Recommendation Algorithm Based On Blending Learning

Recommendation systems in today's world are extremely important for any business and for the users.Recommendation Systems plays a vital role for businesses to increase their revenue;the study shows that business increased their revenues up to 30% by simply implementing Recommendation System Algorithms.Recommendation Systems are Information Filtering Systems,they provide a user customization,according to their own preferences for any product or item.Business organizations uses recommendation systems to learn the behavior of users to understand and learn the market desire and demand which later on used by those business to start marketing campaigns and targeting the audience by true products according to their needs.Therefore RSs helps the businesses to save time,cost and money to make smart decisions and allow businesses in investment for increasing revenue such as advertisement.Businesses such as e-commerce websites like TaoBao,Alibaba and JingDong increased their revenue 35% by simply using recommendation systems which increases the cross sells.Despite many advantages of recommendation systems,getting the accuracy and true predictions are still a bit challenge due to the nature of the data,data sparsity and other factors such as lake of features available.Matrix Factorization is the most popular and widely researched technique.Matrix Factorization uses the dot product which does not satisfy the inequality property.Therefore different techniques proposed to solve the problem such as Metrix Factorization.Although Metric Factorization improved results over Matrix Factorization and solved the dot product inequality property problem.But there is always welcome for new research work to be carried out.Therefore we propose a multi-model ensemble learning technique called blending.This technique consists of two steps.First we train several base models and get the predicted rating of movies,then use a liner regression to combine these results as a second-layer model to get a final rating of movies.The metrics Root Mean Square Error(RMSE)and Mean Average error(MAE)are used for evaluation of different models.Our experimental results indicate that new blending approach is superior to other used techniques.We used models such as SlopeOne,SVD,SVD++ and Metric Factorization(MetF)for our comparison.We conduct our experiments on two publically available datasets i.e.Film Trust and Movie Lens to predict the ratings.We summarizes our results by combining three different models on filmtrust dataset i.e.SlopeOne,MetricF and SVD in first experiment and SlopeOne,MetricF and SVD++ in second experiment we also repeat these same experiment on MovieLens Dataset.Our Model performs superior on both of datasets with respect to MAE and RMSE.Furthermore the research work contains detailed understanding of old and new techniques used in the area of RSs along with algorithms used and different evaluation matrices used for results evaluation in perspective view of Machine Learning and Human Computer Interaction.