Self-supervised contrastive learning for itinerary recommendation
摘要
Itinerary recommendation plays a vital role in enhancing tourists’ experiences by offering personalized travel suggestions for navigating unfamiliar urban environments. Typically, it seeks to recommend a sequence of Points of Interest (POIs) for users by considering various factors, such as time constraints, itinerary length, starting and ending POIs. Traditional itinerary recommendation primarily relies on heuristic approaches to address the orienteering problem, aiming to maximize global rewards under budget constraints, or methods that use stochastic processes to model POI-POI transition patterns. These methods fail to capture transitional regularities and develop a comprehensive understanding of user preferences, which may result in recommendations that diverge from tourists’ actual travel intentions, particularly when trip data is sparse. While recent models based on neural networks, such as Transformers, address some of these challenges, they often overlook practical realities, including noisy implicit feedback, the factors that affect tourists’ demands and the complexity of their visiting preferences. Motivated by the benefits of self-supervised contrastive learning, we propose a Self-supervised Contrastive Learning (SCL) framework to the domain of itinerary recommendation. Specifically, a contrastive learning mechanism is presented, focusing on POI representation and two data augmentation operations are designed based on POI correlations to capture the visiting uncertainties in itinerary generation. Additionally, multiple contextual factors are considered, including sequential information, geographic information and POI-POI transition pattern, to effectively capture and update real-time user context. The experimental results demonstrate SCL achieves significant improvements in terms of three metrics, compared with the best performing baseline method. Recall improved by 3.51%–10.31%, Precision improved by 3.42%–12.57% and Levenshtein Distance Similarity (LDS), defined as the normalized similarity between two itineraries based on the minimum edits (insertions, deletions, or substitutions) required to align them, improved by 3.46%–14.73%. The statistical Friedman-Nemenyi test further confirms the robustness of the SCL model. Furthermore, to provide an intuitive understanding of itinerary recommendations generated by SCL, a case study is presented using a randomized example.
类型
出版物
Expert Systems with Applications