Effective and diverse POI recommendations through complementary diversification models

Highlights

• The DisCovER, a method able to improve diversity and accuracy of POI recommendation.

• An adaptation of DisCovER able to personalize the diversity of POIs recomendation.

• A deep experimental evaluation comparing DisCovER to the state-of-the-art algorithms.

• An adaptation of the Multi-Attribute Utility (MAUT) for the Recommendation field.

Abstract

Nowadays, recommender systems play an important role in several Location-Based Social Networks (LBSNs). The current advances have considered the trade-off between accuracy and diversity to help users to discover and explore new points-of-interest (POI). However, differently from traditional recommendation scenarios, other equally relevant dimensions (e.g., social and geographical user information) have to be considered to understand how the characteristics of services offered by each POI fit the user needs. Specifically, this work sheds light upon naive failures introduced by traditional recommendation methods while they handle this trade-off between diversity and accuracy in POI recommendations. We hypothesize that some efforts on POI recommendations somehow are deviating from basic learnings from the area. In this context, this work addresses four characteristics inherent to the POI domain that previous efforts have failed to recognize: (1) POI categories and locations are complementary dimensions of diversification that should be simultaneously addressed; (2) Diversity is a complex concept that should be modeled by distinct and non-orthogonal models; (3) Distinct users have different biases and willingness to move to fulfill their needs; (4) POI recommendation is a multi-objective task. In order to demonstrate the gains of properly addressing these aspects, we also propose DisCovER, a straightforward re-ordering method that linearly combines geographical and categorical diversification. DisCovER results demonstrate that even simple strategies to exploit simultaneously these complementary dimensions can increase diversification while keeping accuracy high. Differently from state-of-the-art diversification methods, DisCovER does not penalize any quality dimension in favor of others. It allows us to discuss future directions towards more robust user modeling and preference elicitation in POI domains.

Introduction

Location-Based Social Networks (LBSNs) have become important tools for people interested in discovering and exploring new places (Ricci et al., 2011, Lu et al., 2015). LBSNs, such as Foursquare, Flickr, and Weibo, allow users to share their personal experiences on Points-of-Interest (POI), helping other users to decide new places to visit. Nowadays, POI recommendation plays an important role in providing better location-based services in LBSNs for both users and POI owners (Liu, Fu, Yao, & Xiong, 2013). Similar to traditional recommendation domains, handling the trade-off between accuracy and diversity is a major challenge to issue useful recommendations on LBSNs. Differently, however, this domain adds an equally relevant dimension to this challenge: the geographical distance between users and each POI. Besides understanding how the characteristics of services offered by each POI fit the user needs, realizing how far users are willing to move to fulfill these needs is of paramount relevance.

This work sheds light upon naive failures to handle this trade-off between diversity and accuracy on POI recommendations. After conducting an extensive literature review, we hypothesize that some efforts on POI recommendations somehow are deviating from basic learnings from the area. More specifically, this work discusses four characteristics inherent to POI domains that previous efforts failed to recognize: (1) POI categories and locations are complementary dimensions of diversification that should be simultaneously addressed; (2) Diversity is a complex concept that should be modeled by distinct and non-orthogonal models; (3) Distinct users have different biases and willingness to move to fulfill their needs; (4) POI recommendation is a multi-objective task.

In order to demonstrate the gains of properly addressing these aspects, we also propose DisCovER, a straightforward re-ordering method that linearly combines geographical and categorical diversification. DisCovER explores the Intra-List Distance (ILD) (Vargas & Castells, 2011) and Genre Coverage (Vargas, Baltrunas, Karatzoglou, & Castells, 2014) to achieve the categorical-diversification, and the Proportional Geographical Diversification (${\mathit{PR}}_g$) (Han & Yamana, 2017) to achieve the geographical-diversification. Quantitative experiments on DisCovER allow us to discuss future directions towards more robust user modeling and preference elicitation in POI domains.

Regarding the first characteristic, recent works point out that state-of-the-art POI recommender systems harm diversity in favor of accuracy (Han & Yamana, 2017). Since these methods are friend-based and/or geolocation-based, they restrict recommendations to place already known by the social cluster of each user or close to his/her current location (Xu et al., 2015, Xu et al., 2016, Ren et al., 2017, Cai et al., 2018, Zhang et al., 2019, Zhang et al., 2020). There is an increasing consensus in the literature that diversification strategies are essential to enhance POI recommendation quality (McNee et al., 2006, Smyth and McClave, 2001, Wu et al., 2019, Zhang and Hurley, 2008). However, current efforts towards diversifying POI categories and locations are disjoint and non-aligned. On one hand, some works provide more diverse recommendations of POI categories but limited to a small area (Han & Yamana, 2017). On the other hand, recent efforts maximize the geo-diversity of recommendations without concern about the impacts on the categories (Han & Yamana, 2017). DisCovER results on two Yelp Challenge datasets indeed demonstrate that even simple strategies to exploit simultaneously categorical and geographical diversity can increase diversification while keeping accuracy high.

The second characteristic states a well-known fact in many domains (Agrawal, Gollapudi, Halverson, & Ieong, 2009). Distinct authors propose many strategies to assess and model diversity in different domains (Lu and Tintarev, 2018, Schedl and Hauger, 2015, Rao et al., 2013, Han and Yamana, 2017). The results so far seem to consolidate a clear message: there is no single way to model and provide diversity. Thus, why should we use a single strategy to model diversity in POI recommendation? In order to evince the potential gains on combining different diversity models, we contrast DisCovER, which exploits two strategies for assessing category diversification (Genre Coverage and ILD), against modified versions of itself that use only one strategy. The results confirm that Genre Coverage and ILD are equally important to explain the gains provided by DisCovER in the experimental scenarios.

The third characteristic also represents a solid knowledge for the area (Lee et al., 2009, Gonzalez et al., 2008). Although many POI RSs consider personalized user contexts and preferences, they still ignore the fact that users exhibit a distinct desire for diversity, familiarity, or even novelty, affecting the relevance of each quality dimension. By taking into account this user behavior, DisCovER is also able to prioritize quality dimensions that better suit the personalized needs of each user. Finally, we highlight the multi-objective nature of POI recommendations. Rather than prioritizing a single quality dimension, new proposals should balance compromises. To show gains with this assumption, we adapt a traditional metric of multi-attribute utility (MAUT) to the recommendation. Differently from two state-of-the-art diversification methods (PM2 and Binom Vargas and Castells, 2011, Dang and Croft, 2012, Vargas et al., 2014, Han and Yamana, 2017), DisCovER does not penalize any quality dimension in favor of others.

Therefore, the main contributions of this work can be summarized as follows:

• A new strategy for POI recommendation, the DisCovER, which is able to explore simultaneously complementary dimensions of diversification (i.e. categorical and geographical) while keeping accuracy high;

• A personalized version of DisCovER to apply distinct levels of categorical and geographical diversification for each user;

• An adaptation of the Multi-Attribute Utility (MAUT) strategy for the Recommendation field to measure how effective the methods are, considering several evaluation metrics at the same time; and

• A deep experimental evaluation by comparing DisCovER to several state-of-the-art algorithms and considering, simultaneously, several classes of measures, such as accuracy, novelty, and diversity. DisCovER presents results statistically superior to all baselines. This corresponds to an advanced for this kind of research since most of the works consider just one class of metrics.