The significance of the role question goes back to foundational work in bullying research. In the 1990s, psychologists studying school bullying identified a cast of characters beyond the bully and the victim: assistants who support the aggressor, reinforcers who egg the behavior on, defenders who protect the target, and outsiders who simply look away. Cyberbullying inherits this social geometry, but with a twist. Because it unfolds across digital platforms, it is more pervasive, and the bystanders who witness it can respond in seconds, either with hostile retaliation against the bully or with supportive, victim-focused intervention. Prior research has shown that roughly half of university students involved in cyberbullying occupy dual roles, acting as both bully and victim, which makes the social dynamics even harder to untangle.

Most automated detection systems, however, have focused almost exclusively on the bully’s voice. The new study tackles the harder, less-studied problem of identifying victims and bystanders from text alone, a task constrained by the fact that privacy rules typically block researchers from accessing user metadata. The team worked with the AMiCA corpus, a collection of 113,694 English posts gathered from the question-and-answer platform Ask.fm between April and October 2013 by a Belgian research project. Ask.fm was popular among teenagers and frequently associated with cyberbullying, making it an ideal hunting ground. After cleaning, the dataset contained 112,247 posts annotated with roles: non-bully, harasser, victim, and bystander defender. The distribution was starkly imbalanced, with more than 106,000 non-bully posts but only 3,596 harasser posts, 1,354 victim posts, and 425 bystander defender posts.

The researchers pursued two parallel strategies. The first was conventional machine learning, in which features must be engineered by hand from the raw text. This was no small undertaking: the team built roughly 620,000 attributes spanning six feature families. Textual features included word- and character-level n-grams, part-of-speech counts, and named entity frequencies. Sentiment and emotion features drew on tools such as VADER, TextBlob, and the NRC emotion lexicon. Psycholinguistic features came from LIWC 2022 and Empath, while word embeddings ranged from static vectors like Word2Vec, GloVe, and FastText to contextual embeddings from BERT and its many derivatives. Notably, the team also crafted toxicity features using the Detoxify framework, extracting probabilities for labels such as toxic, obscene, threat, insult, and identity hate, an approach no prior role-identification study had taken.

The preprocessing pipeline itself reveals how messy real social media text can be. The researchers merged spaced-out letters, collapsed elongated characters like ‘youuuuuuu’ into ‘you’, preserved emojis with the emot package, resolved slang from online chat dictionaries, and used fuzzy string matching with a 90 percent similarity threshold to de-obfuscate profanity that users had deliberately disguised. Grammar and spelling were corrected with LanguageTool, and stopwords were deliberately retained because negations and pronouns carry crucial contextual signals in harassment scenarios. The class imbalance was handled by randomly downsampling the dominant non-bully class in the training data, while the test set was left untouched to preserve an unbiased evaluation.

The conventional models, multinomial logistic regression and a linear support vector classifier, were then trained with forward feature selection to find the best combinations. Textual features proved the strongest single group, achieving a macro F-measure of roughly 50 percent, while static word embeddings languished near 30 percent because they cannot capture the context-dependent tone that defines cyberbullying. Sentiment and emotion features performed poorly, suggesting that sarcasm and subtle aggression do not map neatly onto traditional sentiment dictionaries. The best logistic regression model, combining textual features, DistilBERT embeddings, psycholinguistic features, term lists, and toxicity scores, reached a macro F-measure of 54.58 percent on the hold-out test, beating the previous benchmark on the same dataset by 1.58 percentage points.

The second strategy, transfer learning, proved decisively more powerful. The team fine-tuned three compact pre-trained language models: DistilBERT, a distilled version of BERT that is 40 percent smaller and 60 percent faster; DistilRoBERTa; and Electra-small, which uses a replace-token-detection objective instead of masked language modeling. DistilBERT emerged as the clear winner. Fine-tuned for four epochs, it achieved a macro F-measure of 70.57 percent on the hold-out test, a roughly 10 percentage point improvement over the previous best result on the AMiCA corpus. At the level of individual roles, it scored an F-measure of 76.32 percent for harassers, 60.67 percent for bystander defenders, and 46.46 percent for victims, with recall improvements across every minority class compared to the conventional approach.

The victim role remains the hardest problem, and the error analysis explains why. Many misclassified victim posts were short, context-free utterances such as ‘yeah funny’ or ‘nice try goodbye honesty hour’, whose meaning depends entirely on the preceding conversation. Sarcasm also fooled the model: posts like ‘apparently you are beautiful’, actually harassment, were labeled as harmless because the literal words are benign. Long, emotionally mixed defender posts were similarly misclassified. Because the model reads each post in isolation, it cannot see the interaction history that would reveal who is responding to whom. The authors argue that incorporating conversational context and, eventually, large language models capable of explaining their predictions could address these failures.

The practical payoff is already visible. The team has released a public role-checker application built on the fine-tuned DistilBERT model, which classifies a submitted post as coming from a non-bully, harasser, victim, or bystander defender. The broader vision is intervention: platforms could use role-aware detection not merely to delete harmful content but to act against perpetrators, route support to victims, and encourage defenders, thereby interrupting the escalation dynamics that make cyberbullying so damaging. Because bystanders’ reactions critically shape whether an episode persists, knowing who is who in an interaction gives moderators leverage that binary detection never could.

The study also carries a methodological lesson for the field. Carefully engineered features still delivered a benchmark-beating conventional model, showing that domain knowledge retains value even in the era of transformers. But the efficiency gap is telling: logistic regression needed nearly nine minutes of cross-validation training with hand-crafted features, while DistilBERT required no feature engineering at all and processed four to five iterations per second during training. As the authors note, the findings may not generalize directly to other languages and cultures, where slang, sarcasm, and indirect communication differ. Still, by moving the field from asking whether cyberbullying occurs to asking how each participant is positioned within it, the work reframes automated moderation as a genuinely social problem, one where the victim’s plea and the defender’s intervention finally count as signals worth detecting.

