Object-Oriented Travel Package Recommendation
Chang Tan,Qi Liu,Enhong Chen,Hui Xiong,Xiang Wu
DOI: https://doi.org/10.1145/2542665
2014-01-01
Abstract:Providing better travel services for tourists is one of the important applications in urban computing. Though many recommender systems have been developed for enhancing the quality of travel service, most of them lack a systematic and open framework to dynamically incorporate multiple types of additional context information existing in the tourism domain, such as the travel area, season, and price of travel packages. To that end, in this article, we propose an open framework, the Objected-Oriented Recommender System (ORS), for the developers performing personalized travel package recommendations to tourists. This framework has the ability to import all the available additional context information to the travel package recommendation process in a cost-effective way. Specifically, the different types of additional information are extracted and uniformly represented as feature--value pairs. Then, we define the Object, which is the collection of the feature--value pairs. We propose two models that can be used in the ORS framework for extracting the implicit relationships among Objects. The Objected-Oriented Topic Model (OTM) can extract the topics conditioned on the intrinsic feature--value pairs of the Objects. The Objected-Oriented Bayesian Network (OBN) can effectively infer the cotravel probability of two tourists by calculating the co-occurrence time of feature--value pairs belonging to different kinds of Objects. Based on the relationships mined by OTM or OBN, the recommendation list is generated by the collaborative filtering method. Finally, we evaluate these two models and the ORS framework on real-world travel package data, and the experimental results show that the ORS framework is more flexible in terms of incorporating additional context information, and thus leads to better performances for travel package recommendations. Meanwhile, for feature selection in ORS, we define the feature information entropy, and the experimental results demonstrate that using features with lower entropies usually leads to better recommendation results.