull of a local optimum by occasionally making dramatic, far-reaching changes to a solution. The second ingredient is simulated annealing, a technique dating back to 1983 that borrows its logic from the cooling of molten metal: it accepts most improvements outright, but also tolerates some worsening moves with a probability that gradually decreases as the search matures, giving the algorithm a controlled way to explore early on and to refine later.

Crucially, the authors did not simply bolt these components together. They describe three main adaptations that coordinate the pieces. ROA’s operators were rewritten for discrete timeslots, as noted. The Lévy Flight mechanism was designed to apply selective perturbations with feasibility checking, so that a long jump does not wreck the hard constraint that no student has conflicting exams; if a perturbation breaks feasibility, it is checked and handled rather than blindly accepted. And simulated annealing’s acceptance criterion was synchronized with ROA’s search phases, so that the willingness to accept worse solutions ebbs and flows in harmony with the underlying swarm dynamics rather than running on an independent schedule. On top of this coordination layer, the team incorporated Kempe chain operators, a neighborhood structure borrowed from graph coloring in which an entire chain of interdependent exam assignments is moved at once, providing a powerful local refinement mechanism that can untangle stubborn scheduling knots.

To find out whether all this engineering actually pays off, the researchers ran a rigorous evaluation on twelve benchmark instances from the well-known Toronto dataset, a standard proving ground for examination timetabling research hosted via the University of Nottingham’s public repository. The evaluation had two prongs: ablation studies, which systematically strip away components to measure each one’s individual contribution, and comparisons against eleven state-of-the-art methods from the literature. The ablation results are striking. The full hybrid configuration improved on a basic ROA implementation by 38.4 percent, with the Lévy Flight component contributing 13.0 percent of that gain and simulated annealing contributing 10.5 percent. These numbers suggest that the hybridization is not decorative; each added mechanism earns its place in the pipeline.

In the head-to-head comparisons, the hybrid algorithm achieved a Friedman rank of 4.59, placing it third overall among the methods tested. Statistical rigor was a priority: the authors applied the Wilcoxon, Friedman, Nemenyi, and Holm tests to the results, and this analysis confirmed that there was no statistically significant difference between the top two methods in the comparison. In other words, while the new hybrid did not claim the crown, it sits firmly within the leading pack, and its performance is statistically indistinguishable from the best performers. For a newly adapted algorithm competing against mature, heavily tuned methods, that is a notable result.

Perhaps just as important as raw solution quality is consistency, and here the algorithm shines. Across all twelve instances, the method demonstrated strong repeatability, with a coefficient of variation as low as 0.01 percent between runs. Stochastic algorithms can sometimes produce wildly different schedules from one run to the next, which is a practical headache for administrators who need to trust that the tool will deliver a comparable result every time. A coefficient of variation in the hundredths of a percent range indicates that the hybrid’s search is remarkably stable, likely a benefit of the coordinated acceptance mechanism that keeps the exploration-exploitation balance steady throughout the run.

The study arrives amid a long and active research tradition. Examination timetabling has been studied algorithmically since at least the mid-1990s, when Carter, Laporte, and Lee laid out foundational strategies, and simulated annealing itself was applied to the problem as early as 1998 by Thompson and Dowsland. Since then, the field has seen tabu search, ant algorithms, artificial bee colonies, genetic algorithms, harmony search, intelligent water drops, firefly algorithms, and cellular memetic approaches, among many others. Surveys published in 2009 and again in 2024 chart the steady evolution of solution methodologies, and the persistence of the problem in the literature reflects its real-world stakes: poorly balanced exam schedules translate directly into student stress and institutional friction.

