基于特征和项目近邻的混合推荐算法研究

Research on Hybrid Recommendation Algorithm Based on Feature and Item Nearest Neighbor

针对传统的协同过滤算法在推荐过程中存在的可扩展性差、推荐准确性低等问题,提出了一种基于动态加权的混合协同过滤算法(ItemBase_ALS collaborative filter,IACF)。该算法将基于项目的协同过滤算法(ItemBase CF)与基于矩阵分解的ALS推荐算法按照一定的权重进行混合,并在分布式平台Spark上得以实现,有效解决了算法扩展性问题。该混合算法首先分别利用ItemBase CF和ALS算法进行初步预测,然后选取能够反映其各自特性的因素,即项目近邻和隐藏特征,按照权重公式进行融合从而得到最终预测结果。通过调整权重比例,可以突出某一算法的特性,满足不同的推荐需求。实验选用MovieLen电影评分数据集,实验结果表明,混合协同过滤算法较之传统单个算法,既能体现其各自特点及变化规律,在可扩展性、准确性上也有所改善。

A hybrid collaborative filtering algorithm (ItemBase_ALS collaborative filtering,IACF) based on dynamic weighting is proposed to solve the problems of poor scalability and low recommendation accuracy in the traditional collaborative filtering algorithm. The algorithm combines the item-based collaborative filtering algorithm (ItemBase CF) and the matrix factor-based ALS recommendation algorithm according to certain weights and is implemented on the distributed platform Spark which effectively solves the problem of scalability. The hybrid algorithm first uses ItemBase CF and ALS algorithms to make preliminary prediction respectively,and then selects the factors that can reflect their respective characteristics,that is,item nearest neighbor and hidden feature,and fuses them according to the weight formula to get the final prediction results. By adjusting the weight ratio,the characteristics of an algorithm can be highlighted to meet different recommendation requirements. The experiment on MovieLen dataset shows that the hybrid collaborative filtering algorithm can not only reflect their own characteristics and change rules,but also improve the scalability and accuracy.