The core problem, as the authors describe it, is that conventional recommendation systems treat a user’s trajectory as a simple sequence of visits and then model only the consecutive correlations between neighboring check-ins. That strategy is vulnerable to what the team calls time bias. Real people deviate from their routines: a commuter catches a later train, a lunch appointment shifts by forty minutes, an errand is skipped one week and doubled up the next. Sequence models that latch onto exact consecutive orderings are easily corrupted by these deviations and by random noise in the data. Worse, most methods ignore periodicity altogether or rely on pre-specified intervals and fixed ranges, hard-coded assumptions that prevent the model from learning the true rhythm of a user’s life adaptively. The researchers also identified a second blind spot: geography. Existing systems tend to reason about spatial proximity at a single, fine-grained scale, even though human mobility is organized hierarchically across neighborhoods, cities, and regions.

Before building their model, the team examined real-world data from Foursquare and found striking statistical evidence for both kinds of regularity. Users’ trajectories showed multi-span temporal patterns, meaning activity tendencies differ across periods such as weekdays versus rest days, and multi-interval patterns, in which a person returns to the same location after a roughly fixed gap, such as every day or every four hours. Spatially, different users roam across regions of very different extent, from a few city blocks to entire metropolitan areas. These observations framed the design brief for MultiPerG: the model needed to discover periodic patterns at multiple intervals and spans without being told what they are, tolerate the small biases that come with human behavior, and read geographic structure at multiple scales simultaneously.

Architecturally, MultiPerG is a hierarchical framework built on the foundations of temporal convolutional networks, or TCNs, a family of sequence models that use causal convolutions, which prevent information from the future leaking into the past, and dilated convolutions, which deliberately skip positions to widen the network’s receptive field. The researchers transform a user’s check-in history into a sequence feature map, a matrix analogous to an image in computer vision, by segmenting the week into equal one-hour gaps over seven days. Candidate locations are likewise converted into a geography feature map by gridding GPS coordinates into equal-area units, and this map is built hierarchically so that each level represents a coarser spatial granularity. A codebook stores the coordinates of every location across all scale levels, allowing the network to look up where any given POI sits within the multi-scale geographic structure.

The heart of the innovation lies in three novel kinds of deformable blocks. The first two handle time. A multi-interval deformable tree layer arranges dilated causal convolutions into a binary-tree structure, with left and right branches preserving the outermost elements of each segment. The convolutions carry learnable deformable offsets, a technique borrowed from deformable convolutional networks in computer vision, which let the sampling positions of each kernel float slightly rather than sitting at fixed intervals. In practice, these offsets absorb the consecutive time bias and accidental fluctuations produced by deviant visits, so the model can still detect that a user returns to a place roughly every day even when the actual visits drift by an hour or two. A second, segment-tree-structured layer works in the complementary direction, combining adjacent segments of the sequence into progressively larger spans so that patterns like day-long or month-long routines emerge adaptively. The outputs of the interval and span layers are then multiplied together to form a joint periodic representation of the user’s habits.

The third block addresses space. The geo-span learning layer stacks two-dimensional dilated causal convolutions across the hierarchical levels of the geography feature map, capturing spatial dependence at scales ranging from fine urban units up to broad regions. The authors note a subtle problem with this design: when locations are discretized onto maps, noise arises because two places that fall into the same unit at one scale may belong to distinct sections at a finer scale. To suppress this, they introduce a custom non-linear activation function called tan, described as the counterpart of tanh in the complex domain via Euler’s formula, whose steep gradient behavior approaching infinity amplifies large-scale influence while diminishing small-scale noise. After amplification and normalization, the outputs of all geographic levels are summed with adjusted weights, and the learned geographic patterns are combined with the periodic sequence patterns to produce, through a query embedding, a probability for every candidate location.

Training uses a cross-entropy loss summed over users, with randomly selected negative samples, and the network is optimized with Adam and mini-batch stochastic gradient descent. The team evaluated MultiPerG against eleven state-of-the-art baselines spanning four methodological families: CNN-based models such as Caser, CosRec, NextItNet and NASR; attention-based models including TiSASRec, STAN and ASPPA; the graph-neural-network-based CyGNet; and RNN-based models LSTPM and PLSPL. Experiments ran ten times on NVIDIA V100 GPUs with baselines fine-tuned for fairness, using four datasets: Foursquare check-ins from New York City and Tokyo, Yelp data from Tucson, and the full Gowalla dataset, filtered to remove inactive users and locations.

