The Research And Application Of Recommendation Based On Group Relationship

With the development of Internet, the requirement from users of information has been satisfied. But the overload of information follows as well. Information retrieval system is considered to be one of the best solution to solve this problem, especially search engine. Different users will get the same page while they search the same keyword in traditional search engine. Actually, the demand of users for information is characteristic of multiplicity and individuation, which cannot be satisfied by traditional search engine. Then, another potential solution for information overload has been proposed, that is personalized recommendation system. It recommends information and product which users may interested in according to historical behavior of users. It attracts wide attention and is rapidly developing,However, the personalized recommendation is not always flawless. Its limitations have been gradually revealed. In reality, people may participate in activities in groups. Therefore, this paper mainly focus on the recommendation system for group-oriented users, which exploit the relationship of groups and prefer to consider the preference of groups.There are two types of group. The first type of group is based on the natural social relationships of users, called social network. Another one is detected by the similarity of users according to the potential properties mined from their historical behavior.Based on these two types of groups, the main research contents are as follows:1. This paper proposes a recommendation algorithm which recommends point-of-interests to users by exploiting check-in records of users, social network between users and information of places that users checked in. And it is parallelized on Spark.2. This paper proposed a content based group recommendation algorithm according to the rating records of users and text information related to items. Particularly, it recommendation to groups which the relationship of users are implicit and need to be detect first. We detect the group by LDA model and then compute the matching score between groups and items. Then predict the rating of users to items with factor model and the matching score.3. This paper proposed Parallel Latent Group Model based on LGM,which improve the basic model by proposing a parallel strategy to parallelized it on Spark and implement it. The scalability and reliability has been greatly improved.4. This paper design and implement a recommendation prototype system. We have finished the front-end and background development and aggregated several algorithms into the system. It provides a friendly visual display operating platform, which simulated the real process of recommendation.