What the Ajman University team’s work illustrates is a broader lesson in modern metaheuristic design: single algorithms rarely dominate, and the most competitive systems are often carefully orchestrated hybrids in which each component compensates for the weaknesses of the others. The rat-inspired swarm provides the population-level exploration, the Lévy Flight injects the occasional bold leap out of stagnation, simulated annealing governs the tolerance for temporary setbacks, and Kempe chains perform the surgical fine-tuning. The result, published in Neural Computing and Applications as volume 38, article 735, is a system that lifts a young optimization method into contention with the state of the art, while offering a template for how discrete scheduling problems can benefit from disciplined hybridization rather than novelty for its own sake.

Subject of Research: A hybrid metaheuristic combining the rat optimization algorithm, Lévy Flights, and simulated annealing for solving the examination timetabling problem.

Article Title: Hybridizing rat optimization algorithm with Lévy Flights and simulated annealing for examination timetabling

Article References: Setteh, M. A., Al-Betar, M. A., & Nachouki, M. (2026). Hybridizing rat optimization algorithm with Lévy Flights and simulated annealing for examination timetabling. Neural Computing and Applications, 38(17), Article 735. https://doi.org/10.1007/s00521-026-12466-5

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12466-5

Keywords: examination timetabling, rat optimization algorithm, Lévy Flight, simulated annealing, hybrid metaheuristic, combinatorial optimization, Kempe chains, Toronto benchmarks, scheduling, swarm intelligence, Neural Computing and Applications, discrete optimization

The work, authored by Sayak Mondal of the School of Computer Science and Engineering at Vellore Institute of Technology in India, begins from an observation that will resonate with anyone who has worked with fine-grained emotion datasets. Real-world semantic structures, particularly those involving entangled emotional labels, exhibit anisotropic correlations and non-uniform separability — technical terms for the fact that meaning is not spread evenly in every direction. A sentence expressing both grief and gratitude, for example, does not sit neatly inside a single emotional category; it occupies a region of semantic space where multiple labels overlap and blur. Flat Euclidean geometry, with its uniform distances in all directions, struggles to capture such tangled structure. Mondal’s answer is to abandon the flat plane and work instead on a curved, deformable hypersphere governed by the mathematics of Riemannian geometry.

The framework unfolds in three stages, each with a distinct mathematical role. First, sentence embeddings produced by standard encoders are mapped into a learned hyperspace, a high-dimensional representation whose geometry is not fixed in advance but shaped during training. Second, the model performs what the paper describes as a Mahalanobis-style metric deformation. The Mahalanobis distance, a classical tool from statistics, generalizes ordinary Euclidean distance by allowing different directions in space to be weighted differently — effectively letting the model stretch the space along some axes and compress it along others. In HyperSpectrum Geometry, this deformation is controlled by a learnable global Riemannian metric tensor, meaning the curvature of the semantic space itself becomes a trainable parameter rather than a fixed property of the architecture.

The third stage carries out angular decision partitioning on a normalized, deformed hypersphere. After the metric deformation, embeddings are projected onto a hyperspherical manifold — the curved surface of a sphere embedded in high dimensions — where classification decisions are made based on angles rather than raw distances. This angular approach draws on ideas popularized in face recognition research, notably additive angular margin losses such as ArcFace, which showed that forcing class representations to occupy distinct angular regions of a sphere can dramatically sharpen decision boundaries. By combining angular partitioning with a deformable metric, HyperSpectrum Geometry effectively gives the model two independent dials: it can reshape the space to reflect anisotropic semantic correlations, and it can carve the resulting curved surface into angular sectors corresponding to different classes.

To test whether this geometric machinery actually matters, the study evaluated the framework on five structurally distinct datasets: GoEmotions, a collection of Reddit comments annotated with fine-grained emotion labels; AG News and DBpedia Classification, both of which organize text into well-separated topical categories; TweetEval Emotion, a benchmark for classifying emotional content in tweets; and Amazon Reviews Multi, a multilingual sentiment corpus. These datasets were chosen deliberately, because they span a spectrum of structural complexity. Topic and ontology datasets such as AG News and DBpedia feature classes that are largely disjoint — a story about sports is rarely also a story about science — whereas fine-grained emotion datasets are notorious for multi-label entanglement, where a single text can plausibly carry several overlapping emotional signals at once.

