Source-linked AI summary
Point-of-Interest Recommendation: Exploiting Self-Attentive Autoencoders with Neighbor-Aware Influence
Chen Ma, Yingxue Zhang, Qinglong Wang, Xue Liu
TL;DR
POI recommenders must model nonlinear user-POI interactions from implicit feedback and incorporate geographic context. SAE-NAD combines a self-attentive encoder with a neighbor-aware decoder, and experiments report improvements over state-of-the-art baselines on three real-world datasets.
Problem
POI recommendation faces challenges modeling nonlinear user-POI interactions from implicit feedback and incorporating geographic context.
Source: §1 · p. 1Method
SAE-NAD combines a self-attentive encoder that distinguishes preference across checked-in POIs with a neighbor-aware decoder using POI embeddings and an RBF distance kernel.
Source: §1 · p. 2Results
Experiments on three real-world datasets validate SAE-NAD's improvements over many state-of-the-art baseline methods.
Source: §6 · p. 10Takeaways & Limitations
The results support combining preference differentiation and geographic influence in an autoencoder-based POI recommender.
Source: §1 · p. 2, §6 · p. 10
Abstract
from arXiv · showhide
The rapid growth of Location-based Social Networks (LBSNs) provides a great opportunity to satisfy the strong demand for personalized Point-of-Interest (POI) recommendation services. However, with the tremendous increase of users and POIs, POI recommender systems still face several challenging problems: (1) the hardness of modeling non-linear user-POI interactions from implicit feedback; (2) the difficulty of incorporating context information such as POIs' geographical coordinates. To cope with these challenges, we propose a novel autoencoder-based model to learn the non-linear user-POI relations, namely \textit{SAE-NAD}, which consists of a self-attentive encoder (SAE) and a neighbor-aware decoder (NAD). In particular, unlike previous works equally treat users' checked-in POIs, our self-attentive encoder adaptively differentiates the user preference degrees in multiple aspects, by adopting a multi-dimensional attention mechanism. To incorporate the geographical context information, we propose a neighbor-aware decoder to make users' reachability higher on the similar and nearby neighbors of checked-in POIs, which is achieved by the inner product of POI embeddings together with the radial basis function (RBF) kernel. To evaluate the proposed model, we conduct extensive experiments on three real-world datasets with many state-of-the-art baseline methods and evaluation metrics. The experimental results demonstrate the effectiveness of our model.
1 INTRODUCTION
Personalized POI recommendation can help residents and tourists discover places while creating opportunities for POIs, but existing approaches face challenges modeling nonlinear interactions, distinguishing check-in preferences, and using spatial context. SAE-NAD combines an attentive autoencoder with a neighbor-aware decoder and performs best on three real-world datasets against state-of-the-art methods.
1 INTRODUCTION
- POI recommendation can help residents and tourists explore unvisited places and create opportunities for POIs to attract visitors.LBSNs provide user-POI interaction data for personalized recommendations.Source: §1 · p. 1
- Existing collaborative-filtering methods based on weighted averages or latent-factor inner products capture only linear user-POI relations.The paper motivates autoencoders as a way to model more complex interactions.Source: §1 · p. 1
- Check-in histories contain POIs with differing importance, while spatial context is critical and sparse implicit feedback mixes negative samples with missing positive samples.Users visit only a small fraction of millions of POIs.Source: §1 · pp. 1–2
- SAE-NAD assigns POIs multidimensional importance scores and uses a neighbor-aware decoder combining embedding inner products with an RBF distance kernel.It also weights visited POIs more highly according to visit frequency while assigning small equal weights to unvisited POIs.Source: §1 · p. 2
- SAE-NAD achieves the best performance on three real-world datasets against state-of-the-art methods.The model combines a self-attentive encoder with a neighbor-aware decoder.Source: §1 · p. 2
2 RELATED WORK
Related work covers collaborative-filtering and geographic approaches to personalized location recommendation, as well as neural recommendation and attention mechanisms. SAE-NAD combines multidimensional attention in an autoencoder with a decoder modeling geographic influence between checked-in and unvisited POIs.
2 RELATED WORK
- Related work reviews personalized location recommendation and attention mechanisms used in recommendation tasks.The section distinguishes personalized location recommendation from attention-based recommendation approaches.Source: §2 · p. 2
2.1 Personalized Location Recommendation
- Personalized location recommendation has used friend-based and model-based collaborative filtering, geographic models, and deep neural networks.Geographic methods include Gaussian distributions, kernel density estimation, and instance- or region-level neighborhoods.Source: §2.1 · p. 2
2.2 Attention Mechanism in Recommendation
- Prior recommendation attention methods model user-item relevance, changing selection criteria, trigger words, multimedia feedback, or review preferences.SAE-NAD instead applies multidimensional attention in an autoencoder and models geographic influence between checked-in and unvisited POIs.Source: §2.2 · p. 3
3 PRELIMINARIES
The preliminaries define POIs, check-ins, and the recommendation task, then introduce notation and autoencoders. They relate autoencoders to matrix factorization and word2vec to situate their use for recommendation.
