DiMuST: Disentangled Multiplex POI Recommendation
- DiMuST is a next-POI recommendation framework that leverages disentangled shared and private representations from multiplex spatial-temporal graphs to effectively model user mobility.
- It integrates social relationships, spatial transitions, and temporal dynamics to overcome misalignment and redundancy found in traditional fusion approaches.
- Experimental results on NYC and TKY datasets demonstrate improved accuracy and MRR compared to baselines, validating the framework's design and effectiveness.
Searching arXiv for the requested topic and supporting papers. arXiv search query: (Li et al., 11 Aug 2025) DiMuST next POI recommendation DiMuST is a next-POI recommendation framework introduced in “Disentangling Multiplex Spatial-Temporal Transition Graph Representation Learning for Socially Enhanced POI Recommendation” (Li et al., 11 Aug 2025). It is designed for the task of predicting the next Point-of-Interest (POI) from a user’s historical check-in trajectory while jointly modeling social relations, spatial transitions, and temporal transitions. The defining technical claim is that many prior methods model spatial and temporal transitions separately and fuse them afterward, which induces representation misalignment, redundant information during fusion, higher uncertainty, and reduced interpretability; DiMuST addresses this by learning disentangled shared and private representations over multiplex spatial-temporal transition graphs and then combining them with socially enhanced user and POI embeddings (Li et al., 11 Aug 2025).
1. Definition, scope, and nomenclature
DiMuST is presented as a socially enhanced POI recommendation model based on disentangled representation learning over multiplex spatial-temporal transition graphs (Li et al., 11 Aug 2025). The task setting is next POI recommendation: given a user’s historical trajectory of check-ins, the goal is to predict the next POI the user will visit. The formal objects used in the method are a POI
where is the POI index, its name, its category, and its longitude/latitude; a check-in
meaning user visited POI at timestamp ; and a user trajectory
$T_{u_i}=\{c_1,c_2,\dots,c_m\}. \) The framework is motivated by three stated drivers of mobility behavior: spatial transitions, temporal transitions, and social relationships. Spatial transitions capture geographically plausible next places; temporal transitions encode time intervals and temporal patterns; social relationships reflect the influence of close social ties on visitation behavior. The central deficiency targeted by DiMuST is the separate modeling of spatial and temporal transitions followed by post-hoc fusion, which the paper states leads to semantically misaligned embeddings for the same POI/node, duplicated shared information, increased uncertainty, and weaker interpretability [2508.07649]. The name DiMuST should be distinguished from similarly spelled acronyms in other arXiv records. A separate 2025 vision paper is explicitly about **MUSt3R**, not DiMuST, and states that the paper never mentions “DiMuST” anywhere [2503.01661]. Another unrelated 2019 porous-media simulator paper notes that **DuMuX 3** is sometimes misread as “DiMuST” [1909.05052]. In the recommendation literature represented here, however, DiMuST denotes the POI recommendation framework of [2508.07649]. ## 2. Architectural organization and modeling pipeline DiMuST is organized into three major components [2508.07649]. The first is social heterogeneous graph representation learning. The framework builds a socially enhanced heterogeneous graph involving users and POIs, uses an entropy-based social strength model rather than raw explicit friendship links, and learns user and POI embeddings from this graph. The second component is the Disentangled Variational Multiplex Graph Auto-Encoder, abbreviated DAE. Two transition graphs are constructed from global trajectories: a spatial transition graph and a temporal transition graph. For each graph, the encoder produces a shared latent distribution and a private latent distribution. Shared distributions are fused using Product of Experts (PoE), while private features are denoised through contrastive constraints. The fused shared features are then concatenated with pooled private features to obtain the final spatial-temporal transition representation. The third component is the recommendation prediction layer. User embedding, POI embedding, and spatial-temporal transition embedding are fused and passed into a Transformer-based encoder and MLP heads. The model performs two prediction tasks: next POI prediction and next visit period prediction [2508.07649]. The key intuition is the separation between shared and private information. Shared information is the common structure present in both spatial and temporal views; private information is view-specific information unique to geography or time. The paper’s stated purpose is to ensure that the shared representation captures the true joint structure of spatial-temporal transitions while the private representation preserves complementary per-view signals without contaminating the shared subspace. This suggests that DiMuST treats disentanglement not as an auxiliary regularizer but as the organizing principle of representation fusion. ## 3. Graph construction and data representation The framework uses three graph structures with distinct semantics [2508.07649]. The social heterogeneous graph %%%%7%%%% includes three kinds of relations: user-user social strength, user-POI preference or check-in associations, and POI-POI recurrence patterns from global trajectories. The paper defines the user set and POI set as \[ U=\{u_1,u_2,\dots,u_N\}, \qquad P=\{p_1,p_2,\dots,p_M\}.i$0 and $i$1 is defined through an Entropy-based Model: $i$2 where $i$3 is the number of co-occurrences of users $i$4 and $i$5 at location $i$6, $i$7, $i$8 is the diversity order, $i$9 indicates location popularity, and $n$0 and $n$1 are linear functions. User and POI embeddings are then written as
$n$2
The multiplex spatial-temporal component consists of two weighted directed graphs constructed from global trajectories $n$3: the spatial transition graph
$n$4
and the temporal transition graph
$n$5
These graphs model POI-to-POI transitions aggregated over all users. A directed edge from $n$6 to $n$7 exists if $n$8 appear consecutively in a trajectory sequence, and the edge weight depends on how plausible or strong that transition is under either spatial distance or temporal interval.