The empirical results reveal a striking pattern that the paper terms a deformation–separability relationship. On the fine-grained, multi-label emotional datasets, the model learned substantially higher global metric deformation, indicating that it needed to bend the semantic space aggressively to separate entangled categories. On the structurally separable topic and ontology datasets, the opposite occurred: classification relied more strongly on angular partitioning, with comparatively little metric deformation required. In other words, the geometry the model chose to learn was not arbitrary — it tracked the intrinsic difficulty and structure of the task. When classes overlap and blur, the space must be curved and stretched to pull them apart; when classes are naturally distinct, simple angular separation on a gently deformed sphere suffices.

This pattern carries implications that extend beyond benchmark scores. It suggests that the amount of metric deformation a trained model exhibits could serve as a diagnostic signal — a measurable fingerprint of how entangled a dataset’s label structure really is. The study also reports several interpretable geometric quantities computed over multiple random seeds, including the global metric deformation, the condition number of the learned metric tensor, and intra-class versus inter-class angular similarity. The condition number, a concept from linear algebra, quantifies how anisotropic the learned metric is: a value near one indicates a nearly uniform space, while larger values reveal strong directional stretching. Because these diagnostics remained stable across seeds, they offer a reproducible lens on representation learning that most black-box classifiers cannot provide.

Stability and efficiency were recurring themes in the evaluation. The author emphasizes that the model delivers lightweight, consistent performance across repeated runs with different random seeds, a nontrivial achievement given that learned metric tensors can be notoriously sensitive to initialization. The framework’s interpretability is presented not as an afterthought but as a core contribution: rather than treating the geometry as an opaque intermediate layer, the paper argues that learnable metric deformation constitutes a useful and underexplored component of interpretable semantic representation learning. In an era when concerns about the opacity of large language models dominate public discourse, approaches that make their internal geometry legible — and that tie that geometry to observable properties of the data — occupy valuable middle ground between pure performance engineering and post-hoc explanation.

The study also situates itself within a broader movement known as geometric deep learning, which has spent the past decade arguing that Euclidean spaces are only one member of a much larger family of possible representational geometries. Hyperbolic embeddings, for instance, have been used to capture hierarchical structures such as taxonomies, where tree-like data can be embedded far more compactly in negatively curved space than in flat space. Graph neural networks have likewise been adapted to non-Euclidean domains. HyperSpectrum Geometry contributes a complementary perspective: instead of committing to a single fixed curvature, it learns the curvature adaptively, letting the data dictate how much deformation is needed. This curvature-adaptive stance may prove especially relevant as language models are increasingly deployed on tasks — emotion recognition, mental-health screening, nuanced content moderation — where label boundaries are genuinely fuzzy rather than artificially clean.

Caveats remain, as with any single-author empirical study. The paper reports that implementation details, training scripts, random seeds, and hyperparameter settings will be made available as supplementary material or through a public repository on request, and independent replication will be needed to establish how broadly the deformation–separability pattern generalizes to other architectures, languages, and domains. The framework was also evaluated with specific encoder backbones, and the interaction between learned metric deformation and the geometry inherited from large pretrained transformers remains an open question. Nevertheless, the central insight stands on its own merits: the geometry of a model’s semantic space is not a fixed design choice but a learnable quantity, and the curvature a model discovers can tell us something honest about the structure of the problem it faces. As text classification systems are asked to navigate ever more nuanced shades of human meaning, frameworks like HyperSpectrum Geometry suggest that the future of machine understanding may depend less on bigger models than on giving them the freedom to bend the spaces in which they think.

