Collaborative filtering with temporal dynamics

Recommendation systems are the systems implementing techniques discovering information in order to make personalized recommendations during product and service interaction. Today, it is important to make custom tailored recommendations, keep customer interest, predict customer preferences and make recommendations accordingly. Collaborative filtering-CF is an effective and easy method used by recommendation systems. Generally CF is the process of filtering information using various techniques among the information sources. In addition, CF is used in a differential area for example; such as identification of data in multiple sensors, data detection, financial data, e-commerce and web applications. Together with offering various recommendations to its customers in accordance with their interests CF meets several difficulties upon the continuous change of the preferences of the customers and the products. New product and services may emerge that change the perception of the customers and the popularity of the products. Seasonal changes or the regional factors may change the perception of the person [1]. For example; user may gain a new point of view in movie recommendation systems according to the type of actor or movie and they can also change the type of movie they prefer in time. Traditional CF techniques tries to estimate the tendency of the users by considering the points they gave in the past for the data sets consisting of numerous user and elements starting from the similarity of the users to each others (user similarity) or the similarity of the products to each others (item similarity). However, these CF techniques remain incapable to respond to the demands of the customers that change in time. In addition every product and customer undergoes different changes. A common model is required that will determine the behaviors of the users, show their time changes for every product and customer. Together with this study the accuracy and the attribute of the recommendations given to the user by analyzing the interest of the users depending on time in order to provide suitable personalized recommendations to the pleasure and appreciation of the users are increased. In this study, the point that customer given the point to that product had tried to be estimated not only according to the points of the users originally, but considering the dates they gave the points. The older a product the more it was given points by that many persons. However, since the habits of the person would change in time, the point they give close to today is more important than the point they gave in the past. In other words, there is the age of the point given for each product by the users. Therefore, on behalf of correcting the deficiencies in traditional CF technique the existing points in this study had planned to decreased according to the age of the points (if the points are old) or strengthened (if the points are fresh). In this study, action had taken over the data of Netflix through the points collected in the last two years in order to measure the success of the methods developed. Netflix is a data set taken from a company established based on movie and video rental in America and used frequently in recommendation systems. The method considered in this study will be performed over the natural data set whose point dates are certain that were given by half a million Netflix customers from 1999 to 2005 for 17770 movies. In addition, in this data set there are records over 100 million that were given points. Moreover, SVD of size reduction methods was measured system performance by applying both netflix and movielens to the data sets. MovieLens data set consists of data collected by GroupLens at the Minnesota University. It is composed 1682 film's of scores given 943 users. It is available user's age, gender, occupation as well as demografic information such simple categorical information on the film types. In order to measure the system success, the results that obtained by the method to be formed by traditional CF and adding time had compared. Due to the system to be developed, a product recommendation system had designed that could produce result depending on the interest and the pleasures of the users that could change in time. Thus, the confidence of the user to the system will increase. If this is an e-commerce site, the company profitability will be increased by keeping the interest of the customer live because of recommending the proper product. Keywords: Data Mining, Recommendation Systems, Collaborative Filtering