Source-linked AI summary
Discovering Collaborative Signals for Next POI Recommendation with Iterative Seq2Graph Augmentation
Yang Li, Tong Chen, Yadan Luo, Hongzhi Yin, Zi Huang
TL;DR
Next-POI recommenders often model check-ins as isolated sequences, overlooking cross-sequence collaborative signals amid sparse POI transitions. SGRec augments sequences with graph context and category-aware learning; it outperforms state-of-the-art methods on two real-world datasets.
Problem
Existing sequential recommenders overlook correlated POI signals across sequences, while sparse POI-POI transitions constrain learned representations.
Source: §1 · pp. 1–2Method
SGRec augments POI sequences with graph context, learns POI embeddings and temporal preferences, and uses a category-transition objective to address sparsity.
Source: §1 · p. 2Results
SGRec significantly outperforms all baseline methods on both Foursquare and Gowalla; on Gowalla, it improves over ST-CLSTM by 4.15% on HR@5 and 7.36% on nDCG@5.
Source: §1 · p. 2, §5.4 · p. 5Takeaways & Limitations
The authors conclude that SGRec is effective and robust, demonstrating strong real-life practicality.
Source: §6 · p. 6
Abstract
from arXiv · showhide
Being an indispensable component in location-based social networks, next point-of-interest (POI) recommendation recommends users unexplored POIs based on their recent visiting histories. However, existing work mainly models check-in data as isolated POI sequences, neglecting the crucial collaborative signals from cross-sequence check-in information. Furthermore, the sparse POI-POI transitions restrict the ability of a model to learn effective sequential patterns for recommendation. In this paper, we propose Sequence-to-Graph (Seq2Graph) augmentation for each POI sequence, allowing collaborative signals to be propagated from correlated POIs belonging to other sequences. We then devise a novel Sequence-to-Graph POI Recommender (SGRec), which jointly learns POI embeddings and infers a user's temporal preferences from the graph-augmented POI sequence. To overcome the sparsity of POI-level interactions, we further infuse category-awareness into SGRec with a multi-task learning scheme that captures the denser category-wise transitions. As such, SGRec makes full use of the collaborative signals for learning expressive POI representations, and also comprehensively uncovers multi-level sequential patterns for user preference modelling. Extensive experiments on two real-world datasets demonstrate the superiority of SGRec against state-of-the-art methods in next POI recommendation.
1 Introduction
Existing sequential next-POI recommenders learn from individual check-in sequences, missing collaborative signals across sequences and facing sparse POI transitions. SGRec augments sequences with graph context and category-aware learning, and outperforms state-of-the-art baselines on two benchmark datasets.
1 Introduction
- Most sequential methods condition recommendations on POIs in one sequence, missing correlated POI contexts across sequences and higher-order collaborative signals.Sparse POI-POI transitions further constrain learned representations and recommendation effectiveness.Source: §1 · pp. 1–2
- Seq2Graph augmentation adds semantic neighbor nodes to chronologically ordered POIs, incorporating correlated-POI context while retaining sequential dependencies.The added context can include correlated POIs and category tags.Source: §1 · p. 2
- SGRec combines neighbor-node context with position-aware attention and a parallel category-transition objective to learn POI embeddings and temporal preferences.The category objective draws on denser category-wise interactions to address POI-level sparsity.Source: §1 · p. 2
- SGRec outperforms state-of-the-art baselines with significant margins on two benchmark datasets.The paper highlights this result as part of its evaluation contribution.Source: §1 · p. 2
2 Related Work
Prior next-POI work models check-ins with Markov chains, matrix factorization, and recurrent networks to capture transitions and dynamic preferences. This paper instead transforms sequences into graphs to model cross-sequence POI correlations.
2 Related Work
- Markov-chain, matrix-factorization, and recurrent-network approaches model user-POI interactions, transitions, and dynamic preferences from check-ins.Examples include time- and distance-specific transition matrices, attention over contextual factors, and time and distance gates.Source: §2 · p. 2
- Earlier work combines first-order Markov chains with matrix factorization to model POI transitions and user-POI interactions.The related-work discussion also describes later RNN approaches to sequential patterns.Source: §2 · p. 2
- Unlike prior RNN-focused sequence mining, the proposed graph-structured sequences let a GNN-based model explore collaborative information across POI sequences.Prior work is described as not explicitly considering these cross-sequence POI correlations.Source: §2 · p. 2
3 Preliminaries
The preliminaries define POIs, check-ins, and chronologically ordered check-in sequences. Next-POI recommendation is framed as recommending K POIs a user is likely to visit next, given a check-in sequence.
