Modelling and efficiency analysis of electronic recommendation systems

The aim of this project was to make an implementation of a non–content based recommendation system, measure its performance and compare it with similar recommendation systems used globally in eCommerce. Baseline predictor through least squares and collaborative filtering were used as main techniques for our implementation. Implementation was tested on real life data acquired from [http://www.grouplens.org/datasets/movielens/]. Data is made up of discrete ratings from 1 to 5 given by users to movies that they watched. Data is divided into a training dataset and testing dataset. Training dataset was used by our recommendation system to train itself and acquire predictive insight. Testing dataset was used to evaluate the efficiency of the recommendation system by calculating root mean square error (RMSE). First, baseline predictor was implemented by solving least squares problem and then collaborative filtering was introduced to improve the predictor. Collaborative filtering exploits movie-movie correlation in the given data, to improve the values of baseline predictor. After running RMSE calculations on baseline predictor using testing dataset, RMSE = 0,900 was achieved. The following test resulted in 42,71% right predictions, 49,16% predictions off by 1 point, 8.19% predictions off by 2 points, 0,8% predictions off by 3 points and 0,04% predictions off by 4 points. After improving the baseline predictor with collaborative filtering and running the same RMSE calculation, RMSE = 0,876 was achieved. The following test resulted in 45,13% right predictions, 47,3% predictions off by 1 point, 7,71% predictions off by 2 points, 0,72% predictions off by 3 points and 0,04% predictions off by 4 points. By adding collaborative filter to the baseline predictor model, RMSE was improved by 2,67% and number of correct predictions increased by 2,42%. Recommendation system used by Netflix has RMSE = 0,8572 [http://www.netflixprize.com], comparing it with our systems RMSE = 0,876 it is seen that our systems RMSE is only 2,19% greater than the RMSE of Netflix recommendation system. This yields the conclusion that our implementation works properly and could be deployed in a similar to Netflix eCommerce application