Prediction of Product Recommendation

Authors: Neha Kale, Mohit Jain

Certificate: View Certificate

Abstract

Recommender Systems (RS) are widely used for providing automatic personalized suggestions for information, products and services. Collaborative Filtering (CF) is one of the most popular recommendation techniques. However, with the rapid growth of the Web in terms of users and items, majority of the RS using CF technique suffer from problems like data sparsity and scalability. In this paper, we present a Recommender System based on data clustering techniques to deal with the scalability problem associated with the recommendation task. We use different voting systems as algorithms to combine opinions from multiple users for recommending items of interest to the new user. The proposed work use K-MEAN clustering algorithm for clustering the users, and then implement voting algorithms to recommend items to the user depending on the cluster into which it belongs. The idea is to partition the users of the RS using clustering algorithm and apply the Recommendation Algorithm separately to each partition. Our system recommends item to a user in a specific cluster only using the rating statistics of the other users of that cluster. This helps us to reduce the running time of the algorithm as we avoid computations over the entire data. Our objective is to improve the running time as well as maintain an acceptable recommendation quality. We have tested the algorithm on the Kaggle Product dataset.

Introduction

In everyday life, we often face a situation in which we need to make choices without sufficient personal experience. Ever increasing volume of information on the web has created the need for automated filtering, refinement and personalized presentation of information to users to help decision making. There have been efforts to design information filtering systems that filter and present information according to The preferences of the individual user. Recommender Systems (RS) form a subclass of information filtering Systems that help the users in their decision making process by suggesting items that the users may prefer. RS are being used in a number of e-commerce sites to help customers in finding suitable products [13]. Such systems should be able to identify the user preferences for items in the application domain. Typical application domains for RS include recommendations for music CDs and DVDs1, 1http://www.dvdcdnow.com/ products2 and books3. Majority of the RS use Collaborative Filtering (CF) techniques [1], [7], [15], to predict the likely preferences of a user based on the known preferences of the other similar users. However, these CF based RS require computations that are very expensive and grow polynomials with the number of users and items in the database. To address this scalability problem, we propose a recommendation method using data clustering techniques and voting algorithms. Our proposed system avoids the costly similarity computations of the CF process by applying a voting4 based recommendation algorithm separately to each cluster. Note that though we partition the users’ space into smaller clusters and applied the voting based Recommendation Algorithm individually to the clusters, it does not mean that two users in different clusters cannot have similarity in the rating patterns. It may also happen that the recommendation quality degrades as we recommend only using the data of a particular cluster. However, our goal is to reduce the overall running time without sacrificing the recommendation quality much. This ensures scalability, allowing us to tackle bigger datasets. To demonstrate the applicability of our method, we are developing a Product Recommender System that will cater to the interests of users. RS for products have many dimensions. Each dimension has a set of attributes or elements. One of these dimensions may describe the type of a product (genre) and contain elements like horror, comedy, tragedy, musical, action, etc. RS usually combine the value/rating of the attributes of every dimension according to some evaluation criteria to obtain a recommendation rating of an item. In the proposed Product Recommender System, a classical voting method is used as the evaluation scheme. Principles of voting theory have been satisfactorily used for many years in multi-agent systems [4] regarding group decision making that maximizes social welfare. So, the use of voting theory in the proposed system promises recommendation that optimizes the user preferences.