The results were decisive. MultiPerG outperformed every baseline on every recall metric at cutoffs of five, ten and twenty recommendations. On the New York and Tokyo datasets it improved Recall@5 by 5.5 and 6.4 percent respectively over the strongest competitor, the LSTM-based PLSPL, and on Gowalla the gains were larger still, at 10.5 percent for Recall@5 and 11.8 percent for Recall@10. The comparison also yielded interesting secondary findings: recurrent models excelled at personalized accuracy but lagged on generalized performance, while models built on temporal convolutional networks with large receptive fields, such as NASR and NextItNet, fared best among the CNN family. Ablation studies confirmed that every module earns its place. Removing the geo-span layer hurt most, underscoring how much geography contributes to next-POI prediction, while removing the interval layer cost slightly more than removing the span layer, suggesting interval periodicity is marginally more influential. Embedding dimensions of roughly 50 to 60 proved sufficient, and a modest geography map size of 128 units performed well.

Perhaps most compelling are the interpretability analyses. Heatmaps of the interval and span layer outputs for 32 New York users over 72 hours revealed a clear four-hour interval pattern during daytime hours, stronger relevance among midnight spans, and an overarching 24-hour daily repetition in both, matching the expected rhythm of human activity. The deformable offsets visibly corrected time bias in the trajectories. Geographic heatmaps, meanwhile, showed clustered high-relevance zones that align with users’ real activity centers: in the New York data the hot zones match the densest check-in areas, and in Gowalla they correspond to Texas and California, the two states contributing the most data. The authors conclude that MultiPerG’s advantage stems precisely from its ability to explore periodic patterns, and they point toward future work on more variations of the geography blocks and more generalized periodicity recovery. For now, the study offers a vivid demonstration that the key to predicting where someone will go next lies not just in where they have been, but in the nested rhythms of time and space that shape every journey.

Subject of Research: Next point-of-interest recommendation using periodic and geographic pattern learning in human trajectory data

Article Title: MultiPerG: Multiple Periodic Geography convolution for next POI recommendation

Article References: Wang, X., Wang, B., Wang, D., Wang, Z., Sun, G., Zhao, B., & Chen, M. (2025). MultiPerG: Multiple Periodic Geography convolution for next POI recommendation. Vicinagearth, 2(1), Article 3. https://doi.org/10.1007/s44336-025-00012-1

Image Credits: AI Generated

DOI: 10.1007/s44336-025-00012-1

Keywords: POI recommendation, human mobility, temporal convolutional network, periodic patterns, geographic patterns, deformable convolution, location-based social networks, sequence modeling, deep learning, trajectory mining, Foursquare, Gowalla

The problem the researchers tackle is familiar to anyone whose phone has suggested a restaurant or a gym. Location-based social services generate enormous streams of check-in data, and turning those streams into accurate next-destination predictions has become a vibrant research area. Deep learning models have steadily improved performance by learning users’ interests and preferences, but the Chongqing team identified two persistent weaknesses in existing approaches. First, many methods built on recurrent neural networks (RNNs) process visits one after another and therefore overlook relationships between a user’s preferences for non-consecutive POI visits. A person might visit a coffee shop every Monday and a bookstore every other Saturday, but a sequential model that only sees immediate neighbors in the sequence can miss those long-range connections, leaving it unable to correctly understand the motivations behind user behavior.

The second weakness concerns short-term interests. Existing methods for analyzing what a user is doing right now, the authors argue, lack an in-depth exploration of both the overall structure and the key details of recent behavior. They fail to effectively combine a global perspective with an analysis of local critical points, the decisive moments that drive what happens next. This limitation produces an inaccurate understanding of users’ short-term interests, fails to capture dynamic changes and key driving factors comprehensively, and ultimately damages the accuracy and adaptability of the model. In practical terms, a recommender that cannot see both the forest and the trees will suggest the wrong place at the wrong time.

STFDN addresses these issues with a two-pronged architecture. The first prong is a residual block-enhanced attention architecture designed to capture the global view of both long-term and short-term user preferences. Residual blocks, a technique popularized in deep computer vision, allow information to bypass layers of processing so that early signals are not degraded as the network deepens. Combined with attention mechanisms, which let the model weigh the importance of different past visits when forming its prediction, this architecture directly tackles the issue of non-consecutive POI preferences. A visit from three weeks ago can exert just as much influence on today’s recommendation as the one from an hour ago, provided the attention mechanism learns that it matters.

