Multi-Aspects based Decision-Making Model for Tourism Stakeholders using Autoencoder and Mutual Information

This paper highlights the need for recommendations of similar users on the basis of multiple aspects of hotels such as food, room, location and so on. These aspects are hidden in the hotel reviews and require some productive representation for generating recommendations. This study utilizes a knowledgeable representation of multiple and diverse aspects of hotels from the latent semantics of hotel reviews posted by users on TripAdvisor. The resultant representation is well versed with users’ preferences as compared to single overall rating used by traditional methods which is quite uncertain and ambiguous. The derived semantics of reviews are transformed into a 3-D user-hotel-feature matrix fed to an autoencoder network to learn user embeddings and predict the semantic scores from these compressed embeddings. The original 3-D matrix is reconstructed from the predicted semantics and further utilized to compute the mutual information shared between each pair of users in a similarity matrix. The experimental evaluation of predictions of autoencoder model for the defined 3-D matrix and 2-D ratings matrix shows smaller MSE for the 3-D representation of the semantics of hotels. The users sharing higher mutual information score with a particular user are recommended as similar users. This aids the decision-making process of tourism or hotel stakeholders to target different types of users according to the multiple latent aspects of hotels.