A HieArarchical LSTM Framework for Capturing Long- and Short-Term Preferences in POI Recommendation

Abstract

Point-of-Interest (POI) recommendation is crucial for improving user experience in location-based social networks (LBSNs). With the growing number of users checking in at various places personalized recommendations are necessary to provide relevant suggestions. Existing methods use long short-term memory (LSTM) networks to model user preferences. However, these methods either consider long- and short-term preferences separately or merge them into a single model without effectively capturing the interactions. This research revisits the problem of long- and short-term preference learning by proposing a hierarchical LSTM (HiLSTM) framework. The framework aims to enhance next POI recommendations by learning representations at two levels: POI-level and semantic-level. Instead of treating these factors independently, HiLSTM integrates them through a structured hierarchical learning approach. One of the key challenges in POI recommendation is handling the sparsity of check-in data. Many users frequently visit new locations. It makes difficult to rely solely on past visits. The proposed model addresses this by introducing a semantic filter. It provides recommendations based on a user’s categorical preferences. By filtering out irrelevant POIs at an early stage, the recommendation process becomes more effective and computationally efficient. To capture long-term user preferences, HiLSTM employs an attention mechanism. Meanwhile, short-term preferences are derived from recent check-ins. It confirms that immediate user intent is not overlooked. The combination of these two components results in a more balanced and accurate recommendation system. These datasets contain check-in records from location-based social networks, enabling rigorous evaluation. The hierarchical structure and attention mechanism contribute to a significant improvement in recommendation precision. This work introduces a novel hierarchical LSTM framework for next POI recommendation.