Subject of Research: Curvature-adaptive Riemannian hyperspherical representation learning for neural text classification

Article Title: HyperSpectrum geometry: a Riemannian hyperspherical framework for curvature-adaptive text classification

Article References: Mondal, S. (2026). HyperSpectrum geometry: a Riemannian hyperspherical framework for curvature-adaptive text classification. Neural Computing and Applications, 38(17), Article 710. https://doi.org/10.1007/s00521-026-12332-4

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12332-4

Keywords: Riemannian metric learning, hyperspherical classification, geometric deep learning, Mahalanobis distance, curvature-aware representation learning, text classification, GoEmotions, semantic embeddings, angular margin loss, interpretable machine learning, natural language processing, Neural Computing and Applications

The core insight of the paper is architectural. Rather than treating forecasting and optimization as two sequential stages that are calibrated once and then left alone, the authors construct a closed-loop system in which the two components continuously inform and correct each other. At the foundation sits a unified representation and pre-processing layer that normalizes multi-source sensor data into a consistent form. From there, a time series prediction model projects forward the values of crucial state variables, and those predicted parameters are handed directly to linear programming and integer programming optimization models. The optimizers then construct the best possible decision scheme subject to the various operational constraints that real IoT environments impose, from energy budgets to task deadlines.

What distinguishes this framework from the conventional static paradigm, in which a prediction model feeds an optimizer and the output is accepted as final, is its rolling-horizon prediction-optimization-feedback mechanism. In the classical approach, any drift in the underlying data, a sudden change in load, a failing sensor, a shift in user behavior, silently degrades both the forecast and the plan built upon it. The new system instead uses the execution feedback generated by adaptive optimization to update the reconstructed input sequence and the optimization parameter states on a rolling basis. Each cycle of execution produces fresh evidence about how well the last decision performed, and that evidence flows back into the prediction model, which in turn feeds a revised plan to the optimizer. The result is a prediction-optimization-feedback loop that achieves rolling optimization and continuous model iteration rather than one-shot planning.

The reported numbers give a concrete sense of how the system performs. In the authors’ evaluation, the normalized load prediction achieved a root mean square error of 0.087, a figure that indicates the forecasting component tracked demand closely even as conditions varied. The optimal decision scheme produced by the optimization layer summed to a total expense of 1.24 multiplied by ten to the relevant power in monetary units, with an energy consumption index of 3.56 multiplied by ten kilowatt-hours. Task completion reached 98.7 percent, and the system decision error converged to approximately 0.063 during the closed-loop iteration process. Those last two figures matter most in practice: high task completion with low decision error suggests the feedback loop is doing exactly what it was designed to do, namely catching and correcting errors before they compound.

Benchmarking is where the study stakes its claim. The authors compared their framework against three recent benchmark methods and report that the proposed approach obtained lower prediction error, lower operating expense and higher stability of the decision. Stability is a metric that is easy to overlook but critical for IoT operators, because a decision system that oscillates between wildly different plans, even if each plan looks good on paper, creates churn that ripples through scheduling, provisioning and maintenance. A closed-loop design that smooths those oscillations by continuously reconciling predictions with observed outcomes offers a more dependable foundation for systems that must run unattended for long periods.

The significance of this work extends beyond any single deployment. IoT systems now underpin smart city traffic management, sustainable supply chains, industrial automation, environmental monitoring and precision agriculture, and in each of these domains the same structural problem recurs: sensor data arrives with mobility and uncertainty, predictions go stale quickly, and decisions made on stale predictions waste energy, money or both. The literature the authors survey reflects a field that has attacked these problems piecemeal, with deep learning time series forecasting on one side and optimization-driven resource allocation, fog-cloud offloading and Lyapunov-based control on the other. By arguing for the deep integration and mutual evolution of machine learning and operations research, the paper positions itself as a template for how those two threads can be woven into a single adaptive architecture.

Technically, the choice of linear and integer programming as the optimization backbone is worth unpacking. Linear programming handles continuous allocation problems, such as distributing computational load or energy across devices, where resources can be divided fractionally. Integer programming handles the discrete decisions that pervade IoT orchestration, such as whether a task is assigned to a particular edge node or deferred entirely. Because both formulations accept the predicted forward parameters of state variables as inputs, the prediction model effectively parameterizes the optimization problem in real time. This is what makes the coupling tight: the optimizer is not working from a static estimate of demand but from a forecast that is itself being retrained and corrected by the outcomes of previous optimization rounds.

