AN EFFECTIVE OPTIMIZED RECOMMENDER SYSTEM BASED ON FEATURE SELECTION TECHNIQUE

Abstract

Task of providing recommendations is achievable through recommender systems (RS). There are various classifications of RS i.e. content, collaborative and hybrid based. Apart from these three, there is one context aware RS. Because, through research it has been proved that, ratings provided by user for a movie or TV program also depends on demographic information, day type, time, mood, and other factors of environment. Considering all these contextual features and providing recommendation increases the computational time complexity of our program to a very high value. Apart from that, RS also suffer from cold start problem, scalability problem and data sparsity problem. Hence, artificial neural networks are used to provide recommendations and make this task achievable. Now, training of Artificial Neural Networks takes a very large time, due to high dimensional features. Because, the number of contextual feature attributes could vary from 24 to 35. In order to reduce time complexity, a perfect subset of these attributes should be considered. But, again reducing such a high dimensional contextual attributes is a kind of combinatorial optimization problem. This problem could be solved by using Ant Colony Optimization (ACO). ACO with heuristic information as either covariance or fuzzy values are used as heuristic function. At the end, back propagation algorithm is used, for training the neural network, only on those feature subset obtained via ACO with covariance or fuzzy c means measures. While testing the RS, around 80% of the data set is classified as training data set and rest 20% of the data set is classified as testing data set. Accuracy of recommender system is determined, on testing data set. Finally, mean absolute error has also been calculated and results are analysed by comparing the accuracy of recommender system with previous approaches.