Subject of Research: Automatic identification of participant roles in cyberbullying using machine learning and transfer learning

Article Title: Automatic classification of participants’ role in cyberbullying

Article References: Teng, T. H., Varathan, K. D., & Crestani, F. (2026). Automatic classification of participants’ role in cyberbullying. Heliyon, 12(15), Article e45491. https://doi.org/10.1016/j.heliyon.2026.e45491

Image Credits: AI Generated

DOI: 10.1016/j.heliyon.2026.e45491

Keywords: cyberbullying, natural language processing, machine learning, transfer learning, DistilBERT, text classification, bystanders, online harassment, feature engineering, Ask.fm, toxicity detection, class imbalance

The starting point of the work is the additive rule ensemble, a prediction model composed of a small set of if-then rules whose outputs are summed, each weighted by a coefficient. In the classical formulation, a rule’s condition is a conjunction of simple threshold propositions such as a single variable exceeding a value, for example age greater than 45 or body mass index above 30. Geometrically, each such rule carves an axis-parallel region out of the input space, a rectangle or box whose faces align with the coordinate axes. This structure is what makes the models so readable: a human can mentally evaluate each threshold comparison, combine them with logical conjunctions, and trace exactly why a prediction was made. The authors note that these properties, known as simulatability and modularity in the explainable AI literature, are precisely what large ensembles like random forests lose, since even though forests can be expressed as rule ensembles, their sheer scale defeats human comprehension.

The catch is that axis-parallel rules are only compact when the input features are well curated. If the true decision boundary of a problem runs diagonally through the data, a box-shaped rule cannot capture it in one stroke; instead, the ensemble must stack many small axis-parallel regions in a staircase pattern, inflating both the number of rules and their length. The Monash team’s solution is to generalize the atomic proposition from a single-variable threshold to a weighted linear inequality, where a learnable, sparse weight vector combines several input variables before comparison with a threshold. Geometrically, this turns the decision regions from axis-parallel boxes into general polyhedra with oblique, slanted faces. Crucially, the weights are kept sparse, meaning only a handful of variables contribute to any single inequality, so each proposition remains something a person can plausibly compute in their head.

The authors illustrate the payoff with a diabetes risk model built from the 2017-18 CDC NHANES survey. A conventional axis-parallel rule ensemble and their oblique version achieved approximately identical log loss, but the oblique model did so with a markedly simpler rule set. The key was that, combined with log transforms, the learnable linear conditions allowed the model to represent ratio concepts automatically, in this case effectively reconstructing body mass index from height and weight, a feature that would otherwise have to be supplied by manual feature engineering. This is the deeper significance of the method: it equips rule ensembles with a linear, and therefore still principally interpretable, form of representation learning, letting the model discover interaction features like products and ratios on its own rather than depending on a domain expert to hand them over.

Making this idea algorithmically practical required solving a hard optimization problem inside every boosting round. The method builds on fully corrective gradient boosting, which starts from an empty model and iteratively adds the rule condition that maximizes a boosting objective, here the gradient sum, while re-optimizing all rule weights by convex optimization after each addition. The authors’ central technical insight is that finding the optimal oblique proposition for the gradient sum objective reduces to a weighted linear classification problem: the labels are the signs of the loss gradients over the training examples, the weights are the gradient magnitudes, and the goal is to find a sparse hyperplane that separates them. Because directly minimizing the 0/1 loss is intractable, they substitute an l1-regularized logistic loss, using a bisection search over the regularization parameter to land on a solution with exactly the desired number of non-zero weights, then refit the selected coefficients without regularization to remove estimation bias. The efficient LibLinear solver handles the underlying convex problems.

On top of this building block, the team designed a search strategy that avoids a subtle but important pitfall. A naive adaptation of greedy rule growing would always prefer oblique cuts over adding new simple propositions, encouraging dense linear transformations and wasting the complexity budget. Instead, their method systematically enumerates candidate rule conditions at every complexity level from one up to a maximum, allowing previously learned propositions to be refitted with higher sparsity as the budget grows, and then selects among the level-wise candidates using validation risk rather than the raw boosting objective. This ensures the search can discover combinations of genuinely simple propositions instead of prematurely committing to complicated ones.

The computational analysis shows the approach is surprisingly cheap. The total worst-case running time is linear in the number of training examples, in contrast to axis-parallel methods that rely on presorting data for fast cut-point search, and the bisection search over regularization values behaves essentially as a constant factor in practice, requiring fewer than 20 iterations across all benchmark datasets. For the small ensembles targeted by interpretable modeling, the cost of refitting rule weights is negligible relative to condition learning, so LLTBoost runs without substantial overhead compared to traditional axis-parallel rule boosting, a claim the timing experiments confirm directly.

The empirical evaluation compared LLTBoost against four families of competitors: traditional gradient boosting of axis-parallel rules, a RuleFit-style generate-and-select pipeline built on forests of oblique trees, neural rule ensembles that combine tree-extracted rules with continuous optimization, and explainable boosting machines that model the target as sums of univariate step functions plus pairwise interactions. Because every method can trade accuracy against complexity, the authors compared full risk-complexity Pareto fronts, measuring model complexity as the sum of rule counts, proposition counts, and non-zero weight entries, and normalizing test risk by the variance or entropy of an empty model so results are comparable across datasets. LLTBoost achieved the lowest overall test risk on the majority of the 14 datasets within a complexity budget of 100, and when the authors scanned each method’s Pareto front for the minimum complexity needed to hit risk-reduction targets of 25, 50, and 75 percent, LLTBoost required the least complexity on most datasets, with the advantage widening at the demanding 75 percent target.

