Sequencing of items in personalized recommendations using multiple recommendation techniques

Highlights

• Proposed approach has high precision value for small top-n recommendations.

• Sequencing of items in recommendation list is made on basis of popularity.

• Both ratings and opinions of users are used about items.

• Handles item side cold start and gray sheep problems in fairly simple manner.

• Experiment shows around 20% improvement in precision for small n values.

Abstract

Recommendation System (RS) is a piece of software that gives suggestions according to the interest of users in many domains like products in e-commerce, tours, hotels, entertainment etc. In any of the established e-commerce website hundreds of products are available under the same category. RS helps buyers to find the right product based on buyer’s past buying pattern and item information. Currently many established approaches for item recommendations like content based filtering, collaborative filtering, matrix factorization, etc., exist. All these approaches create a big list of item recommendations for the target user. In general most users prefer to see only top-n recommendations, where the value of n is small and just ignores remaining recommendations. It means good RS must have high precision value for smaller values of n but at present almost all recommendation systems to the best of authors’ knowledge are having high recall value and low precision value. It clearly means that top-n recommendations generated by these systems have very few items that may be liked by the target user. The proposed approach generates recommendations by combining features of content based filtering, collaborative filtering, matrix factorization and opinion mining. The proposed RS dynamically keeps track of user’s inclination towards different types of items with respect to time. It analyzes user’s opinions about products and finds the product popularity in the market by its own unique way. In the proposed approach, items are arranged in such a way that almost all preferred items by target user comes under top-n recommendations. The experimental results show that top-n recommendations generated by the proposed approach for smaller value of n have high precision value when compared with other traditional benchmark recommendation methods.

Introduction

Availability of Internet and other related technologies have given platform for the expansion of e-commerce websites. Every e-commerce website has its own Recommendation System (RS) that finds user’s interest implicitly or explicitly and recommends items to users depending on their interest. RS helps e-commerce websites in retaining their existing customers and attracting new customers. Most extensively used recommendation approaches are content based filtering and collaborative filtering.

Content based filtering recommends items to users based on their past likings. It creates the profile of each user using attributes of items that user has rated high earlier and compare it with other non rated items in order to recommend new interesting items (Claypool et al., 1999). Widely known examples of such systems are InfoFinder and NewsWeeder (Krulwich, Burkey, 1996, Lang, 1995). Major drawbacks of content based filtering is that it does not provide quality assessment and serendipitous recommendations.

Collaborative Filtering (CF) addresses these shortcomings of content based filtering, it is based on human judgments (ratings). Collaborative filtering recommends items to the target user that are already items of interest for other like minded users. Sometimes CF systems are not able to make suitable recommendations to the target user because no ratings or only a small number of ratings are available for new user or new item, this situation is known as Cold Start (CS) problem in recommender system. CS problem occurs when new item or new user is introduced into the system (Balabanovic, Shoham, 1997, Claypool, Gokhale, Miranda, Murnikov, Netes, Sartin, 1999, Sarwar, Konstan, Borchers, Herlocker, Riedl, 1998). CF system also suffers from the Gray Sheep (GS) problem. In this problem a group of users/ user will not receive correct recommendations because their views are different from any group of people. These users rarely get benefit from the collaborative filtering although the recommendation system has enough ratings to make recommendations (Balabanovic, Shoham, 1997, Claypool, Gokhale, Miranda, Murnikov, Netes, Sartin, 1999). Many researchers have tried to solve these problems by the use of hybrid approaches of content based filtering, collaborative filtering and matrix factorization techniques. The success of collaborative filtering relies on finding the similar users of the target users. Collaborative filtering performs well if there is sufficient data available in the rating matrix. In general any single user rates few items in comparison to large number of available items, this situation leads to the sparseness problem in the collaborative filtering. The insufficiency of data in rating matrix hinders CF process in generating good quality recommendations. Matrix factorization is another technique of recommendation that alleviates the sparse rating problem of the collaborative filtering.

Matrix factorization is another recommendation technique that finds latent features of both users and items inferred from their rating pattern (Koren, 2008). This technique is very similar to singular value decomposition method. It uses user-item rating matrix to find the latent features.

With the arrival of Web 2.0, almost all e-commerce websites are giving opportunity to users to write their own experiences about products in the form of reviews. Now users are reading reviews of other users before making any new purchase. All prominent websites are having hundreds of reviews about popular items, so it becomes very difficult for the user to read every review before making any new purchase. Opining mining is alleviating this pain of users by finding the semantic orientation of reviews i.e. positive or negative.

Most users prefer to see only top-n recommendations from e-commerce websites. Almost all buyers prefer smaller value of n. But most established recommendation techniques generate recommendations with high recall and low precision for smaller value of n. It means many recommended items that are of interest of the target user may come at the very end of the recommendation list. The proposed RS tries to keep user’s preferred items at the beginning of the recommendation list, which results in high precision value for smaller values of n. The proposed approach creates the profile of each user dynamically using content based filtering. It uses collaborative filtering to generate item recommendations list. It uses collaborative filtering as a classifier. The proposed approach finds the popularity of all items in the market and converts reviews into ratings using opinion mining. It uses matrix factorization to predict good quality of approximate missing ratings and generates recommendations (Alami, Nfaoui, & Beqqali, 2015). The proposed recommendation technique also addresses the item- side cold start problem and preserves gray sheep users interests in a fairly decent manner.

The key contributions of this paper are as follows:

• Most users prefers to see only small size recommendation list with their items of interest in it. In this approach we have proposed a technique that increases the precision value of top-n recommendations, specially for small value of n. It means that the most of the items in small size list recommended by the proposed approach are liked by users.

• With the arrival of Web 2.0, millions of product reviews are available in e-commerce websites.To the best of authors knowledge in comparison to other approaches found in literature our approach uses these reviews to generate high precision top-n recommendations using opinion mining in co-operation with collaborative, content based filtering and matrix factorization technique.

• The proposed approach uses people’s opinions to calculate popularity of different items available in e-commerce website and uses it to generate high precision top-n recommendations.

• Now-a- days almost all e-commerce websites are having data about user’s ratings, reviews and descriptions of items that they are selling. Our approach uses only ratings, reviews and items description data to generate recommendations. This makes our approach suitable to almost all the e-commerce websites.

• The proposed approach is developed without using user’s tags and social network related data in generating user’s top-n recommendations because it is almost unavailable to most e-commerce websites.

• Proposed approach has a provision that changes the profile of each user dynamically with the changing interest of each user. Our approach does not use complex calculations and handles item side cold start and gray sheep problems also in fairly simple but robust manner.

The paper is further organized as follows: Section 2 discusses background knowledge and related work, Section 3 throws light on proposed approach, Section 4 explains practical implementation and evaluation. Finally Section 5 concludes the paper.