Research On Collaborative Filtering Based On Similarity Of Item Attributes

Increased content on the web has impacted the data overload and created diversed information.A huge amount of data and various items lead users into confusion in terms of decision making.With poor decision making,certain information could be lost and bizarre.The application of recommendation system is the solution to the problem of filtering information and predicting the right decision for users.Collaborative Filtering is one of the popular recommender systems techniques which had been greatly improved since its invention.Collaborative Filtering technique principle is based on the identification of other neighbors' similar ratings.A Recommender system does not immediately simplify decision making process of the user because challenges and problems exist in the recommender system itself,such as cold start,data sparsity,scalability,Gray Sheep,Shilling Attack,Diversity,and Long Tail.Based on above identified problems and recent research,this paper focuses on the utilization of item attributes similarity.By comparing the result of item attributes similarity with several techniques that have been invented before,this research studies the advantages and disadvantages of item attributes prediction.In order to able to explore the features of item attributes similarity,we analyzed the proposed method into several common challenges of recommendation system as listed before.The result shows that proposed method is able to handle Gray Sheep,Shilling Attack and Diversity and Long Tail problems better than the compared traditional recommender systems.Proposed methods are applied in three phases of recommender system,which are Information Collection Phase,Learning Phase,and Predicting / Recommendation Phase.The information collection phase is the data preprocess phase.Purpose of this phase is to produce the required dataset from the available dataset.In the learning phase,system examines the similarity between users based on the distance similarity and genre similarity.The distance similarity is calculated by using Euclidean algorithm and the genre similarity is calculated using Jaccard Index Similarity.Lastly,in the Predicting/Recommendation phase,the system will calculate the item prediction based on the similarity and rating of other users.The experimental evaluation of the system is testing the proposed methods CFSIA-1 and CFSIA2 compared to other methods.By using Mean Absolute Error and Root Squared Mean Error metric measures,tests are performed on several environments.Test of the proposed methods is resulted in the method strength and weakness analysis.Further examination of the proposed methods has created the optimization of CFSIA-1.In conclusion,the optimized method of CFSIA-1 has better accuracy compared to several traditional methods.Optimized CFSIA-1 is also proved able to face several recommendation system challenges better than any other methods.