The authors are candid about the limits of their complexity metric. Counting non-zero weights treats an oblique inequality as roughly equivalent to a simple threshold, but the cognitive effort of mentally evaluating a weighted sum of several variables is genuinely higher than checking whether one number exceeds another. Quantifying that difference, they note, is a question for interdisciplinary work with cognitive science. Still, they point to practical mitigations within machine learning itself, such as restricting weight values to small integer ranges in the style of integer scoring systems, which becomes especially appealing in combination with log transforms because it lets rules express clean products and ratios of variable powers.

The broader message is that a modest change to the atomic unit of a rule, from a single threshold to a sparse linear inequality, can ripple through an entire modeling framework, delivering simpler models at equal accuracy and reducing the field’s dependence on manual feature engineering. With a Python implementation publicly available on GitHub and a single interpretable complexity parameter governing the trade-off, LLTBoost lowers the barrier for practitioners who need models that both perform well and can be read, checked, and defended by the people who use them. In an era when regulatory and ethical pressure increasingly demands explainable automated decisions, techniques that shrink the accuracy gap between transparent and opaque models may prove among the most consequential tools in applied machine learning.

Subject of Research: Sparse oblique rule boosting for interpretable additive rule ensembles in machine learning

Article Title: Sparse oblique rule boosting for simpler additive rule ensembles

Article References: Behzadimanesh, S., Le Bodic, P., Webb, G. I., & Boley, M. (2026). Sparse oblique rule boosting for simpler additive rule ensembles. Data Mining and Knowledge Discovery, 40(6), Article 109. https://doi.org/10.1007/s10618-026-01241-8

Image Credits: AI Generated

DOI: 10.1007/s10618-026-01241-8

Keywords: machine learning, interpretable AI, gradient boosting, rule learning, additive rule ensembles, sparse linear transformations, oblique decision boundaries, data mining, statistical learning, model complexity, feature engineering, LLTBoost

The commercial stakes are considerable. The global fresh strawberry market has been estimated at roughly 20.7 billion US dollars in 2024, with projections exceeding 30 billion dollars by the mid-2030s as year-round demand grows. In the United Kingdom alone, official statistics value the strawberry crop at 389 million pounds for 2024. At daily resolution, even modest forecasting errors ripple through the entire operation: harvest crews are scheduled against predicted volumes, packing lines and storage capacity are booked in advance, and supermarkets hold contracts that assume reliable supply. A model that misses a peak picking day by a wide margin creates labour inefficiencies on one hand and post-harvest waste on the other.

What makes strawberries so hard to forecast is that daily yield is not a string of independent numbers but a biologically structured time series. Production follows a recognisable lifecycle: output starts at zero before harvest begins, climbs as plants enter peak fruiting, and falls back to zero as the season ends. That rise-peak-decline shape is governed by cultivar, planting density, accumulated temperature and humidity, and the decisions of farm managers. The researchers argue that a realistic forecasting framework must satisfy three conditions at once. It must preserve the natural lifecycle, including genuine zero-yield periods at season boundaries. It must enforce strict temporal alignment so that only information available before harvest can be used as a predictor. And it must test models on unseen future seasons rather than interpolating within the data it has already seen.

That last condition turns out to be the study’s sharpest methodological point. In agricultural machine learning, it is dangerously easy to inflate performance through information leakage, for example by letting weather recorded after a harvest event act as a predictor, or by using random train-test splits on strongly seasonal data. The team deliberately excluded aggregate productivity indicators such as yield per hectare and percentage picked from the feature set, because those variables indirectly encode the answer. Environmental sensor readings from the farm’s SoilMoistureSense platform, covering temperature and relative humidity, were aggregated strictly within calendar-day boundaries and merged by date so that no future information could slip into the pipeline.

The dataset itself is a portrait of messy commercial reality rather than a tidy laboratory experiment. It was assembled from three sources: annual weekly planning spreadsheets, 189 individual daily forecast files recording both the farm’s own operational predictions and actual harvested weights, and sub-daily environmental sensor exports. After cleaning and harmonisation across years, the modelling dataset contained 2,280 daily observations, which rose to 3,540 after the researchers appended seven zero-yield days at the start and end of each harvest cycle. That padding step, an ablation analysis showed, mattered a great deal: removing it degraded the best model’s mean absolute error from about 340 kilograms to nearly 611 kilograms, while extending padding to 14 days lowered absolute error further but reduced explained variance. The seven-day configuration offered the best balance.

Feature engineering was deliberately biological. Rolling averages of prior yield over 1, 3, 7, 14 and 21 days captured what the authors call production memory, the tendency of strawberry harvests to persist across neighbouring days as overlapping cohorts of fruit mature and ripen. Rolling seven-day temperature averages and 14-day humidity averages represented accumulated environmental exposure, reflecting how heat and moisture stress modulate fruit development. Plant counts from the planning spreadsheets served as a static capacity indicator for each field. Together these features let the models learn whether production is rising, plateauing or declining, without ever violating temporal causality.

The model comparison spanned thirteen approaches, from simple ridge regression through gradient boosting frameworks such as XGBoost, LightGBM and CatBoost, to compact neural networks including a multilayer perceptron, a small LSTM and a temporal convolutional network, and finally TabPFN, a pretrained tabular foundation model. The results were striking. Under a conventional random 80/20 split, TabPFN achieved the lowest error with a mean absolute error of 340.18 kilograms and an R-squared of 0.733, while the neural sequence models performed worse than a constant mean predictor, producing negative R-squared values. Leave-one-out cross-validation told the same story, with TabPFN again leading at a mean absolute error of 338.09 kilograms. The authors note that with only six seasons of data, the strong inductive biases of tree ensembles and pretrained tabular models matter far more than architectural depth, which is precisely why the from-scratch deep learning baselines struggled.