The rolling-horizon mechanism also echoes a long tradition in control theory and operations research, where receding-horizon control has proven effective for systems that must be steered under uncertainty. What the authors add is the explicit feedback of execution results into the reconstructed input sequence of the learning model, closing the loop in a way that makes the entire pipeline self-correcting. In effect, the system treats every decision as an experiment whose outcome refines the next forecast. Over successive iterations, the decision error reported in the study settles toward its low value, which the authors interpret as evidence that the collaborative mechanism between prediction and optimization converges rather than destabilizes.

The authors are careful to frame their contribution as both theoretical and architectural. They write that the paper reveals the universality of the deep integration and mutual evolution of machine learning and operations research, and that the framework provides theoretical support and scalable architecture modes for building stable and adaptive intelligent decision-making systems for complex IoT environments. Scalability is the operative word: because the pre-processing layer unifies multi-source data and the loop structure is agnostic to the specific prediction or optimization algorithms plugged into it, the design could in principle be adapted to settings ranging from smart grids to logistics networks without redesigning the fundamental control structure.

For practitioners, the study offers a pragmatic lesson: the next generation of IoT intelligence will likely come less from ever-larger forecasting models and more from the discipline of how predictions, optimizations and real-world feedback are wired together. As billions of devices continue to join the network, the systems that manage them will need to make thousands of small decisions per second under uncertainty, and a closed-loop architecture that learns from its own choices, as demonstrated by Shen and colleagues, points toward a future in which prediction and optimization are no longer separate stages but a single, continuously evolving decision engine. The research, published in the special issue on Machine Learning and Big Data Analytics for IoT Security and Privacy, arrived online on 30 September 2026 as volume 38, article 760 of the journal.

Subject of Research: A closed-loop intelligent decision-making framework integrating machine learning prediction and operations research optimization for Internet of Things systems

Article Title: Intelligent decision-making system combining machine learning and operations research models in Internet of Things systems

Article References: Shen, Q., Yang, C., Bai, W., & Wang, C. (2026). Intelligent decision-making system combining machine learning and operations research models in Internet of Things systems. Neural Computing and Applications, 38(19), Article 760. https://doi.org/10.1007/s00521-026-12461-w

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12461-w

Keywords: Internet of Things, machine learning, operations research, closed-loop decision-making, rolling horizon control, linear programming, integer programming, time series prediction, optimization, sensor data, adaptive systems, Neural Computing and Applications

The core problem the team set out to solve is a familiar one in multimodal machine learning. Most existing pipelines process each modality independently, using a vision network for frames, an audio network for sound and a language model for captions or titles, and only combine the outputs in a late fusion stage. That design is convenient, but it comes at a cost. Information that only makes sense across modalities, such as the way a piece of music reinforces the mood of a scene, or the way a caption disambiguates a visually ambiguous object, can be lost before fusion ever happens. The separate encoders may also learn feature distributions that are poorly aligned with one another, so that the fusion layer receives streams that are difficult to reconcile. The result, the authors argue, is inefficient representation learning and a ceiling on classification accuracy for tasks such as microvideo category recognition.

MTA departs from that recipe by building extraction and fusion into a shared Transformer structure. Transformers, which rely on self-attention to weigh relationships between elements of a sequence, have become the backbone of modern artificial intelligence, powering everything from large language models to video vision transformers such as ViViT and multimodal self-supervised systems such as VATT. What distinguishes MTA from a conventional Transformer is what happens in its fusion layer. Instead of applying a static attention mechanism that treats all inputs uniformly, MTA incorporates a dynamic modality-aware attention mechanism. In practice, this means the model learns to adaptively attend to modality-specific information, shifting its focus depending on what a particular clip demands. For a clip in which the soundtrack carries most of the semantic signal, the audio stream can receive greater weight; for a clip driven by on-screen text or visual content, the mechanism can redirect attention accordingly. The dynamic weighting also helps align the heterogeneous feature distributions produced by different encoders, reducing the mismatch that plagues late-fusion designs.

