STIM: Spatio-Temporal Periodic Interest Modeling

• STIM is a spatio-temporal modeling framework that captures recurring user behavior through grouped review points and forgetting-curve masking.
• It dynamically weights historical interactions based on temporal and spatial context to improve local-life ranking and conversion prediction.
• The architecture integrates adaptive query MoE and hierarchical multi-interest extraction to aggregate sparse, irregular user histories effectively.

Searching arXiv for recent and foundational papers on STIM and closely related spatio-temporal periodic modeling for POI recommendation. I’m checking arXiv for papers on “spatio-temporal periodic interest modeling”, “next POI recommendation periodic interest”, and related hierarchical or center-based spatio-temporal models. Spatio-Temporal Periodic Interest Modeling (STIM) denotes a class of recommendation and mobility models that treat user behavior as jointly conditioned by time, location, and recurrence, and also refers specifically to a long-sequence local-life recommendation architecture that operationalizes this idea through forgetting-curve-based masking, spatio-temporal grouping, adaptive expert queries, and hierarchical multi-interest extraction. In the local-life setting, STIM is designed for ranking tasks such as click-through rate and conversion prediction, where user histories are long, sparse, irregular, and strongly coupled to request-time and request-location context. Within the broader literature, closely related systems range from explicit periodic next-POI models based on day, hour, and area matching to hierarchical or center-based models that recover recurrent behavior more indirectly (Hu et al., 4 Aug 2025).

1. Conceptual foundations

The STIM formulation for local-life service recommendation starts from the ranking problem

$$\hat{y} = g\left(f\left(B; \Theta_f\right),C; \Theta_g\right),$$

where $B$ is the user’s historical behavior sequence, $f$ is the behavior-sequence encoder, $C$ denotes additional context and profile features, and $g$ is the prediction head. Its motivating claim is that local-life recommendation differs from standard e-commerce or content recommendation in two specific ways: long behavior sequences are sparse and irregular, and user behavior exhibits strong spatio-temporal dependence. A user’s previous behavior may be as recent as yesterday or as old as a year, while local-life demand is shaped by time, location, and item or service type jointly (Hu et al., 4 Aug 2025).

The model is framed through an analogy to human memory. The paper explicitly links recommendation behavior to a forgetting process, the recency effect, and cyclicality of memory. In this view, recent interactions should dominate current intent, but repeated interactions under similar temporal or spatial conditions may signal stabilized habits rather than noise. STIM therefore treats periodic interest not as abstract seasonality alone, but as recurring consumption tendency under similar time and location contexts (Hu et al., 4 Aug 2025).

Related work shows that this general idea has multiple operationalizations. ST-PIL explicitly models daily, hourly, areal, and hourly-areal periodic interest for next POI recommendation (Cui et al., 2021). STACP treats recurring behavior as temporal-state-conditioned movement around different activity centers, using working time and leisure time as discrete states rather than an explicit cyclic latent model (Rahmani et al., 2020). STAR-HiT, by contrast, is not an explicit periodic-interest model, but its hierarchical spatio-temporal subsequence discovery can reveal daily and weekly regularity in user trajectories (Xie et al., 2022).

2. Periodic interest as grouped recurrence

In STIM, periodicity is recovered through grouped “review points.” Historical behaviors are assigned to coarse temporal and spatial groups, and a current request is matched against those groups to determine which past interactions should be treated as periodic anchors. The temporal groups are an hourly partition with morning $3{:}00$–$10{:}00$, midday $11{:}00$–$16{:}00$, and night $17{:}00$–$B$0, together with a weekly partition into weekday and weekend. The spatial grouping maps behavior locations to geohash6 and then splits them into three groups based on the first two geohash characters. Historical events falling into the same hour, week, or geohash group as the current request are treated as review points (Hu et al., 4 Aug 2025).

