Timeliness Sensitive Collective Matrix Factorization Personalized Recommendation

Hai-yan ZHAO,Ying WANG,Qing-kui CHEN,Jian CAO
DOI: https://doi.org/10.3969/j.issn.1000-1220.2017.09.020
2017-01-01
Abstract:Timeliness is a very important factor for many products in market. Generally,people choice new products and out-of-dated products based on different reasons. Therefore,these differences should be considered in recommendations. Unfortunately,most current recommendation approaches don′t take this factor into consideration. Based on the analysis of product timeliness,a timeliness-sensitive recommendation approach is proposed. It is based on the collective matrix factorization algorithm,which divides the products into sev-eral matrices,sharing parameters among features when an entity involves in multiple relations. Then,using the gradient descent algo-rithm to train them together and measures the error between the real values and predictions. This approach not only considers the inner relations between products with different timeliness,but also overcomes the problem of sparse in other recommendation algorithms. Simultaneously,it can find the implicit relations between the divided product sets through discovering characteristic vectors in a more comprehensive way. The approach uses the characteristic of products as regularization,which can cope with over-fitting problem which arises during model training on the sample data. Moreover,the approach also uses popularity as another regularization in the model, which helps make the result to be better and more accurate. The experimental results show that this approach achieves better perform-ances against other ones. Therefore,the timeliness-sensitive recommendation approach achieves higher prediction accuracy than facto-ring each matrix separately and demonstrates its superiority,as well as the benefit of considering the timeliness of products.
What problem does this paper attempt to address?