A second conceptual innovation in the paper is the notion of categorical modality. In many multimodal classifiers, the fused representation is compared against class labels only at the very end, through a final classification layer that treats labels as inert symbols. MTA instead facilitates direct interaction between the fused multimodal features and label semantics during prediction. By giving the label side a learned semantic representation that can exchange information with the fused features, the model can exploit the relationships between categories themselves, for example when two microvideo categories share visual or textual vocabulary. The authors report that this end-to-end interaction improves learning effectiveness, because the prediction stage is no longer a passive readout but an active participant in shaping the shared representation.

The third pillar of the study is an investigation of parameter-sharing strategies. Sharing parameters across the branches of a multimodal network is a double-edged sword: it can regularize the model, reduce its size and encourage features that generalize across modalities, but excessive sharing can blur modality-specific distinctions. The researchers examined two sharing strategies within the MTA framework and found that both further boosted model performance, suggesting that a carefully chosen degree of weight sharing complements the dynamic attention mechanism rather than competing with it. This finding is practically significant for deployment, because parameter sharing reduces memory footprint and training cost, which matters when models must process the enormous volumes of short video that platforms handle daily.

Evaluation was carried out on a category-balanced benchmark dataset, a deliberate choice that removes the distorting effect of imbalanced class distributions, in which a model can appear accurate simply by favoring frequent categories. Across this benchmark, MTA consistently outperformed state-of-the-art multimodal fusion methods in classification accuracy. The authors present this as evidence of the robustness and scalability of the approach for real-world microvideo understanding tasks, and they have released their code publicly on GitHub, allowing other researchers to reproduce the results and build on the architecture.

The significance of the work becomes clearer when set against the broader landscape of multimodal research. Recent years have produced spectacular large multimodal models, from Flamingo and BLIP-3 to Llama 3 and the Qwen vision-language family, which achieve broad capabilities through massive pretraining. Yet specialized tasks such as microvideo category recognition, scene recognition and venue classification still benefit from architectures tailored to the specific structure of the problem, where clips are short, modalities are tightly interleaved and label sets are well defined. Prior work on microvideos has explored joint sequential-sparse modeling, neural multimodal cooperative learning, gated fully convolutional sequence models and attention-enhanced joint learning networks, each attacking pieces of the fusion problem. MTA’s contribution is to unify extraction and fusion within a single Transformer backbone while making the fusion itself adaptive, rather than fixed.

The applications that motivated the research are far-reaching. Content recommendation systems depend on accurate representations to surface clips that match user interests, and better multimodal understanding translates directly into more relevant feeds. Regulatory compliance is an equally pressing driver, since platforms must detect content that violates policies even when the violation is conveyed by the combination of audio and imagery rather than by either alone. Multimedia search, meanwhile, requires representations that capture what a clip is about, not merely what it looks like. A model that can weigh modalities dynamically and reason over label semantics offers a stronger foundation for all three use cases than pipelines that lose cross-modal information before fusion.

Technically, the design reflects several lessons from the Transformer literature. Self-attention allows every element of a sequence to attend to every other element, which makes it a natural fit for aligning features across modalities that unfold at different rates and granularities. Prior multimodal Transformers, such as the Multimodal Transformer for unaligned language sequences, demonstrated that cross-modal attention can bridge streams that are not temporally aligned. MTA extends this line by making the attention itself modality-aware and dynamic, so the alignment process is conditioned on the content being fused. Combined with the categorical modality mechanism, the architecture treats classification not as an afterthought but as an integrated stage of representation learning, echoing the end-to-end philosophy that has driven progress across deep learning.

