[Paper Review] Learning Spatiotemporal-Aware Representation for POI Recommendation

Bei Liu, Tieyun Qian+4 more|arXiv (Cornell University)|Apr 28, 2017

Recommender Systems and Techniques9 references3 citations

TL;DR

This paper proposes a spatiotemporal-aware (STA) representation model for point-of-interest (POI) recommendation that treats time and location as a joint $<$time, location$>$ pair to model user-POI relationships via translation vectors in a knowledge graph embedding framework. The method achieves state-of-the-art performance, significantly outperforming baselines in accuracy, robustness to data sparsity, and handling of cold-start POIs by learning shared semantic representations through translational dynamics.

ABSTRACT

The wide spread of location-based social networks brings about a huge volume of user check-in data, which facilitates the recommendation of points of interest (POIs). Recent advances on distributed representation shed light on learning low dimensional dense vectors to alleviate the data sparsity problem. Current studies on representation learning for POI recommendation embed both users and POIs in a common latent space, and users' preference is inferred based on the distance/similarity between a user and a POI. Such an approach is not in accordance with the semantics of users and POIs as they are inherently different objects. In this paper, we present a novel spatiotemporal aware (STA) representation, which models the spatial and temporal information as \emph{a relationship connecting users and POIs}. Our model generalizes the recent advances in knowledge graph embedding. The basic idea is that the embedding of a $$ pair corresponds to a translation from embeddings of users to POIs. Since the POI embedding should be close to the user embedding plus the relationship vector, the recommendation can be performed by selecting the top-\emph{k} POIs similar to the translated POI, which are all of the same type of objects. We conduct extensive experiments on two real-world datasets. The results demonstrate that our STA model achieves the state-of-the-art performance in terms of high recommendation accuracy, robustness to data sparsity and effectiveness in handling cold start problem.

Motivation & Objective

• To address the limitations of existing POI recommendation methods that treat users and POIs as equivalent entities in a shared latent space.
• To model the joint influence of time and location on user check-in behavior more effectively than methods using only spatial or temporal factors.
• To improve recommendation performance under data sparsity and cold start conditions by learning robust, joint spatiotemporal representations.
• To develop a method that naturally captures the semantic relationship between users and POIs through translation vectors in a unified embedding space.

Proposed method

• The model represents the joint spatiotemporal context as a $<$time, location$>$ pair and models it as a translation vector from user embeddings to POI embeddings.
• It generalizes knowledge graph embedding techniques, specifically using a translational mechanism where $\vec{u} + \vec{r}_{t,l} \approx \vec{v}$, with $\vec{r}_{t,l}$ being the translation vector for time and location.
• The recommendation is performed by computing the similarity between the translated user embedding and candidate POI embeddings, selecting the top-k POIs with highest similarity.
• The model is trained end-to-end using a negative sampling strategy to optimize the ranking loss for personalized POI recommendations.
• An extended version, STA-C, is proposed to handle cold-start POIs by leveraging shared $<$word, location$>$ pairs to infer translation vectors.
• The method avoids direct user-POI distance computation in a common space, instead using structured relational knowledge to model user preferences.

Experimental results

Research questions

• RQ1 Can modeling time and location as a joint $<$time, location$>$ pair improve POI recommendation accuracy compared to using either factor alone?
• RQ2 Does a translation-based embedding mechanism that connects users to POIs via spatiotemporal relations outperform conventional user-POI similarity methods in sparse data regimes?
• RQ3 To what extent does the proposed STA model mitigate the data sparsity problem in POI recommendation?
• RQ4 How effective is the STA model in handling cold-start POIs compared to existing baselines?
• RQ5 Can the translational structure of the model generalize well across diverse real-world LBSN datasets?

Key findings

• On the Foursquare dataset, STA achieves a 20.00% improvement in accuracy@1 over the GE baseline when training data is reduced by 20%, indicating superior robustness to data sparsity.
• On Gowalla, STA reduces accuracy drop by 15.73% compared to GE under the same 20% data reduction, further confirming its robustness.
• The STA model reduces accuracy degradation by 20.00% on Foursquare and 15.73% on Gowalla under data sparsity, outperforming GE which sees drops of 31.69% and 18.91% respectively.
• For cold-start POI recommendation, STA-C consistently outperforms GE on Foursquare, with a significant margin due to its ability to leverage shared $<$word, location$>$ pairs for translation vector estimation.
• The model’s performance remains stable and superior even with limited check-in records, demonstrating its effectiveness in handling cold-start scenarios.
• The ablation study confirms that the joint modeling of time and location as a single relational entity is critical to the model’s performance, as opposed to treating them separately.