The base retention function is the exponential Ebbinghaus curve

$B$1

where $B$2 is memory retention and $B$3 is the characteristic time constant. STIM then reinterprets review effects for recommendation. The paper explicitly states that recommendation scenarios are opposite to classic memory review: review points farther from the current request should receive lower post-review retention, and the farther they are, the faster interest should decay. The resulting mask is obtained by mapping time to the compressed sequence and applying normalization,

$B$4

with separate hour, week, and geohash masks later remapped by a one-layer MLP into $B$5 (Hu et al., 4 Aug 2025).

This construction lets the same mechanism emphasize both recent and periodic spatio-temporal features. Recent behaviors retain the largest weights because decay starts from the behavior nearest to request time. Periodic behaviors receive local increases in weight whenever a historical action matches the request’s hour, week, or geohash grouping, but the review-induced rise does not return to the original peak, preserving the model’s recency bias. The result is a soft continuous weighting scheme rather than a hard selector: events adjacent to matched review points can also retain appreciable weight because nearby actions may causally influence the matched behavior (Hu et al., 4 Aug 2025).

A closely related but more explicit periodic retrieval design appears in ST-PIL. There, periodic interest is defined by reorganizing history according to day of week, similar hour slots, current area, and their intersection, so that periodicity becomes a context-conditioned history filtering problem rather than a pure recency problem (Cui et al., 2021).

3. Architecture of the STIM system

The STIM pipeline begins with long-sequence compression. A target-relevant subsequence $B$6 is extracted by a General Search Unit inspired by SIM, after which masking and interest extraction operate on the compressed sequence rather than the full history. The current request’s time and location are used to generate the dynamic masks described above, while the request context also drives a query-generation stack called Query MoE (Hu et al., 4 Aug 2025).

Query MoE defines three base experts,

$B$7

and further decomposes the time expert into hour and week sub-experts,

$B$8

Request-conditioned gates are then computed as

$B$9

with a holiday-enhanced week weight

$f$0

The hour and week branches are combined into

$f$1

and the main query output is

$f$2

The model also defines additional pairwise-combination query outputs $f$3 to $f$4, intended to capture explicit cross-feature interactions such as temporal-item and spatial-item coupling, although the exact pairwise formulas are not fully specified in the paper (Hu et al., 4 Aug 2025).

Interest extraction is performed by the Hierarchical Multi-Interest Network Unit. Given key sequence $f$5, query $f$6, and mask $f$7, the shallow interaction stage computes

$f$8

followed by masked query-key matching,

$f$9

The deep interaction stage converts masked scores into head-specific weights,

$C$0

and aggregates transformed key features as

$C$1

The resulting spatio-temporal multi-interest features are concatenated with other context $C$2 and passed to a prediction head such as DNN or MMoE. Training uses binary cross-entropy,

$C$3

This makes periodic interest operationally equivalent to the historical signal that survives both request-conditioned matching and forgetting-curve weighting (Hu et al., 4 Aug 2025).

4. Position within the periodic POI and mobility literature

STIM belongs to a broader family of spatio-temporal periodic models, but its inductive bias is distinctive: it uses grouped review points and a recommendation-specific forgetting process rather than only embeddings or only recurrent retrieval. The nearest explicit predecessor in the provided literature is ST-PIL, which builds one long-term daily periodic interest and four short-term interests—recent sequential, areal, hourly, and hourly-areal—then fuses them through inter-level attention (Cui et al., 2021).

Model	Periodic mechanism	Characterization
STIM	Forgetting-curve masks + grouped review points + Query MoE	Target-conditioned local-life ranking
ST-PIL	Weekday prototypes + hourly/areal/hourly-areal sequence retrieval	Explicit multi-granularity periodic interest
STACP	Working-time/leisure-time activity centers	Coarse recurring spatio-temporal preference
STAR-HiT	Hierarchical subsequence discovery with spatio-temporal context	Indirect periodicity only
GSTM-HMU	Fourier periods $C$4h/$C$5d/$C$6d + CTM + LCB	Periodicity-aware generative mobility model

ST-PIL models periodicity by reorganizing history around matching contexts. Its long-term module compresses all history into seven weekday-specific prototypes, and its short-term module constructs four LSTM-encoded sequences $C$7 corresponding to recent, same-area, similar-hour, and same-area-plus-similar-hour histories. This is a direct multi-granularity periodic-interest design in which periodicity is represented through context-aligned retrieval rather than through decay-based masking (Cui et al., 2021).