The second, and more novel, prong moves the analysis into the frequency domain. The researchers designed a Fre-sequence module and a Fre-temporal module to identify dependencies between what they call energy concentration points, which they describe as analogous to dominant frequencies in signal processing. In ordinary language, these are the salient periodic patterns in a user’s behavior: the weekly coffee ritual, the monthly shopping trip, the after-work gym session. By transforming sequences of visits from the time domain, where events follow one another, into the frequency domain, where periodicities stand out as peaks of concentrated energy, the modules can spot these rhythms and model the dependencies between them. This enables an in-depth exploration of users’ short-term interests in the frequency domain, pairing the global structural view with a fine-grained reading of the critical points that actually drive movement.

The idea of analyzing sequential data through its frequency content has been gaining traction in adjacent fields. The authors draw on recent work showing that frequency-domain multilayer perceptrons can be more effective learners in time-series forecasting, suggesting that human mobility, like stock prices or weather, carries structure that is easier to read once the noise of raw chronology is stripped away. Applying that insight to POI recommendation is a distinctive move, because mobility data is not merely a time series. Each visit carries spatial coordinates, temporal context and semantic content, and a useful model must fuse all three. STFDN’s spatiotemporal framing, reflected in its name, is intended to do exactly that, layering frequency-domain analysis on top of representations that respect where and when each check-in occurred.

To test the approach, the team ran extensive experiments on two benchmark datasets, comparing STFDN against a crowded field of competitors that includes recurrent models, attention-based architectures and spatio-temporal gated networks. The reported result is straightforward: the model outperforms other state-of-the-art algorithms. The authors’ contribution statement indicates that Huang led conceptualization, methodology, software, visualization, validation, data curation and the original draft, while Zhang supervised the work and contributed formal analysis, review and editing, with He, Li and Li handling validation, resources, investigation and formal analysis. The work was supported by the National Natural Science Foundation of China and the Natural Science Foundation Project of Chongqing.

The lineage of this research stretches back more than a decade. Early POI recommenders relied on matrix factorization enriched with geographical and social influence, exploiting the observation that people tend to visit places near places they already like. Successive generations added sequential modeling, starting with factorizing personalized Markov chains and moving through attentional recurrent networks such as DeepMove, contextual attention architectures, and temporal and multi-level context attention models. More recently, spatio-temporal attention networks and transformer-style recommenders have mined relationships between visited locations directly. Each generation has chipped away at the same core question: how much of a person’s future is explained by their last stop, and how much by the deeper pattern of their life?

STFDN’s answer is that both matter, but in different ways, and that a single model can hold both views simultaneously. The residual-enhanced attention branch preserves the long arcs of preference that stretch across weeks, while the frequency modules compress recent activity into its most meaningful periodic components and let the model reason about them explicitly. This division of labor is what the authors say was missing from prior short-term interest analyses, which tended to treat recent history as either an undifferentiated blob or a fine-grained but myopic list. The frequency-domain lens offers a middle path, keeping the details that repeat and discarding the incidental noise.

The implications extend beyond restaurant suggestions. Accurate next-location prediction underpins urban planning, traffic management, epidemic modeling and personalized travel services, and any method that better captures the periodic structure of human movement could feed those applications. The field also continues to broaden its toolkit, with neighboring work by some of the same authors exploring latent diffusion models for social recommendation, smooth diffusion models for multimodal recommendation, and hierarchy-aware diffusion with knowledge-enhanced contrastive learning. For now, STFDN stands as a demonstration that sometimes the best way to understand where people are going is to stop watching their footsteps and start listening to the rhythm underneath them. The study, published as volume 68, article 262 of Knowledge and Information Systems, was received in April 2025, revised in June 2026 and published on 22 September 2026.

Subject of Research: Spatiotemporal frequency-domain deep learning for next point-of-interest recommendation

Article Title: Spatiotemporal frequency-domain network for next POI recommendation

Article References: Spatiotemporal frequency-domain network for next POI recommendation. (n.d.). https://doi.org/10.1007/s10115-026-02883-2

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02883-2

Keywords: POI recommendation, frequency domain, recommender systems, deep learning, attention mechanism, user preferences, human mobility, spatiotemporal modeling, recurrent neural networks, short-term interests, check-in data, Knowledge and Information Systems