The study, received in April 2025 and accepted in July 2026, was supported by the 111 Project, the Natural Science Foundation of Liaoning Province and the National Natural Science Foundation of China. Its limitations are those inherent to the field: the benchmark is category-balanced, while real-world feeds are anything but, and the reported data are available only on request due to privacy and ethical restrictions. Nevertheless, the consistent gains over existing fusion methods, the public code release and the architectural clarity of the approach position MTA as a meaningful step toward short-video AI systems that genuinely watch, listen and read at the same time. As platforms grapple with billions of daily microvideo views, architectures that fuse modalities adaptively within a shared Transformer structure are likely to shape how the next generation of recommendation, moderation and search systems makes sense of the world’s shortest, densest media.

Subject of Research: Multimodal microvideo understanding using a shared Transformer architecture with dynamic modality-aware attention

Article Title: Multimodal microvideo understanding based on the shared transformer structure

Article References: Zhong, Z., Zhang, W., Yu, H., & Zhu, Z. (2026). Multimodal microvideo understanding based on the shared transformer structure. Neural Computing and Applications, 38(19), Article 758. https://doi.org/10.1007/s00521-026-12409-0

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12409-0

Keywords: microvideo understanding, multimodal fusion, transformer, attention mechanism, deep learning, video classification, multimedia search, content recommendation, neural networks, representation learning, Neural Computing and Applications, Multimodal

The core insight of the work is that most of the records and most of the columns in a typical banking dataset are redundant. When a bank profiles a customer, it may hold hundreds of variables representing that individual, from transaction counts and account ages to demographic markers, and many of these variables overlap heavily. At the same time, thousands or millions of customer records often cluster into patterns that a handful of representative examples can capture. The computational cost of many classification algorithms grows with both the number of instances and the number of features, so every redundant record and every superfluous variable inflates training time, memory usage, and storage requirements. For an industry that runs classification tasks continuously across enormous customer bases, the savings from aggressive but careful reduction can be substantial.

To compress the number of instances, the researchers turned to a self-organizing map, an unsupervised neural network architecture introduced by Teuvo Kohonen in the 1980s. A SOM maps high-dimensional input data onto a usually two-dimensional grid of neurons, preserving the topology of the original data space: similar customers end up near each other on the map. Each input record activates a best matching unit, the neuron whose weight vector most closely resembles the record. By training the map and then using the neuron weight vectors as prototypes, the method replaces thousands of raw records with a far smaller set of synthetic representatives that summarize the structure of the data. In their experiments, this prototype generation step reduced the Census Income benchmark dataset by 99.3 percent and the proprietary banking dataset by 99.8 percent, meaning that fewer than one in a hundred original records needed to be retained to preserve the information the classifier requires.

Compressing instances is only half the battle. The second axis of reduction concerns the features themselves, and here the team proposed five weighting procedures that assign each variable a degree of relevance before the classification step. The most novel of these is a feature weighting technique based directly on the self-organizing map, dubbed FWSOM. Because the SOM has already learned a low-dimensional representation of the data, its internal structure encodes information about which variables drive the organization of the map. FWSOM exploits this by analyzing how strongly each feature contributes to the positioning of records on the trained map, producing weights that can then be folded into a weighted Euclidean distance metric used by the k-nearest neighbors classifier. Variables judged irrelevant receive low weights, effectively muting their influence without requiring a separate feature selection pass.

The remaining four procedures adapt permutation importance, a model-agnostic technique popularized by random forests and widely used in interpretability tools, to the peculiarities of SOM output. Permutation importance works by shuffling the values of one feature at a time and measuring how much the model’s performance degrades; a feature whose shuffling wrecks the predictions is clearly important. The catch is that the standard formulation assumes a conventional supervised model, not a topology-preserving map. The researchers therefore designed four tailored criteria: performance-based PI, which tracks changes in overall accuracy; error-based PI, which monitors shifts in classification error; prediction changes-based PI, which counts how often predictions flip when a feature is permuted; and F1-score-based PI, which focuses on the harmonic mean of precision and recall, a metric better suited to imbalanced datasets where one class, such as defaulters or fraud cases, is rare. Each variant offers a different lens on feature relevance, and the team evaluated all of them empirically.