The most deployment-relevant test, however, was forward-year validation: train on 2020 through 2024, then predict the entirely unseen 2025 season. Under this protocol TabPFN remained the strongest performer with a mean absolute error of 438.53 kilograms and an R-squared of 0.696, followed by Random Forest, Histogram Gradient Boosting and Extra Trees. The rise in error relative to the random split is not a failure, the researchers argue, but an honest measure of genuine interannual variation in weather, phenology and management. Permutation importance analysis on the 2025 test season confirmed that the models rely on biologically sensible variables: recent yield lags dominate the ranking, and interaction features combining recent production with temperature and humidity contribute measurably, suggesting that weather modulates an established production trajectory rather than acting as an independent driver.

The study is candid about its limits. All data come from a single farm, so the work demonstrates temporal generalisation to an unseen season within one commercial setting, not transfer across farms or regions. The available sensors captured only temperature and humidity, and the authors identify the absence of solar radiation, day length, irrigation records and growing degree days as a likely explanation for a recurring error mode: the systematic underestimation of peak harvest days, when similar recent conditions can conceal very different yield potential. Adding phenology-aware signals such as flower or truss counts, whether recorded manually or by computer vision, is flagged as a plausible route to better peak capture.

Even so, the implications are significant for an industry built on thin margins and perishable goods. The findings suggest that daily yield forecasting can be integrated into commercial decision-support systems using the data farms already collect, without exotic sensors or vast historical archives. In small, discontinuous seasonal datasets, the study concludes, strong inductive bias and regularisation beat model complexity, and rigorous forward-year evaluation is the only honest yardstick of what a forecasting system will actually deliver when the next season begins. For growers weighing the promise of agricultural AI, that combination of realism and rigour may be the most valuable harvest of all.

Subject of Research: Daily strawberry yield forecasting with artificial intelligence in commercial horticulture

Article Title: Daily strawberry yield forecasting using artificial intelligence in commercial horticulture

Article References: Ahmed, S., & Caldwell, N. H. (2026). Daily strawberry yield forecasting using artificial intelligence in commercial horticulture. Smart Agricultural Technology, 15, Article 102583. https://doi.org/10.1016/j.atech.2026.102583

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102583

Keywords: strawberry, yield forecasting, machine learning, TabPFN, deep learning, precision agriculture, time series, commercial horticulture, feature engineering, forward-year validation, IoT sensors, UK farming

The review, conducted by Atefeh Abedzadeh Attar, Mohammad Hossein Moattar and Esmaeil Kheirkhah of Islamic Azad University in Mashhad, Iran, systematically analyzed 69 studies published between 2020 and 2026 that apply machine learning (ML) and deep learning (DL) techniques to electroencephalography (EEG) and event-related potentials (ERPs) for depression diagnosis. The authors searched Scopus and Web of Science, screened 187 initial records down to the final set, and organized the field around a complete analytical pipeline: signal preprocessing, feature engineering, feature selection, and classification. Their central message is that artificial intelligence can extract objective, reproducible biomarkers of depression from brain signals that clinicians currently cannot see with the naked eye.

Why EEG? Unlike MRI and fMRI, which are expensive, immobile and offer poor temporal resolution, EEG is cheap, portable and captures neural activity at millisecond precision. It records five canonical frequency bands—delta (0.1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz) and gamma (30–40 Hz)—each tied to different cognitive states. ERPs, the brain’s stereotyped voltage responses to specific stimuli, add another layer: depressed patients show characteristic abnormalities in components such as P1 and P2 and heightened sensitivity to negative stimuli. Together, resting-state EEG and stimulus-evoked ERPs offer a complementary window into the neural signatures of the disorder, something the review emphasizes is often missed by studies that examine only resting-state data.

Before any algorithm can learn, the raw signals must be cleaned. EEG recordings are contaminated by eye movements, muscle activity and power-line interference, so the reviewed studies relied heavily on finite impulse response filters, discrete wavelet transforms, independent component analysis, notch filters and even Kalman filters to strip out noise. A second, less obvious problem is dimensional explosion: a single recording can yield thousands of features from only dozens of patients, inviting overfitting. The authors document how researchers tamed this with principal component analysis, locally linear embedding, correlation and variance thresholds, and in one extreme case compressed a set of 10,800 features before modeling even began.

Feature engineering emerges as the heart of the enterprise, and the review organizes it into three families. Handcrafted approaches extract interpretable quantities directly: power spectral density across frequency bands, synchronization measures such as the self-synchronization index and the phase lag index, functional connectivity matrices, and nonlinear descriptors like Lempel–Ziv complexity and various entropies. End-to-end deep learning models—convolutional neural networks, CNN-LSTM hybrids, attention mechanisms, graph convolutional networks and even spiking neural networks—learn their own representations from raw or transformed signals. Between the two sit hybrid methods, which feed manually engineered features such as STFT spectrograms or directed connectivity measures into deep networks, combining human domain knowledge with automated pattern discovery.

The performance numbers are striking. On the machine learning side, accuracies ranged from 82.68 to 100 percent, with support vector machines and K-nearest neighbors dominating the literature; boosting methods such as XGBoost reached 98.92 percent, and one lightweight LightGBM framework hit 97.42 percent using only three prefrontal electrodes—evidence that wearable-scale screening is technically feasible. Deep learning models ranged from 77.78 to 100 percent, with CNN-LSTM architectures reaching 99.9 percent in a real-time wearable system called DepCap, and attention-based and graph-based models consistently exceeding 95 percent. Reported biomarkers converge on frontal and temporal regions, elevated theta and alpha power, interhemispheric asymmetry in delta, alpha and beta bands, and disrupted functional connectivity in parietal-occipital networks.

But the review delivers a sharp caution about those headline numbers. Several studies reporting perfect 100 percent accuracy relied on validation schemes that may leak subject-specific information: when EEG recordings are chopped into segments and randomly split into training and test sets, segments from the same person can appear on both sides, letting the model memorize individuals rather than learn the disorder. One study’s accuracy fell from 100 percent under ordinary 10-fold cross-validation to 83.96 percent under leave-one-subject-out cross-validation, a dramatic illustration of the problem. The authors argue that subject-independent validation, transparent data partitioning and strict leakage controls should be mandatory before any accuracy claim is taken at face value.