STACP offers a weaker but still relevant periodic interpretation. It assumes users move around different activity centers in different temporal states and defines a center-based contextual score over a user’s temporal-state-specific centers $C$8, with the practical implementation using working time and leisure time. It is therefore better described as context-conditioned spatio-temporal preference modeling over recurring temporal partitions than as a full periodic dynamics model (Rahmani et al., 2020).

STAR-HiT occupies a different position. It is a spatio-temporal hierarchical sequence model for next POI recommendation whose core mechanism is adaptive subsequence partitioning and recursive aggregation. It does not define an explicit periodicity component such as day-of-week or hour-of-day recurrence modules, but its case study reports discovery of daily regularity and weekly regularity. Its relevance to STIM lies in the idea that repeated routines may be better represented as adaptively discovered semantic subsequences than as raw POI repetitions (Xie et al., 2022).

A more recent mobility model, GSTM-HMU, incorporates periodicity more explicitly through a Fourier feature bank

$C$9

and integrates these rhythms with location, category, cell structure, continuous-time memory, and long-term lifestyle prompts. This suggests a different design axis from STIM: explicit cyclic phase encoding rather than grouped review-point masking (Luo et al., 23 Sep 2025).

5. Broader methodological analogues

Outside recommendation proper, localized spatio-temporal modeling has treated periodicity as a tensor-product basis problem. A streaming prediction framework for scalar spatio-temporal processes defines

$g$0

where $g$1 is a temporal periodic basis and $g$2 is a localized spatial basis with finite support. This allows different spatial regions to express different temporal periodicities without requiring globally stationary behavior. Although it is a scalar regression model rather than a ranking model, the underlying principle is closely aligned with STIM: periodic behavior is not globally uniform, but spatially conditioned (Osama et al., 2018).

A second analogue appears in large-scale spatio-temporal forecasting. UltraSTF assumes a user-specified period length $g$3 and separates periodic structure into cross-period dynamics and intra-period shape variation. The cross-period module forecasts aligned phases across cycles, while the shape-bank component stores a small learned dictionary of period-length motifs and reconstructs within-period adjustments through attention-like activation. This does not model recommendation interest directly, but it makes a useful distinction between recurrence across cycles and structure within a cycle, a distinction that is also implicit in STIM’s separation between review-point recurrence and hierarchical multi-interest extraction (Yeh et al., 28 Feb 2025).

These broader models support two technical readings of periodic interest. One reading treats periodicity as explicit cyclic basis encoding or harmonic structure. Another treats it as the retrieval, masking, or aggregation of historically recurrent contexts. STIM belongs chiefly to the second group, whereas GSTM-HMU and the localized covariance model give stronger examples of the first (Luo et al., 23 Sep 2025).

6. Empirical evidence and deployment

The STIM paper reports results on two industrial-scale datasets. On Ele.me, with 14 million users, 7 million items, maximum sequence length 50, and 146 million samples over 8 days, STIM attains CTR AUC $g$4 and CTR GAUC $g$5. On TRec, with 786 million users, 162 million items, maximum sequence length 20,000, and timespan 365, STIM reports CTR AUC $g$6, CTR GAUC $g$7, CTCVR AUC $g$8, and CTCVR GAUC $g$9. Compared with the sub-optimal model on TRec, the reported gains are $3{:}00$0 CTR AUC, $3{:}00$1 CTR GAUC, $3{:}00$2 CTCVR AUC, and $3{:}00$3 CTCVR GAUC, all statistically significant with paired $3{:}00$4-test at $3{:}00$5. The online A/B test reports a statistically significant $3{:}00$6 improvement in gross transaction volume, and the method is described as deployed in a large-scale system serving hundreds of millions of daily active users (Hu et al., 4 Aug 2025).