The adjacency entry $n$9 is defined by a Gaussian-decay-like weighting: $c$0 with
$c$1
For the spatial transition graph, the reported hyperparameters are $c$2, $c$3, $c$4, and $c$5. For the temporal transition graph, they are $c$6, $c$7, $c$8, and $c$9 (Li et al., 11 Aug 2025).
4. Disentangled variational multiplex graph auto-encoding
The DAE is the central technical contribution of DiMuST (Li et al., 11 Aug 2025). Given the two graph views $l=(\lambda,\phi)$0, corresponding to spatial and temporal transition graphs, the encoder with non-shared parameters produces both private and shared latent distributions: $l=(\lambda,\phi)$1 For each view it outputs a shared distribution $l=(\lambda,\phi)$2 and a private distribution $l=(\lambda,\phi)$3, and the paper explicitly states Gaussian sampling: $l=(\lambda,\phi)$4
The shared posteriors from the two views are fused through Product of Experts: $l=(\lambda,\phi)$5 The stated purpose of PoE is to preserve intrinsic spatial-temporal correlation, reduce redundant information in fusion, and avoid simply stacking inconsistent features. The paper describes PoE as principled, simpler, and numerically stable (Li et al., 11 Aug 2025).
The reconstruction objective combines private and shared latents: $l=(\lambda,\phi)$6 The decoder is not given as a separate explicit architecture formula beyond the likelihood term
$l=(\lambda,\phi)$7
Disentanglement is additionally enforced through a correlation-based regularizer between shared and private features. The paper states that if $l=(\lambda,\phi)$8 and $l=(\lambda,\phi)$9 are independent, then
$c = (u,t,p),$0
which motivates the loss
$c = (u,t,p),$1
Here $c = (u,t,p),$2 and $c = (u,t,p),$3 are measurable functions. This is described as a Pearson-correlation-based regularizer; lower correlation implies better disentanglement (Li et al., 11 Aug 2025).
5. Private-feature denoising and prediction layer
A further mechanism of DiMuST is contrastive denoising of private features (Li et al., 11 Aug 2025). The paper argues that private representations contain both useful complementary view-specific information and undesirable noise or spurious edges. Because the transition graphs have no explicit labels, pseudo-labels for node pairs $c = (u,t,p),$4 are inferred using similarity in the shared representation: $c = (u,t,p),$5 where $c = (u,t,p),$6. High-similarity pairs are treated as semantically close and likely same-class; low-similarity pairs are treated as noise. This yields complementary edges $c = (u,t,p),$7 as positives and noisy edges $c = (u,t,p),$8 as negatives.
