An improved recommender system based on normalization of matrix factorization and collaborative filtering algorithms
Recommendation System (RS) came to lime light when the information on the internet started growing to the extent that it became time consuming to get the required information. There are different techniques used in RS. Some works are based on user past knowledge known as Content Based (CB) while mor...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2015
|
| Subjects: | |
| Online Access: | http://psasir.upm.edu.my/id/eprint/65551/1/FSKTM%202015%2025IR.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Recommendation System (RS) came to lime light when the information on the internet started growing to the extent that it became time consuming to get the required information. There are different techniques used in RS. Some works are based on user past knowledge known as Content Based (CB) while more popular techniques referred to as neighborhood models (CF and MF) are based on finding similar users for recommendation. Existing techniques have certain drawbacks such as user getting the same information. This problem is known as stability versus plasticity (in CB). Another problem called cold start gives wrong recommendations amongst new users as data of new users is not enough for recommendation. Other limitations include too much dependence on other users or no consideration of user personal preferences (CF and MF). There is a technique known as normalization which develops models like user involvement in subject matter or user likeness according to the details of item
to predict ratings to user. Normalization shows good results but it is truly personalized from single user perspective and lacks other user’s opinion for the recommendation. Some researchers combine different techniques into hybrid to overcome the problems in RS, but there is very limited work that has investigated the effect of hybridizing normalization technique on neighborhood models. Therefore, this research is dedicated to combining the normalization technique with neighborhood models (CF and MF) to produce
CF+N (collaborative filtering and normalization) and MF+N (matrix factorization and normalization). The hypothesis is that the tendency of normalization technique to simplify the data combined with the accuracy of the neighborhood models can improve the accuracy of the RS. This hybrid technique rates user personal preferences more than other user’s recommendation towards the final recommendation, while still considering user’s personal recommendation as important input in the process. Several experiments have been conducted on the movielens dataset where 80% of data is used as training set while 20% is used as test set. The experiments are designed to perform the comparisons with the existing works that target to solve the existing problems in RS. There are three categories of evaluation of RS predictive accuracy metrics, classification accuracy metrics and rank accuracy metrics. This study follows MAE and RMSE from predictive accuracy metrics for evaluation of results since the main focus of the study is to reduce errors in
RS. Results show that MF+N unite well as hybrid technique where the gray sheep is handled by MF and normalization manages cold start, mood changes, stability versus plasticity and difference of opinion. On the contrary, CF+N technique requires some enhancements as the results were below expectations because of the tendency of CF to produce big differences in the prediction of raw data. It is concluded that the resultant hybrid techniques can perform well if the variables provided to normalization by neighborhood model (MF and CF) do not have big differences in order for the hybrid normalization model to outperform every algorithm in comparison. |
|---|