The ablation studies are also central to the STIM identity. Among forgetting-curve forms, exponential outperforms power and logarithmic variants. The best reported hyperparameters on TRec are $3{:}00$7, $3{:}00$8, $3{:}00$9, and $10{:}00$0, with $10{:}00$1 having the strongest effect. In the dynamic masking ablation, review points are identified as the most important enhancement; using same-group review points is better than using identical-value review points, and the one-layer MLP remapping is also beneficial. Query MoE ablations show that four-expert weighting, holiday enhancement, and pairwise expert combinations each improve performance, while the HMIN-Unit slightly but consistently exceeds ordinary multi-head attention on TRec (Hu et al., 4 Aug 2025).

The broader periodic-interest literature provides corroborating evidence. ST-PIL reports state-of-the-art next-POI performance on NYC and TKY, with full-model Acc@1 reaching $10{:}00$2 on NYC and $10{:}00$3 on TKY, and the short-term periodic interests based on area and hour-area alignment proving especially influential (Cui et al., 2021). STACP shows statistically significant gains over prior geographical and temporal baselines on Gowalla and Foursquare, and its temporalized-center ablation supports the claim that users exhibit spatio-temporally centered behavior rather than merely geographically centered behavior (Rahmani et al., 2020). STAR-HiT reports superior performance on Foursquare NYC, Foursquare US, and Gowalla, with its strongest evidence for periodic relevance coming from case studies that separate daily and weekly regularity in learned subsequences (Xie et al., 2022).

7. Interpretive significance, limitations, and recurrent misconceptions

A common misconception is that spatio-temporal periodic interest modeling is equivalent to adding time embeddings to a recommender. The literature provided does not support that simplification. STIM combines grouped recurrence priors, forgetting-curve-based soft masks, adaptive expert queries, and hierarchical multi-interest extraction; ST-PIL reorganizes historical logs into multiple context-matched periodic subsequences; and GSTM-HMU couples cyclic time with continuous-time memory and semantic preference prompts. In each case, periodicity is tied to structured history selection or state evolution rather than to timestamp encoding alone (Hu et al., 4 Aug 2025).

A second misconception is that periodic interest is necessarily explicit. ST-PIL and STIM are explicit in different senses, but STAR-HiT shows that periodic-like behavior can also emerge indirectly from spatio-temporal hierarchical sequence modeling. The paper explicitly states that STAR-HiT is not a true explicit periodic-interest model, even though it discovers daily and weekly regularity. This distinction matters because a model may recover routines without representing cycle phase directly (Xie et al., 2022).

The principal limitations of STIM are also clearly delimited in the source material. Its temporal and geohash groupings are heuristic and domain-specific; periodicity is coarse, being captured through hour, week, and geohash group matching rather than continuous alignment or learned periodic kernels; several formulas in the paper are partially under-specified or typeset incorrectly; and the exact definitions of pairwise expert combinations $10{:}00$4 are not given. The paper also does not provide explicit latency or serving-cost measurements, despite emphasizing industrial deployment (Hu et al., 4 Aug 2025).

Across the broader field, related limitations recur in different forms. ST-PIL relies on handcrafted periodic sequence construction, STACP depends on manually defined temporal states, STAR-HiT lacks explicit cycle representation, and GSTM-HMU does not define explicit periodic interest states even though it uses cyclic Fourier features. A plausible implication is that future STIM systems will increasingly combine explicit periodic phase encoding with adaptive retrieval or hierarchical subsequence discovery, joining the stronger parts of these currently separate design traditions (Cui et al., 2021).