Leveraging multi-aspect time-related influence in location recommendation

Saeid Hosseini, Hongzhi Yin, Xiaofang Zhou, Shazia W. Sadiq, Mohammad Reza Kangavari, Ngai-Man Cheung

TL;DR

This paper introduces MATI, a novel probabilistic model that leverages multiple temporal granularities (e.g., hour of the day, day of the week) to improve location recommendations. By considering how users and locations interact across various time scales, MATI addresses the sparsity and complexity of temporal data in LBSNs, leading to more accurate and personalized suggestions.

What Problem Does This Paper Address?

Recommending relevant Point-of-Interest (POI) locations to users in Location-Based Social Networks (LBSNs) is a challenging task, primarily due to the sparsity of user-location data. While existing methods have incorporated temporal influences, they often focus on a single temporal granularity (e.g., weekday vs. weekend) or a limited set of time-related features. This approach is insufficient because user behavior is influenced by a complex interplay of various temporal factors at different scales (e.g., the specific hour within a day, the day of the week, and how these combine). The paper aims to address this by developing a model that can effectively capture and utilize these multi-aspect temporal influences to enhance recommendation accuracy.

Methods

The core of the proposed solution is the Multi-aspect Time-related Influence (MATI) model, a probabilistic generative model. Key methodological components include:

  • Multi-aspect Temporal Slabs: The paper proposes extracting "temporal slabs" that represent combinations of different temporal granularities. This is achieved by first computing similarities between temporal slots (e.g., hours) within a single latent factor (uni-aspect temporal slab) and then combining these uni-aspect slabs to form multi-aspect temporal slabs. Hierarchical Agglomerative Clustering (HAC) is used for this merging process.
  • Expectation-Maximization (EM) Algorithm: An EM algorithm is employed to infer latent parameters of the MATI model. This iterative process helps to estimate the complex temporal correlations between users and locations, even with sparse data.
  • Hybrid Decision Method: Recognizing that not all users or scenarios benefit equally from temporal influence, a hybrid approach is introduced. This method first assesses whether a user's behavior is temporally sensitive before applying the multi-aspect temporal influence. If not, it defaults to non-temporal methods.
  • Temporal Subset Property (TSP): The model leverages the TSP, which states that finer temporal granularities are subsets of coarser ones (e.g., minutes are part of hours, hours are part of days). This property allows for a generalized model that can incorporate an infinite number of temporal latent factors.

Datasets

The authors evaluate their proposed method on two publicly available LBSN datasets:

  • Foursquare: A widely used dataset for location-based recommendations.
  • Brightkite: Another large-scale LBSN dataset.

The paper notes that both datasets suffer from sparsity, with Brightkite having a higher "cold start ratio" (users with very few check-ins).

Key Insights and Ideas

  • Temporal Granularity Matters: The paper strongly advocates for considering multiple temporal granularities simultaneously, arguing that a single temporal dimension is insufficient to capture the nuances of user behavior.
  • Sparsity Mitigation through Temporal Slabs: The concept of multi-aspect temporal slabs is a novel way to aggregate sparse temporal data, making it more amenable to recommendation algorithms.
  • Hybrid Approach for Robustness: The inclusion of a hybrid decision method is crucial for practical application, ensuring that temporal influences are only applied when they are genuinely beneficial, thus avoiding potential noise.
  • Generalizability of the Model: The MATI model's ability to handle an "infinite number of temporal latent factors" through the TSP suggests a highly flexible and extensible framework.

Key Results

  • The proposed MATI model consistently outperforms other baseline methods (UBCF, USG, USGT) in terms of F1-score across both Foursquare and Brightkite datasets.
  • The hybrid decision method, by selectively applying temporal influence, leads to improved performance.
  • The paper demonstrates that the MATI model can significantly improve recommendation accuracy, particularly for active users, by capturing complex temporal correlations. For instance, it shows an increase in success rate by 4.1% and 6.6% in Brightkite and Foursquare datasets, respectively, at a recommendation level of @5.

Limitations and Future Work

  • Assumption of Stable Temporal Behavior: The current MATI model assumes that users' temporal behavior is stable across their check-in history. However, users' mobility patterns can change due to holidays, travel, or other events.
  • Dynamic Temporal Aspects: Future work aims to adapt the approach to study dynamic multivariate temporal aspects by incorporating correlation networks and considering how temporal slots are affected by their containing latent factors.
  • Behavioral Differences Across Days: The authors plan to address the fact that behavior during different hours of the day can vary significantly depending on the day of the week.

Summary updated

Abstract

Point-Of-Interest (POI) recommendation aims to mine a user’s visiting history and find her/his potentially preferred places. Although location recommendation methods have been studied and improved pervasively, the challenges w.r.t employing various influences including temporal aspect still remain unresolved. Inspired by the fact that time includes numerous granular slots (e.g. minute, hour, day, week and etc.), in this paper, we define a new problem to perform recommendation through exploiting all diversified temporal factors. In particular, we argue that most existing methods only focus on a limited number of time-related features and neglect others. Furthermore, considering a specific granularity (e.g. time of a day) in recommendation cannot always apply to each user or each dataset. To address the challenges, we propose a probabilistic generative model, named after Multi-aspect Time-related Influence (MATI) to promote the effectiveness of the location (POI) recommendation task. We also develop an effective optimization algorithm based on Expectation Maximization (EM). Our MATI model firstly detects a user’s temporal multivariate orientation using her check-in log in Location-based Social Networks(LBSNs). It then performs recommendation using temporal correlations between the user and proposed locations. Our method is applicable to various types of the recommendation models and can work efficiently in multiple time-scales. Extensive experimental results on two large-scale LBSN datasets verify the effectiveness of our method over other competitors. Categories and Subject Descriptors: H.3.3 [Information Search and Retrieval]: Information filtering; H.2.8 [Database Applications]: Data mining; J.4 [Computer Applications]: Social and Behavior Sciences
BibTeX
@article{hosseini2019leveraging,
  author = {Saeid Hosseini and Hongzhi Yin and Xiaofang Zhou and Shazia W. Sadiq and Mohammad Reza Kangavari and Ngai-Man Cheung},
  title = {{Leveraging multi-aspect time-related influence in location recommendation}},
  journal = {World Wide Web},
  year = {2019},
  doi = {10.1007/s11280-018-0573-2},
}
Figure