Recommendation System Algorithms

Recommendation System Algorithms

Bhavya Chaudhary

Master Trainer
WCSC-ITE Singapore

bhavyachaudhary26@gmail.com

Kaushal Mehta

Associate Professor
BPIBS, Government Of Delhi
kpu_713@yahoo.com

Associate Software Engineer
Accenture

mohandhwani@gmail.com

ABSTRACT

Recommendation systems have re-shaped E-commerce sites from an incipient to serious business implementations. Almost all big commerce websites use one or a combination of recommendation systems to personalize the online market store. Also, the system makes it easy for the customers to find the product on their wish list. In order to sail through the competitive world and operate in the challenging environment it is imperative for the websites to serve the customers well and maintain a robust relationship with them. The system learns from customer and displays the most valuable products from the available stock. In this paper we present explanation of how the recommendation system improves the customer base, increase the sale and Gross Merchandise Value. Also, the paper gives a detailed comparison of the recommendation system and algorithms used by major e-commerce players.

Keywords

E-commerce, Filtering, Customer base, Recommendation System, Collaborative Filtering.

1. INTRODUCTION

E-commerce today has provided the customers with variety of options augmented with new level of customization and an abundance of information that is to be processed by the customer before making a selection. The recommendation system comes to the rescue and, provides flexibility and ease to the customer while making a choice. The products can be recommended by keeping various factors into consideration like overall sellers on a site, demographics, past purchase history of the customer. The system enhances the sales in following ways:

Ease to browse: Today, in the fast paced world, no one has the time to browse through a wide variety of products. The system can help customers find products they wish to purchase.

Additional Recommendation: By suggesting additional products the order size can be increased. If the recommendations are valuable then the sale base can be enhanced. The site often recommends additional products during the process of checkout.

Nexus with the customer: Customers returns to the site where the requirement is best matched. The more a customer uses the system, more it will be able to improve the efficiency of the algorithm running at the backend. The system adds value to the Customer Relationship Model.

The most common approaches to solve the recommendation problem are traditional collaborative filtering, cluster models and search-based methods.

[pic 1]

                RECOMMENDATION SYSTEM

1. RECOMMENDATION SYSTEM ALGORITHMS  

The prior most task is to find a set of customers who’s purchased and rated items are similar to user’s purchased and rated items. The items which are already purchased are eliminated and the rest are recommended. Two most popular algorithms for the above are collaborative filtering and cluster models. Apart from these the algorithms that do not focus over customers but items include search-based algorithms.

1. TRADITIONAL COLLABORATIVE FILTERING

In traditional collaborative filtering algorithm a customer is treated as an N-dimensional vector. N is the distinct items in the catalog. If the purchase has been made or the item is rated positively the vector is positive else the vector is negative. To obtain the best-selling product, the vector is multiplied with inverse of purchase frequency making less known items more relevant.

Recommendations are generated keeping a few customers as the basis which can be used to predict the rest. One way to generate the recommendations can be to rank all the item in accordance with the purchase by similar customers.

This way of recommendation is expensive. The worst case complexity is O(MN), where M is the customer base and N is the number of items. However, the average customer vector being sparse, the algorithmic performance should be closer to O(M+N). O(M) processing time is required for all the customer. Further, each customer has significant percentage of catalog items requiring a processing time O(N). Thus, final performance comes out to be O(M+N). For varied count of the number of customers as well as of items the algorithm would face performance issues.

In order to deal with the performance issue the data size can be reduced by sampling. A few customers can be discarded as well as segregation on the basis of popular or unpopular items can be done. We have various for the reduction technique but all of them reduce the quality of recommendation in several ways. First, a small customer data-set is less similar to the user. Second, item set partitioning is restricted to some specific areas. Third, if the items are segregated as popular and unpopular then they will never appear as recommendations.

1. CLUSTER MODELS

The goal of algorithm followed here is to assign the user to the cluster comprising of similar customers. In order to find the customers similar to the user the entire customer base is divided into segments. Further, the purchase and ratings of the customers in a particular segment are used to generate recommendations.

The segments can be generated on the basis of a few learning algorithms or through manual detection. Using this similarity, the clustering algorithm follows greedy approach to expedite the process. The algorithm starts with an initial base with one customer selected from each cluster. The customer base is repeatedly compared and often, new segments are created either by segregation or merging.

Some algorithms classify the user into more than one segment and justify the strength of the relationship. Cluster models work better than collaborative filtering as the comparison is restricted and not done with the entire base. All the customers in a segment are considered for making the recommendation because of the similarity. The expense for both the above mentioned algorithms turns out to be almost same.

1. SEARCH BASED METHODS

Search-based algorithm treats the problem by constructing a query to fetch the popular and relevant items of the same traits and genre. It works well if the purchase history of the customer is small because the algorithm needs to be scaled properly for efficient performance. In case the purchase history is large the performance quality of the system is not at par with the requirement. A recommendation system should ideally be able to help the customer fetch relevant, unique and interesting products. The system which we have right now fails to satisfy this goal.