The two AI paradigms also carry distinct trade-offs. Machine learning models are interpretable and computationally cheap, letting clinicians trace which biomarkers drove a decision, and they perform well on small datasets—but they depend on laborious, error-prone manual feature pipelines. Deep learning models are robust to noise and capable of real-time monitoring, early detection and severity tracking, yet they demand large datasets, heavy computation, and suffer from black-box opacity that undermines clinical trust. The review also flags ethical concerns: most studies enrolled fewer than 100 participants, raising doubts about generalizability across populations, and over-reliance on automated systems risks eroding the human empathy central to mental health care. The authors call for explainable AI techniques, adherence to reporting standards such as TRIPOD and PROBAST, and patient-centered deployment.

Looking forward, the review sketches a roadmap for the field: larger and more diverse public datasets with standardized acquisition protocols, multimodal models that fuse EEG with speech, eye-tracking and clinical records, federated learning to share knowledge across hospitals without moving sensitive patient data, and foundation models pre-trained on large biomedical corpora to overcome the small-data bottleneck. It also highlights a practical dual trend—increasingly sophisticated architectures on one hand, and radically simplified systems using as few as two frontal electrodes on the other—that could bring objective, AI-assisted depression screening out of the laboratory and into clinics, wearables and underserved communities worldwide. If those validation and equity challenges are met, the authors conclude, brain-signal-based AI could transform depression diagnosis from a subjective art into an objective, scalable science.

Subject of Research: Machine learning and deep learning approaches for diagnosing depressive disorders from EEG and ERP signals

Article Title: A review on machine learning and deep learning approaches for depressive disorder diagnosis based on EEG and ERP signals

Article References: Abedzadeh Attar, A., Moattar, M. H., & Kheirkhah, E. (2026). A review on machine learning and deep learning approaches for depressive disorder diagnosis based on EEG and ERP signals. Discover Artificial Intelligence, 6(1), Article 1308. https://doi.org/10.1007/s44163-026-02334-5

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02334-5

Keywords: depression, EEG, ERP, machine learning, deep learning, brain signals, diagnosis, biomarkers, neural networks, mental health, feature engineering, clinical AI

The research team, led by Hyeong-Yun So, Hyeon-Cheol Yoon, and Tae-Gyun Kim of the National Integrated Drought Center at the National Disaster Management Research Institute, together with Se-Jeong Lee of the Korea Institute of Hydrogical Survey, focused on the domestic and industrial water supply sector of South Korea’s drought warning system. Unlike the meteorological sector, which relies on indices such as the Standardized Precipitation Index, or the agricultural sector, which monitors reservoir storage and soil moisture, the domestic and industrial sector bases its four warning stages—attention, caution, alert, and severe—on dam-specific water supply adjustment criteria tied to storage volumes and storage ratios. Predicting those storage levels quantitatively, months in advance, therefore translates directly into earlier and better-calibrated drought warnings for cities and industry.

The study area was Juam Dam itself, a multipurpose dam in the Boseonggang River basin, a tributary of the Seomjingang River, with a total storage capacity of 457 million cubic meters. During the prolonged meteorological drought that stretched from the second half of 2022 into the first half of 2023, Juam Dam’s storage fell to the severe stage, disrupting major industrial complexes at Yeosu and Gwangyang and prompting consideration of domestic water supply restrictions in Gwangju. The researchers assembled 32 years of daily hydrological and meteorological observations, spanning 1991 to 2022, from the National Water Resources Management Information System and the Korea Meteorological Administration’s Weather Data Portal, and reserved the drought-stricken first half of 2023 as an independent test period.

A central challenge emerged immediately from correlation analysis. Using the Pearson correlation coefficient after min–max normalization, the team found that even the most strongly related input variable, dam outflow, showed a correlation with storage of only r = 0.253, followed by inflow, humidity, wind speed (negatively correlated), temperature, and rainfall. In other words, the linear relationships between the available observations and the target variable were weak across the board. This is precisely the regime in which deep learning models, capable of capturing temporal lags and complex nonlinear interactions, outperform traditional regression and conventional machine learning approaches—but only if the input representation is designed well. The team therefore applied feature engineering, generating lagged variables, arithmetic transformations, and moving averages from the raw observations, and built three distinct training datasets. Dataset A contained only the original variables; Dataset B incorporated all expanded features; Dataset C retained only the strongly correlated expanded features derived from inflow, outflow, and antecedent storage, deliberately excluding meteorological parameters.

Each dataset was paired with four deep learning architectures, producing twelve candidate models. The Long Short-Term Memory network, a recurrent architecture with forget, input, and output gates and a cell state that preserves long-term information, served as the field’s standard baseline. The Bidirectional LSTM added a backward pass over the sequence, combining past and future context. The hybrid CNN-LSTM stacked one-dimensional convolutional feature extraction in front of recurrent sequence learning. The fourth architecture, the Temporal Convolutional Network, replaced recurrence entirely with causal one-dimensional convolutions—ensuring that predictions depend only on current and past inputs—and dilated convolutions, which insert gaps between filter elements to expand the receptive field exponentially across time. That dilation mechanism allows the TCN to learn patterns over very long horizons while remaining fully parallelizable during training, a structural advantage over step-by-step recurrent processing.

