Nearest biclusters collaborative filtering framework with fusion

Highlights

• Propose to fuse the item-based CF and user-based CF.

• New similarity measure to obtain the Nearest Biclusters.

• Our approach greatly outperforms classical collaborative filtering techniques.

Abstract

Collaborative filtering is one of the widely used recommendation technique. It provides automated and personalized suggestions to consumers for selecting variety of products by examining their preferences. However, sparsity is one of the major weaknesses of this prosperous approach. This problem inherently occurs in the system due to ever increasing number of users and items. This affects the performance of a recommender system as the accuracy of prediction decreases. Thus, there is a need for a technique that can perform efficiently under sparse environment and this work proposes one such technique. The memory based CF techniques can be user-based or item based. In both cases, the user-item rating matrix can provide only partial information to predict unknown ratings. This is due to the sparsity inherent to rating data. Hence, we propose to fuse the item-based CF and user-based CF. Subsequently, Neighborhood formation is a crucial step in Collaborative filtering technique. Therefore, this paper adopts the biclustering approach for neighborhood formation. This approach, allows a degree of overlap between biclusters (i.e. a user or item is included in more than one clusters). Therefore, a new similarity measure is proposed that obtains a bicluster that has strong partial similarity with an active users’ preferences. Experimental results demonstrate that proposed approach generates better accuracy of rate prediction compared to the tradition item-based, user-based and some state of the art approaches.

Introduction

A Recommender System (RS) [1], [2], [3], is an information-filtering tool that helps to mitigate parts of the information overload generated via explosion of available information on the Internet. It attempts to reduce the amount of information available to the user by presenting information on items and products that are likely to be of interest to the user. For example, sites like Netflix prescribe films, Flipkart and Amazon prescribe products of interest to the users. It collects information on the preferences of its user for a set of items i.e. songs, electronic products, books etc. and makes use of this collected information [4] to provide recommendations to other users. In everyday life, we often rely on suggestions of like-minded individuals, or other trusted sources. A Recommender System is an automated form of this word-of-mouth phenomenon. The operations of a RS are generally based on Collaborative filtering (CF) [5] technique. CF based recommendation system is inspired by human social behavior. According to this, if the users have similar taste for a set of items in past, then they will share common taste in future.

The CF algorithms can be categorized into Memory based and Model based algorithms [6], [7]. The Model based algorithms learn a model in an offline phase and then uses it as a “model” to perform recommendations [8]. On the other hand, Memory based algorithms perform recommendations based on similarity between users or items [9]. Memory based algorithms are further categorized into: User-based CF and Item-based CF algorithms [9]. The User-based CF algorithms preform recommendations by using the previous ratings of most similar users on the specific item. On the other hand, Item-based CF algorithms gives recommendations by finding similar items [10].

Although, CF is one of the most popular and successful approach for recommendation, it experiences a serious limitation specifically sparsity problem. The sparsity problem [11] refers to the situation where the ratio of ratings that need to be predicted to ratings already obtained is very high i.e. to predict more than 90% of the ratings from less than 10% of available data. For instance, a RS used by an online store can make everything available to the customer that exists. As the number of products and users increases tremendously, most of the items even very popular ones has been given feedback or purchased only by few users.

A CF based recommendation system initially determines the most similar users or items and then makes recommendations to individual users. It is based on the other user’s past purchasing history with the fact that the system has only limited amount of available data i.e. less than 1% [12], [13]. Typically, the neighborhood can be formed by using a similarity measure or by adapting a clustering approach to finding the most similar users or items. Due to sparsity, it is hard to define the similarity between items or users, rendering CF useless. Even if the framework succeeds in computation of similarity, there might be a possibility that this similarity is not truthful because of insufficient amount of information processed.

To overcome this problem, the work here proposes a new CF approach algorithm based on user/item preference biclustering to address the sparsity issue. Initially the biclustering approach is adapted to simultaneously cluster the rows and columns of the user/item rating matrix. This helps the system to find the quality neighbors. As a result, we obtain the clusters of users/items with strong partial similarity within cluster. Then the nearest bicluster of active user/item is computed. To obtain the nearest bicluster of each user, we propose a new similarity approach based on Mean Measure of Divergence that computes individual's personal habits to express preferences. The resultant cluster helps to identify the neighbor users or neighbor items. The similarity between users in case of User-based CF or items in case of Item-based CF is computed respectively. Finally, the resultant prediction of each model is combined using weighted sum of item-based CF and user-based CF approaches to generate recommendation results for target users.