The experimental design paired a public benchmark with real-world data. The Census Income dataset, drawn from the UCI Machine Learning Repository, is a classic classification problem in which the goal is to predict whether an individual earns above a certain threshold, a task structurally similar to many banking applications such as creditworthiness assessment. Alongside it, the researchers used a proprietary dataset from GAMCO covering real bank customers, giving the evaluation a dose of industrial reality that benchmark studies often lack. Performance was measured using standard metrics including accuracy, balanced accuracy, sensitivity, and false positive rate, ensuring that the compressed models were judged not merely on how fast they ran but on whether they still classified correctly, including on the minority classes that matter most in risk management.

The results were striking on both fronts. Permutation importance proved the more aggressive feature pruner, cutting 82.40 percent of features from the bank dataset and 21.49 percent from the income dataset. FWSOM, while more conservative, still removed 6.4 percent of features from the banking data and 14.29 percent from the income data, with the advantage that it integrates seamlessly into the SOM-based pipeline rather than requiring a separate model. Combined with the near-total prototype reduction, the full methodology delivered datasets that were orders of magnitude smaller than the originals. The practical consequences cascade: lower storage requirements, reduced memory footprints during training, faster classification of new customers, and simpler data analysis overall. For banks operating under tight latency constraints and rising cloud computing costs, a classifier that needs a fraction of the data to reach comparable decisions translates directly into money saved.

What makes the approach particularly appealing for the financial sector is its interpretability. Permutation importance produces an explicit ranking of which variables drive predictions, which aligns with regulatory expectations that automated decisions be explainable. Related research has applied permutation-based methods to default prediction and to identifying biomarkers in medicine, underscoring the technique’s versatility. By fusing that interpretability with the topology-preserving compression of the SOM, the Spanish team has built a pipeline in which the same structure that shrinks the data also reveals which customer attributes actually matter. The work builds on the group’s earlier prototype generation method using a growing self-organizing map, published in the same journal in 2023, extending it from instance reduction alone to a joint treatment of rows and columns.

The methodology also fits into a broader research landscape. Feature selection has been tackled with genetic algorithms, particle swarm optimization, Harris hawks optimization, sine-cosine algorithms, and variational autoencoders, each bringing computational overhead of its own. Prototype selection, meanwhile, has been explored through multi-armed bandits, geometric medians, and multilabel instance-based methods. The SOM-based approach distinguishes itself by addressing both dimensions of reduction within one unsupervised framework, and by being validated on proprietary banking data rather than benchmarks alone. The authors acknowledge limitations: the banking dataset and the code are proprietary and not publicly available, which restricts independent replication, and the balance between reduction and accuracy must be tuned per application. Still, the reported figures suggest that in data-heavy industries, the smartest model may be the one trained on far less. As machine learning spreads further into finance, healthcare, and beyond, techniques that strip away redundancy while preserving signal are likely to become as important as the algorithms they feed.

Subject of Research: A data reduction methodology combining self-organizing maps and permutation importance for k-nearest neighbors classification in the banking sector

Article Title: New data reduction methodology for classification applied to the banking sector

Article References: Ruiz-Moreno, S., Núñez-Reyes, A., García-Cantalapiedra, A., & Pavón, F. (2026). New data reduction methodology for classification applied to the banking sector. Neural Computing and Applications, 38(19), Article 757. https://doi.org/10.1007/s00521-026-12471-8

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12471-8

Keywords: machine learning, self-organizing map, k-nearest neighbors, data reduction, permutation importance, feature selection, banking, prototype generation, dimensionality reduction, Neural Computing and Applications, data, reduction