The private-representation contrastive objective is given as
$c = (u,t,p),$9
where $u$0 is the private representation of node $u$1 in graph $u$2, $u$3 is sampled from complementary edges $u$4, $u$5 is sampled from noisy edges $u$6, $u$7 denotes cosine similarity operation, and $u$8 is a temperature parameter. The stated effect is to pull together private node pairs corresponding to complementary edges and push apart private node pairs corresponding to noisy edges. After denoising, the model average-pools private representations across graphs and concatenates them with shared representations to form the final spatial-temporal transition representation
$u$9
The prediction layer fuses user, POI, and spatial-temporal representations as
$p$0
This fused representation is fed into a Transformer-based encoder-decoder: $p$1 Two MLP heads produce logits for next-POI prediction and next visit period prediction: $p$2
$p$3
Optimization is formulated as classification with cross-entropy losses $p$4 and $p$5. The total objective is
$p$6
with
$p$7
The paper states that there is no separate social-regularization term in the final loss; social information is incorporated structurally through $p$8 and learned end-to-end (Li et al., 11 Aug 2025).
6. Experimental results, ablations, and limitations
DiMuST is evaluated on two benchmark datasets from Foursquare, NYC and TKY, spanning check-in data from April 12, 2012 to February 16, 2013, about 10 months (Li et al., 11 Aug 2025). Daily check-ins of each user are consolidated into trajectories, sparse users and POIs are filtered out, and the split is 80% training, 10% validation, and 10% test. The compared baselines are PRME, ST-RNN, STGCN, PLSPL, STAN, GETNext, AC-TSR, FEARec, DiffuRec, and CrossDR-GEN. The reported metrics are Acc@1, Acc@5, Acc@10, Acc@20, and MRR.
On NYC, DiMuST reports:
- Acc@1 = 0.2584
- Acc@5 = 0.5366
- Acc@10 = 0.6370
- Acc@20 = 0.7375
- MRR = 0.3787
The strongest baseline reported in the table is CrossDR-GEN with:
- Acc@1 = 0.2493
- Acc@5 = 0.5211
- Acc@10 = 0.6289
- Acc@20 = 0.7103
- MRR = 0.3705
On TKY, DiMuST reports:
- Acc@1 = 0.3021
- Acc@5 = 0.5621
- Acc@10 = 0.6645
- Acc@20 = 0.7471
- MRR = 0.4108
The strongest baseline reported in the table is again CrossDR-GEN with:
- Acc@1 = 0.2385
- Acc@5 = 0.5021
- Acc@10 = 0.6003
- Acc@20 = 0.6698
- MRR = 0.3570
The authors summarize that on NYC, DiMuST achieves an average 2.55% improvement over the strongest baseline CrossDR-GEN, and an 11.70% gain over recent state-of-the-art methods on key metrics. The gains on TKY are described as especially large (Li et al., 11 Aug 2025).
The ablation study evaluates w/o HRL, w/o DAE, w/o spatial, w/o temporal, and w/o PoE. The paper reports that both the social module and DAE are important, and that DAE has the largest effect. Removing DAE reduces performance notably, for example Acc@5 by 11.09% and Acc@10 by 7.27%. Using only spatial or only temporal information also hurts strongly, which the paper takes as evidence that the two are inherently intertwined (Li et al., 11 Aug 2025).
Sensitivity analysis varies 9, 0, 1, and 2. The reported finding is that proper 3 and 4 are crucial for good disentanglement and denoising, while 5 and 6 strongly affect Acc@10 and Acc@20 because reconstruction dominates latent-space quality in the VGAE backbone. A disentanglement analysis on NYC shows low cross-correlation in the off-diagonal blocks of the correlation map between shared and private representations. A social-strength case study visualizes user pairs with high versus low social strength and shows that high-strength pairs have more overlapping and clustered POI visits, supporting the validity of the EBM-based social strength measure (Li et al., 11 Aug 2025).
The paper claims three main contributions: a new POI recommendation framework integrating social collaboration, user-POI interactions, and spatial-temporal transition patterns; the first introduction of a disentangled variational graph autoencoder into POI recommendation for multiplex spatial-temporal graphs, separating shared and private latent distributions and fusing them adaptively; and strong empirical gains on real datasets (Li et al., 11 Aug 2025). At the same time, it does not provide a dedicated limitations section. The stated assumptions and open questions include the conditional independence or PoE assumption for shared-posterior fusion, dependence of pseudo-label quality on shared-representation similarity, fixed graph-construction hyperparameters 7, and the simplification of the social heterogeneous graph by treating heterogeneous nodes and edges as homogeneous during embedding. A plausible implication is that DiMuST’s empirical behavior is tightly coupled to the quality of graph construction and the separation between shared and private latent structure rather than to sequential modeling alone.