3 Preliminaries
- A check-in is a user, POI, category, and timestamp tuple; each POI has an associated category label.A POI denotes a specific location, such as a cinema.Source: §3 · p. 2
- A check-in sequence consists of chronologically ordered records for one user within a short time interval, such as 24 hours.The passage identifies the final record as the user's current location.Source: §3 · p. 2
- Given a user's check-in sequence, next-POI recommendation aims to recommend K POIs likely to be visited at the user's next move.Source: §3 · p. 3
4 Methodology
SGRec augments check-in sequences with cross-sequence POI neighbors, then combines category-aware graph aggregation, temporal preference encoding, and joint POI-category prediction. Its training objective optimizes both prediction tasks.
4.1 Seq2Graph Augmentation
- Each sequence gains incoming first-order POI neighbors from the user’s other sequences, with a sampled subset refreshed each training epoch.The augmentation retains chronological sequence edges and includes each POI among its own neighbors.Source: §4.1 · p. 3
4.2 Category-Aware Graph Attention Layer
- Category-aware graph attention conditions neighbor aggregation on category-pair edge relations, producing POI embeddings that encode neighbor POIs and category information.SGRec also incorporates category-wise dependencies and spatial factors during aggregation.Source: §4.2 · p. 3
4.3 User Temporal Preference Encoding
- Position-aware attention estimates each check-in’s influence on the last POI using relative position, then forms a sequence-level short-term preference embedding.The model assigns position embeddings according to each check-in’s offset from the last one.Source: §4.3 · p. 4
4.4 Next POI and Category Prediction
- SGRec jointly predicts the next POI and its category, using category transitions as an auxiliary task while performing only POI prediction at inference.The category task targets denser category-wise transitions than POI-level interactions.Source: §4.4 · p. 4
4.5 Model Optimisation
- Training applies cross-entropy to both prediction tasks, with a weighted category loss intended to connect the upcoming category with user preferences under data sparsity.The category-loss weight is controlled by η.Source: §4.5 · p. 5
5 Experiments
SGRec outperforms the compared baselines on both datasets, with especially strong results on sparse Gowalla; ablations and hyperparameter analyses examine its components and settings.
5.1 Datasets and Evaluation Metrics
- Experiments use Foursquare and Gowalla check-in sequences, filtered for active users and popular POIs, split 80/10/10 by user, and capped at 20 check-ins.Sequences are segmented at gaps longer than one day for Gowalla and three days for Foursquare; sequences shorter than 10 check-ins are discarded.Source: §5.1 · p. 5
- HR@K measures top-K performance and nDCG@K overall ranking; the last check-in is the target, and Seq2Graph augmentation is not applied at test time.The remaining check-ins in each test sequence are fed to the model for prediction.Source: §5.1 · p. 5
5.2 Baseline Methods
- Comparisons include FPMC-LR, GRU, TMCA, ST-CLSTM, and SR-GNN, representing Markov/matrix-factorization, recurrent, and graph-based recommendation approaches.SR-GNN is included as a sequential-recommendation GNN baseline because no GNN-based next-POI recommender is identified in the passage.Source: §5.2 · p. 5
5.3 Implementation Details
- SGRec uses 120-dimensional embeddings, neighbor sampling rate γ=0.2, category-loss weight η=0.2, and two or three CA-GAT layers on Foursquare and Gowalla, respectively.Training uses Adam with batch size 64, learning rate 0.005, and L2 regularization strength 1e-5.Source: §5.3 · p. 5
5.4 Analysis on Recommendation Effectiveness
- SGRec significantly outperforms all baselines on both datasets and improves over Gowalla’s second-best ST-CLSTM by 4.15% on HR@5 and 7.36% on nDCG@5.SGRec provides around 35% average improvement over the best competitors on Gowalla, where interaction density is 0.001 versus 0.052 on Foursquare.Source: §5.4 · p. 5, Table 2 · p. 6
5.5 Ablation Study
- Removing Seq2Graph augmentation or CA-GAT reduces performance on both datasets, while removing category prediction or position-aware attention also degrades results.The category-prediction ablation supports the effectiveness of learning category-level transitions for alleviating POI-level sparsity; the position-aware ablation highlights sequential information and temporal preferences.Source: §5.5 · p. 6, Table 3 · p. 6
5.6 Hyperparameter Analysis
- SGRec achieves best HR@20 at D=120 and η=0.4; performance grows as neighbor-sampling rate γ rises from 0% to 10%, while 2 and 3 CA-GAT layers perform best on both datasets.At γ=0%, iterative embedding updates are disabled; on Foursquare, performance drops slightly when γ≥30%.Source: §5.6 · p. 6
6 Conclusion
SGRec addresses missing collaborative information and sparse POI interactions in sequential recommenders. It adds Seq2Graph augmentation and category-awareness, and experiments on two real-world datasets show effectiveness and robustness.
6 Conclusion
- SGRec combines Seq2Graph augmentation for collaborative signals in POI embeddings with category-awareness to strengthen user-preference modeling.The paper reports experiments as demonstrating SGRec’s effectiveness and robustness.Source: §6 · p. 6