Abstract:

With the rapid development of event-based social networks （EBSN）， online platforms， such as Meetup and Douban， attract more and more groups to create， discover and share offline social events， such as concerts， exhibitions， and parties. A suitable venue is essential for the groups to organize a successful event. Therefore， the point of interest （POI） recommendation has become an effective solution to alleviate the information overload that identifies attractive and interesting venues from multiple options. However， event POI recommendation for groups in EBSN is very challenging compared to the traditional recommendation tasks， e.g.， movie recommendation. One reason is the lack of history for a particular event which arises a serious cold-start problem， the other is it involves complex interactions between multiple entities （such as users， groups， events， POI， etc.）. In this paper， multiple entities and their interactions in EBSN are considered， and an event POI recommendation algorithm is proposed based on heterogeneous information network （HIN） and attention mechanism to recommend appropriate POI for groups to host events. First， a principled method is developed to obtain the latent representation of the Meetup entities （groups， events， and POIs） via embedding， which incorporates both qualitative and quantitative information. Then， to explicitly characterize meta-path based context for improving the modeling of the interaction， a priority based sampling technique is used to select high-quality path instances and effective representations of the groups， events， POIs， and meta-path based context is learned for implementing a powerful interaction function. In the original embedding method， each meta-path indeed receives equal attention， which lacks the ability to capture varying semantics from meta-paths in different interaction scenarios. Meanwhile， an effective embedding method for modeling meta-path based context should be interaction-specific， which is able to provide highly discriminative semantics in various complicated recommendation scenarios. Thus， by combining the two parts of attention components， the original representations for the groups， events， POIs， and meta-path based context is improved in a mutual enhancement way， which is called the Co-Attention mechanism. Finally， experiments are conducted on two real-world datasets， namely， Meetup-NYC and Meetup-CHI. Extensive experimental results based on the real-world dataset have demonstrated the superiority of our model in recommendation effectiveness. Our model works especially well for recommending new POIs， which contains little prior history of organizing events. It is believed that our proposed neural model provides a promising approach to utilize HIN information and attention mechanism for improving event POI recommender systems.

Key words: event-based social networks, POI recommendation, heterogeneous information networks, attention-based neural networks