The comparative results were decisive. Models trained on Dataset B, with its full set of engineered features, reduced test RMSE by roughly 14 percent relative to Dataset A, while Dataset C—stripped of meteorological variables—performed dramatically worse, with test RMSE increasing by about 34 percent over Dataset A. The lesson was counterintuitive but important: although temperature, humidity, and wind speed are physically coupled to inflow and outflow through evapotranspiration, removing them deprived the models of nonlinear context about long-term storage depletion. Deep learning architectures, the authors conclude, thrive on multi-dimensional interactions among many weakly correlated variables rather than on a few strongly correlated linear inputs. Across architectures, the TCN dominated, achieving average RMSE values of 12.150 in training, 8.823 in validation, and 21.337 in testing, while the LSTM fared worst at 44.818, 17.296, and 57.950 respectively—reductions of roughly 73, 49, and 63 percent in the TCN’s favor. Qualitatively, however, the LSTM reproduced the temporal shape of storage variation most faithfully, and the TCN delivered the closest numerical agreement, a tension the team resolved by optimizing the TCN further.

Optimization proceeded through a two-stage grid search over the TCN’s hyperparameters, which are known to strongly influence its behavior: activation function, dilation rate, filter size, kernel size, learning rate, batch size, number of epochs, and the lookback period of historical input. The first stage, generating 256 model configurations, converged on the ReLU activation, a dilation rate sequence of 1, 2, 4, and 8, and 200 training epochs. The second stage revealed a striking finding: the optimal lookback window was 1,460 days—approximately four years. The authors attribute this to the operational rhythm of South Korean dam management, in which the standardized Dam Water Supply Adjustment Criteria are revised every two to four years. A four-year memory allows the model to internalize these cyclical regulatory shifts and the human-controlled discharge patterns they produce. Notably, the configuration with the lowest training and validation error showed unstable oscillatory predictions and poor generalization, so the final hyperparameters were selected from the first-stage results to guarantee robustness over raw precision.

The final optimized model was evaluated four times across 2023, each with a different forecast origin, and its predictions were compared against Juam Dam’s operational warning thresholds, where consecutive drought stages are separated by roughly 4.5 to 5 percent of total storage. The model captured the severe drought of early 2023, reproduced the storage recovery toward the end of June, tracked spring and flood-season dynamics, and matched the stable conditions of late year. Its one clear weakness was the abrupt storage surge in mid-July driven by an extreme rainfall event totaling 306 millimeters in five days, which the model underestimated. Quantitatively, the four evaluations yielded an average Mean Absolute Error of 28.491, an average RMSE of 35.440, and an average error rate of 6.23 percent of total storage, with the best quarter achieving an error rate of just 3.68 percent. In earlier comparative testing, the TCN model also achieved a 78.9 percent hit rate within a 5 percent error margin of observed storage, with a correlation coefficient exceeding 0.6 in the Taylor diagram analysis.

The authors are candid about what a 6.23 percent error rate means operationally. Because warning stages are densely spaced, a forecast of this precision could misclassify a specific stage if used for short-term, deterministic actions such as triggering water restrictions, which demand errors below roughly 3 percent. But the purpose of a six-month outlook is different: it functions as an early-stage monitoring tool for macro-level storage trends and precursory signs of impending shortage. Within that framing, capturing a downward storage trajectory within about 6 percent provides water managers an extended proactive window to secure alternative supplies and establish conservative safety margins well before crisis points arrive—the very lead time that was missing during the Juam and Obong crises.

The research team outlines an ambitious path forward. They plan to extend the framework to all 33 dams covered by South Korea’s domestic and industrial drought forecasting system, developing dam-specific datasets and models that reflect each reservoir’s unique hydrology, and to investigate stretching predictions to a twelve-month horizon. An ensemble framework is planned that would combine the TCN’s error-suppressing precision, the LSTM’s skill at reproducing temporal trends, and the Transformer’s capacity for long-range context, with performance assessed using additional metrics such as the Nash-Sutcliffe Efficiency and the Kling-Gupta Efficiency. To address the models’ struggle with the non-stationary extremes of a changing climate—intense short-duration rainfall and prolonged drought alike—the team intends to accumulate training data on unprecedented events and generate synthetic, physics-informed virtual datasets. If those efforts succeed, the six-month dam storage forecast demonstrated here could evolve from a research result into a routine instrument of national drought preparedness, giving water managers the one resource drought victims never have enough of: time.

Subject of Research: Midterm drought forecasting using deep learning prediction of dam reservoir storage dynamics

Article Title: Midterm drought forecasting based on dam storage prediction using deep learning algorithms

Article References: So, H.-Y., Yoon, H.-C., Kim, T.-G., & Lee, S.-J. (2026). Midterm drought forecasting based on dam storage prediction using deep learning algorithms. Environmental Earth Sciences, 85(16), Article 407. https://doi.org/10.1007/s12665-026-13138-2

Image Credits: AI Generated

DOI: 10.1007/s12665-026-13138-2

Keywords: drought forecasting, deep learning, temporal convolutional network, LSTM, dam storage prediction, feature engineering, grid search, hyperparameter optimization, water resource management, South Korea, Juam Dam, hydrological modeling

The research team, led by Payal Mahajan and Aniket A. Muley of Swami Ramanand Teerth Marathwada University in Nanded, together with Madhav R. Fegade of Digambarrao Bindu Arts, Commerce and Science College in Bhokar, built their framework on daily mandi-level price records from the AGMARKNET portal, the official agricultural marketing database of the Government of India. The commodities spanned the breadth of Indian farming: Arhar, Bajra, Cotton, Maize, Paddy, Sesamum, Soyabean, Sunflower and Wheat. Records dating from 2001 through October 2025 were cleaned, validated and aggregated into monthly average modal prices, the most commonly observed market price for each commodity, producing long monthly price series that capture decades of market behaviour.

One of the study’s central contributions is methodological rather than predictive. In many published forecasting papers, preprocessing steps such as outlier removal are performed on the entire dataset before the model is tested, which silently leaks information from the future into the training process and inflates apparent accuracy. The Maharashtra team avoided this trap entirely. Extreme prices were capped at commodity-specific percentiles, but the quantile limits were estimated only from training data available at each sequential forecasting step and then applied to the following test month. In other words, the pipeline was designed so that no future observation could ever influence a past forecast, a discipline that many machine learning benchmarks in economics and agriculture still lack.

