DEVELOPMENT OF A MOBILE CONTEXT-AWARE SYSTEM FOR RECOMMENDING PLACES OF INTEREST

Abstract

The key target of recommender systems is usually to suggest items based on user preferences, but user preferences vary in different contexts (such as different moods) and in different locations. Some existing tactics towards recommender systems majorly considered recommending the most relevant items to individual users and do not take into consideration any contextual information such as user’ mood, budget, companion and location. Meanwhile, importance of these contextual information has been identified in many disciplines, such as information retrieval, data mining, and mobile computing. In this research work a mobile context-aware system that recommends places of interest to users is developed. For a new user, the system collects context data (on user’s mood, budget size of user and companions joining the user) that depicts the user’s current context. Each of these context data has weight in the database; the weights are fetched and used for calculation in order to give recommendation to user with no ratings using Bayesian algorithm. Then the new user is encouraged to rate a series of Places of Interest (POIs). These ratings are estimated to provide improvement of the quality of subsequent recommendations. For an existing user, the system switches to the collaborative-filtering part. In this case, POIs are recommended based on where other users have visited in similar context conditions. The recommender system stores the ratings for each POI in each context for each user. User’s context similarities are calculated using cosine similarity algorithm. POIs locations (such as restaurants, hotels and landmarks) were obtained from Google maps. Only locations that have already been Google mapped were considered in this research work. Evaluation of the system was conducted using questionnaire distributed to general users, IT inclined users and software developers. Discounted cumulative gain was used to evaluate the ranking quality. The normalized discounted cumulative gain (nDCG) values from the experiments show that the ranking of POIs in the recommendations list are very good. The average scores (in %) of comparative analysis of two related existing works and the proposed systems gave 76%, 80% and 83% respectively, in term of accuracy and efficiency. This result shows that the proposed system is more accurate and efficient than existing works.