| As the rapid development of economy, especially the popularization of the Internet, electronic commerce industry has risen sharply. Nowadays, online shopping has become a fashion and direction of future development. However, as the amount of information quickly increasing, it is difficult for users to choose proper products. The task of recommender system(RS) is to provide a method to solve this problem, namely it helps users to find their favorite products more quickly and effectively.Cold-start is a difficult problem of RS, and has gradually been focused on by researchers as a classic problem. Recently, most cold-start recommender systems have used to deal with the problems where either the user or the item is new. However, the third problem of cold-start, namely to recommend new items for new users, has seldom been considered. Due to the historical records of the current user and item are unknown, it is more challenging to solve the problem for this situation.In this paper, we present a recommendation algorithm based on granular association rules to solve the cold-start recommendation problem, especially recommend new items to new user. Since current granular association rules cannot deal with the multi-value and numerical attributes problems, we first propose the solutions to these problems, and then apply them to recommendation algorithm. The main research results are listed as follows.First, we deal with multi-value attribute problem of granular association rules. At present, there are many multi-value data types. For example, the genres of a movie are both action and adventure. In this paper, we adopt scaling-based approach to deal with multi-value attribute problem of information system. During granular association rules are mined, the negative granules are filtered out, and the active ones are preserved. So that it is beneficial to reduce negative rules, and then avoid being recommended.Second, we deal with numerical attribute problem of granular association rules. At present, the research works of granular association rules only consider the nominal data type; however there are a large number of numerical data in the real world. In this paper, we adopt Equal Width approach, Equal Frequency approach and K-Means approach to deal with the numerical data. So we can mine more and stronger rules, which will be used for RS later.Third, we define three measures to evaluate the recommendation algorithm, namely accuracy measure, significance measure and diversity measure. Here accurate measure can reflect the quality of recommendation algorithm, and significance and diversity measures can reflect the performance of personalized recommendation algorithm.Fourth, we design the cold start recommendation algorithm based on granular association rules. This part is the core of this paper. First, we regard new users and new items as different, such as “35 year old male manager”, “romantic action movie”, and “comedy movie released in 2014”. These are all information granules. According to existing users, existing items in the many-to-many entity relationship system, we mine association rules mining by satisfying the conditions of different measures of granular association rules. According to these rules, the new user can match the appropriate user granules. Then we choose recommended granules through the confidence-based and significance-based approaches. Finally, we match the granules of new items through recommended ones, and recommend the corresponding items. We also study Top-N cold-start recommendation based on granular association rules.Experiments are undertaken on the Movie Lens dataset. Results indicate that multi-value and numerical attributes can be effective to deal with in granular association rules. The cold-start recommendation algorithm based on granular association rules has successfully been applied in three kinds of cold-start recommendation problem, and obtains good accuracy. At the same time, our algorithm is effective to obtain diversity and Top-N recommendation. |
PDF Full Text Request