The feature engineering pipeline was deliberately rich. Lag features captured prices from one, two, three, six, nine, twelve, eighteen and twenty-four months earlier, allowing the models to weigh both recent momentum and annual cycles. Rolling means over three, six, twelve and twenty-four month windows summarised short-, medium- and long-term trends, while rolling standard deviations of the same lengths encoded volatility, telling the models whether the market had been calm or turbulent in the preceding months. Seasonal calendar variables were encoded with sine and cosine transformations of the month number, a standard technique that preserves the cyclical nature of the year so that December and January are treated as neighbours rather than as numerically distant values. All prices were log-transformed to stabilise variance and dampen the influence of sudden spikes.

Four tree-based ensemble algorithms were put to the test: Random Forest, which aggregates decision trees trained on bootstrapped samples; Extra Trees, which injects additional randomisation into split selection for robustness; Histogram Gradient Boosting, a fast boosting method optimised for structured numerical data; and XGBoost, a regularised gradient boosting system that sequentially corrects the errors of weak learners. Models were trained with fixed hyperparameters and evaluated with Mean Absolute Error, Root Mean Squared Error, the coefficient of determination and Mean Absolute Percentage Error, using an expanding-window protocol in which each month from January 2021 to October 2025 was forecast using everything known before it, with the training set growing as actual prices became available.

The headline result is stark. The one-month naive forecast, which simply predicts that the coming month’s price equals the previous month’s, achieved the lowest MAPE for every single commodity, ranging from just 2.30 percent for Paddy to 4.91 percent for Sunflower. The best machine learning models, however, consistently outperformed the twelve-month seasonal naive benchmark, which predicts that this month’s price will match the same calendar month one year earlier. This distinction matters. It shows that the engineered lag, rolling and seasonal features do capture genuine nonlinear temporal structure beyond annual repetition, but not enough extra information to improve on the sheer stubbornness of short-run price continuation in aggregated monthly data.

Why is the naive benchmark so formidable? Monthly agricultural prices at the state level adjust gradually. Supply arrives in bulk through regulated markets, demand shifts slowly, and administrative factors such as minimum support prices dampen sudden swings. In such a regime, the previous month’s price already embeds most of the information needed for a one-step-ahead forecast, leaving little room for algorithms to extract additional signal from historical prices alone. The study’s authors note that abrupt movements driven by rainfall deficits, market arrivals, input costs, policy interventions or export-import shocks simply cannot be anticipated from the price series itself, which caps what any historical-price-only model can achieve.

The experiments also revealed meaningful differences among commodities and validation strategies. Switching from fixed-window training to expanding-window retraining, in which models were periodically refreshed with newly observed prices, reduced forecast errors for all nine commodities, confirming that regular updating is worthwhile even when it does not overturn the naive benchmark. Paddy and Wheat proved relatively easy to forecast, reflecting their lower relative price variability, while Arhar, Cotton, Sesamum, Soyabean and Sunflower, commodities with higher volatility and more erratic price behaviour, remained challenging. Sensitivity analyses with different training start years showed that no single configuration was uniformly best, underscoring that commodity-specific market dynamics, rather than any universal algorithmic choice, shape forecastability.

The researchers are transparent about the limitations of their design. Exogenous drivers such as rainfall, temperature, market arrivals, fuel and transport costs, inflation, festival demand and government policy were deliberately excluded, making the study a controlled assessment of what historical prices alone can deliver. Deep learning architectures such as LSTM and GRU networks were also left out, since a fair comparison would require a separate chronological tuning and architecture-selection study under the same leakage-safe principles. The authors frame both omissions as clear directions for future work, alongside probabilistic forecasting that would quantify uncertainty rather than produce single point estimates.

For farmers, traders and policymakers, the practical message is double-edged but valuable. First, for month-ahead planning, the cheapest forecast on the table, carrying last month’s price forward, is remarkably hard to beat, and any proposed machine learning service should be benchmarked against it before being trusted. Second, meaningful improvements in agricultural price forecasting will not come from more elaborate algorithms chewing on the same price history, but from richer information: weather data, crop arrivals, production estimates, cost indices and policy signals integrated into leakage-safe evaluation frameworks like the one this study demonstrates. In an era when artificial intelligence is often oversold, this Maharashtra experiment stands out as a model of scientific candour, showing exactly where machine learning helps, where it does not, and how the next generation of forecasting systems should be built.

Subject of Research: Machine learning models for forecasting monthly agricultural commodity prices in Maharashtra, India

Article Title: Forecasting agricultural commodity prices using machine learning models

Article References: Mahajan, P., Muley, A. A., & Fegade, M. R. (2026). Forecasting agricultural commodity prices using machine learning models. Discover Informatics, 1(1), Article 15. https://doi.org/10.1007/s44564-026-00015-0

Image Credits: AI Generated

DOI: 10.1007/s44564-026-00015-0

Keywords: agricultural commodity prices, machine learning, price forecasting, time series, Random Forest, XGBoost, AGMARKNET, Maharashtra, feature engineering, naive benchmark, ensemble learning, agricultural economics

The work, led by Roger Arnau, Jose M. Calabuig, Nuria Ortigosa and Luiza Petrosyan, addresses a stubborn problem in agricultural economics: price series for different crop varieties cover seasons of wildly different lengths, start at different times of year, and fluctuate on very different scales. Comparing such series directly with conventional clustering tools, which typically rely on Euclidean distance between raw data points, tends to group varieties that merely share similar absolute price levels while ignoring whether their prices rise, fall or oscillate in the same way. The researchers’ solution is to abandon raw prices altogether and instead describe each variety-province-season combination with a compact set of five interpretable features: the duration of the marketing season, the variance of prices, the slope of a linear trend fitted to the weekly prices, the R-squared quality of that fit, and a novel Q-ratio that captures price amplitude per week of season.

