چكيده به لاتين
A massive amount of data is being sent daily between social network users which allows companies to extract and retrieve useful information from them. A large part of these data is composed of texts with these characteristics: high amount of noise, limited length, abbreviated words, and informal content. Hence, the semantic understanding of such texts would be very complex. Also, experiments show that short-text content is affected by the time they are sent in. Another part of this data is spatial data and geographic coordinates that suffer from excessive sparsity. Hence, the systems that use textual, temporal and spatial data types which were transmitted on social networks at the same time, such as location or tourism recommendation systems, face a variety of challenges to achieve the best results. In this research, a system is designed and implemented which uses a time-sensitive word embedding method for better short-text semantic understanding by taking into account the temporal patterns of words used. Then we propose a group of methods which convert various textual items such as short-text, content, concept, cluster, and user into vectors. It allows the combination of different textual items by executing mathematical operators on their vectors. Then to use spatial data, different users are defined at four levels of the alley, street, neighborhood, and city with the corresponding location vectors. The system, according to the obtained vectors, computes the textual-temporal similarity and spatial similarity of the users and combines them at the end. Finally, a web interface is implemented to support the user interactions with the recommender system to recommend POIs visited by a user to the most similar user to her/him. The implemented system was tested by a series of experiments. Based on the results of the textual-temporal embedding method, the maximum accuracy of the word-analogy test on the word vectors derived from the CBOW embedding model has increased from 12.1 percent to 13.6 percent. Also, the concept-based similarity and the textual-temporal similarity which are introduced in this study, with the values of 44.66 and 85.36 out of 100 for the concept-based and normal weighted precisions, exceeded the results from other existing methods and showed better performance in detecting similar users. The implemented Visual and adaptive recommender system has also met the objectives of this research work.
