A visiting sequence recommendation framework: Enhanced by dynamic landmark and stay time.

The content shared via social media significantly influences tourists' travel planning process and even destination decisions. By analyzing the shared text, photos, and videos on social media, we can gain insight into what tourists like and how they move. Many researchers have studied tour recommendation problems and proposed various solutions based on location-based social media to address this issue. Despite achieving some progress, research on personalized tours still reveals some difficulties. In this paper, we propose a framework for recommending a visiting sequence containing orderly landmarks and stay time (LST) to meet the user's specified constraints by using user-generated content in a photo-sharing social media. First, the landmarks, also known as POIs, are extracted from the social media photo dataset by clustering photos close to each other, where the closeness can be derived by the geotags in photos. Second, implicit preferences (topics) among the photos are derived by a topic modeling approach called Latent Dirichlet Allocation (LDA). These topics are then used to generate landmark and user feature vectors. Third, users are assigned to the class whose user vectors are similar, and the visiting sequences in the class are used as the base for constructing the prediction model. Fourth, we take advantage of the long-short term memory (LSTM) model, an enhanced version of RNN, and develop an LST-LSTM prediction model to predict the next landmark and its stay time. Finally, the LST-LSTM model keeps generating the most likely landmark and stay time by the proposed time-based tour recommender until the user's specified constraints are researched. The experiments show that the proposed LST-LSTM predictor performs better than the baseline L-LSTM predictor, which does not consider stay time in all metrics. Moreover, our experiment shows that the proposed recommender can help users visit more landmarks and spend less time when compared with the baseline recommender. [ABSTRACT FROM AUTHOR]