Once each observation is encoded in this feature space, the team applies k-medoids clustering, also known as Partitioning Around Medoids, using correlation dissimilarity rather than Euclidean or Manhattan distance. Unlike k-means, which represents each cluster with a mean that may not correspond to any real data point, k-medoids selects actual observations as cluster representatives, making the method less sensitive to outliers and fully deterministic, requiring no random seed. The correlation distance, defined as one minus the Pearson correlation between feature vectors, considers two observations similar if their variables vary proportionally, even when their absolute magnitudes differ. In practice, this means two citrus varieties are grouped together when their prices rise and fall at the same times, regardless of whether one sells at twice the price of the other.

Determining the right number of clusters, a classic hyperparameter challenge in unsupervised learning, was handled with multiple lines of evidence. The Silhouette method and the Elbow method both pointed to three clusters, and subsample consensus analysis, in which the clustering was repeated 500 times on random 80 percent subsets of seasons, produced its lowest proportion of ambiguous clustering, about six percent, at exactly that value. Bootstrap resampling with 1,000 repetitions yielded 95 percent confidence intervals for the internal validation indices, while the GAP statistic was treated as non-diagnostic because it favored a single cluster under the 1-SE rule. The convergence of the other criteria on three groups gave the researchers confidence that the structure was genuine rather than an artifact of a single algorithm run.

The choice of distance metric proved decisive. When the correlation-based k-medoids was compared against Euclidean and Manhattan alternatives using standard internal validation indices, the correlation approach dominated: it achieved a Dunn2 index of 1.60 versus 0.62 for Euclidean and 0.56 for Manhattan, and a Calinski-Harabasz score of 393 versus 158 and 204 respectively. An ablation study further showed that neither the five-feature representation nor the correlation distance alone explains the improvement; the gain emerges from their synergy. When raw weekly prices were used instead of features, even dynamic time warping, a sophisticated technique for aligning time series of unequal length, failed to match the combined approach, partly because very short seasons of five or six weeks produce pathological alignments.

The three resulting clusters translate directly into market narratives. The first group, dominated by mandarins and early clementines, is characterized by short seasons, high price volatility and a clear downward drift, with prices falling by a median of 1.4 euro cents per week and a strong linear trend. The second group, populated largely by orange varieties with long marketing windows and prices often agreed in advance, shows remarkable stability: near-zero trend slopes, the lowest price variance and the lowest Q-ratio. The third group is the most erratic, with a median R-squared of only 0.167, indicating that prices swing up and down in ways no linear model can capture, a signature of external shocks such as weather disruptions or sudden demand shifts.

Crucially, the cluster assignments were validated against an independent external criterion derived from official Valencian agricultural sector reports, in which each season was labeled up, flat or down based on weekly price changes. The agreement was moderate but statistically significant, with accuracy of 0.52, a Macro-F1 of 0.52, Cohen’s kappa of 0.28 and a permutation p-value of 0.001, and most disagreements occurred between the flat cluster and its adjacent neighbors, exactly where borderline seasons would be expected. The clustering structure also aligned with documented market events: the 2018-2019 season of overproduction and delayed harvesting pushed most mandarins into the declining cluster, torrential rains in late 2016 preceded a sharp price drop for Navelina oranges, the COVID-19 pandemic in 2020 produced atypical fluctuating patterns as demand for vitamin C surged, and the farmer protests that blocked roads in Castellón in early 2024 drove varieties such as Ortanique and Clemenvilla into the declining cluster in that province while they remained stable in Valencia.

Perhaps the most sobering finding is the sheer instability of cluster membership over time. On average, about 63 percent of variety-province pairs shifted clusters between consecutive seasons, peaking at 70 percent in 2018-2019. Every one of the 111 varieties tracked across all seasons changed groups at least once, which the authors interpret as evidence that agricultural commodity prices are subject to sharp, recurring fluctuations driven by weather, logistics, demand shocks and policy events. This volatility complicates long-term forecasting based on historical prices alone, but it also underscores the value of a monitoring tool that can flag, in near real time, when a variety’s behavior departs from its usual regime.

The researchers emphasize that the framework is deliberately simple and portable. Five features suffice, the algorithm is deterministic, the processed dataset of 528 observations has been made publicly available for reproduction, and the same pipeline could be transferred to other crops, regions or time periods with modest domain-specific tuning, such as defining season boundaries or selecting derived variables. The main limitations are that the model uses only prices at origin, without weather, trade or production-volume data, and that publicly available prices are aggregated by province, limiting microeconomic resolution. Even so, the fact that the clusters independently recovered the fingerprints of floods, a pandemic and road blockades suggests that a shape-based view of price dynamics, built on correlation rather than magnitude, can extract real economic signal from nothing more than weekly price records, providing a reproducible template for market surveillance across the agri-food sector.

Subject of Research: Correlation-based k-medoids clustering of weekly citrus price dynamics in the Comunitat Valenciana, Spain

Article Title: Enhancing unsupervised learning with correlation-based k -medoids: A case study on citrus price dynamics

Article References: Arnau, R., Calabuig, J. M., Ortigosa, N., & Petrosyan, L. (2026). Enhancing unsupervised learning with correlation-based k-medoids: A case study on citrus price dynamics. Machine Learning with Applications, 26, Article 100988. https://doi.org/10.1016/j.mlwa.2026.100988

Image Credits: AI Generated

DOI: 10.1016/j.mlwa.2026.100988

Keywords: unsupervised learning, k-medoids, correlation distance, citrus prices, agricultural economics, clustering, time series, machine learning, Comunitat Valenciana, price volatility, feature engineering, market monitoring