3.1 Definition and Notation
- A POI is a uniquely identified site with an identifier and geographic coordinates, and a check-in records a user's visit and its timestamp.POI recommendation uses check-in records to suggest places a user is interested in but has not visited.Source: §3.1 · p. 3
- The check-in matrix records visit frequency, while a binary rating matrix marks whether each user has visited each POI.The paper uses “POI” and “location” interchangeably.Source: §3.1 · p. 3
3.2 Autoencoders
- An autoencoder maps input data into a latent representation and reconstructs it, with training aimed at minimizing reconstruction error.The encoder and decoder each use an activation function.Source: §3.2 · p. 3
- With an identity decoder activation, an autoencoder can be rephrased in a form closely related to matrix factorization.The paper relates the representation to user and item latent factors.Source: §3.2 · p. 3
- With one-hot input, an autoencoder resembles word2vec's skip-gram model, and its output can express how likely a user who checked in at one POI is to check in at another.The paper describes this relationship using the skip-gram output and a softmax probability.Source: §3.2 · pp. 3–4
4 METHODOLOGIES
SAE-NAD uses a stacked autoencoder to model user–POI interactions, with a self-attentive encoder for user representations and a neighbor-aware decoder for geographical influence. Its methodology also weights implicit-feedback examples and describes how the network is trained and used for recommendation.
4.1 Model Basics
- The stacked autoencoder encodes users into hidden representations and reconstructs their ratings to capture non-linear user–POI interactions.The hidden representation and reconstructed ratings are the encoder and model outputs, respectively.Source: §4.1 · p. 4
4.2 Self-Attentive Encoder
- Multi-dimensional attention assigns checked-in POIs different importance levels across aspects, then aggregates their weighted embeddings into user representations.The model aggregates the resulting aspect-specific representations into one representation for the encoder.Source: §4.2 · pp. 4–5
4.3 Neighbor-Aware Decoder
- The neighbor-aware decoder combines POI-embedding similarity with an RBF kernel so checked-in POIs exert more influence on nearby unvisited POIs.The geographical kernel uses POI coordinates; the decoder combines this influence with the user preference representation.Source: §4.3 · pp. 5–6
4.4 Weighted Loss for Implicit Feedback
- The implicit-feedback loss treats unvisited POIs as negative examples with a shared weight and increases visited-POI weights with check-in frequency.The frequency-based weights represent different confidence levels in preferences for visited POIs.Source: §4.4 · p. 6
4.5 Network Training
- Training minimizes the regularized objective by gradient descent with backpropagation and uses Adam to adapt the learning rate.The described mini-batch algorithm applies the encoder, decoder, and loss to batches before backpropagating.Source: §4.5 · p. 6
- At prediction time, the model reconstructs a user's rating vector and recommends untrained POIs with the largest prediction scores.The input is the user's binary rating vector.Source: §4.5 · p. 6
5 EXPERIMENTS
Experiments compare SAE-NAD with POI recommendation baselines on three datasets, then examine its components and parameter sensitivity. SAE-NAD leads the reported benchmark comparisons, while ablations and sensitivity tests highlight the roles of attention and geographical influence.
5.1 Datasets
- The evaluation uses Gowalla, Foursquare, and Yelp, with 20% of each user’s visits held out for testing after dataset-specific filtering.The remaining visits form training data; random selection is repeated six times, with results averaged across five partitions.Source: §5.1 · pp. 6–7
5.2 Evaluation Metrics
- Performance is measured with Precision@k, Recall@k, and MAP@k.These metrics assess the precision and recall of top-k recommendations and mean average precision.Source: §5.2 · p. 7
5.3 Methods Studied
- Baselines include implicit-feedback factorization, geographical POI recommenders, and deep models, including WRMF, BPRMF, RankGeoFM, IRENMF, PACE, and DeepAE.SAE-NAD is evaluated as an autoencoder with a self-attentive encoder and neighbor-aware decoder.Source: §5.3 · p. 7
5.4 Parameter Settings
- Experiments use 50-dimensional latent representations, with d_a=20 and γ=60 selected by grid search.The architecture is [N, 500, 50, 500, N] on Gowalla and [N, 200, 50, 200, N] on the other datasets.Source: §5.4 · p. 7
5.5 Performance Comparison
- SAE-NAD achieves the best performance on all evaluation metrics across Gowalla, Foursquare, and Yelp.The performance comparisons are shown in Figures 2–4.Source: Figure 2 · p. 8, Figure 3 · p. 8, Figure 4 · p. 8
- The authors attribute SAE-NAD’s comparisons to its geographical influence and nonlinear interaction modeling, while PACE leads the remaining baselines.The paper contrasts SAE-NAD with geographical factorization methods and BPRMF, and attributes PACE’s baseline performance to neural embeddings and context graphs.Source: §5.5 · pp. 8–9
5.6 Impacts of Self-Attentive Encoder and Neighbor-Aware Decoder
- Across all three datasets, SAE-WAE outperforms WAE, and NAD-WAE outperforms both SAE-WAE and WAE.Table 3 reports these component comparisons; the authors connect the attention gain to representative POIs and the NAD gain to correlations between checked-in and unvisited POIs.Source: Table 3 · p. 9, §5.6 · p. 9
5.7 Sensitivity of Parameters
- As d_a increases, performance improves and stabilizes; γ=0 performs poorly, while larger γ strengthens correlations between POIs.Parameter trends are presented for Gowalla and Foursquare; the authors report similar trends for Yelp.Source: Figure 5 · p. 9, Figure 6 · p. 9, §5.7 · pp. 9–10
6 CONCLUSION
SAE-NAD uses a self-attentive encoder to distinguish preference degrees across checked-in POIs and a neighbor-aware decoder to model geographical influence on unvisited POIs. Experiments on three real-world datasets showed improvements over many state-of-the-art baselines.
6 CONCLUSION
- SAE-NAD combines a self-attentive encoder with a neighbor-aware decoder for POI recommendation.Source: §6 · p. 10
- The self-attentive encoder assigns importance score vectors to checked-in POIs to distinguish user preference degrees.Source: §6 · p. 10
- The neighbor-aware decoder models geographical influence from checked-in POIs on unvisited POIs to differentiate user reachability.Source: §6 · p. 10
- Experiments on three real-world datasets validated improvements over many state-of-the-art baseline methods.Source: §6 · p. 10