基于神经因子分解机的推荐多样性提升方法

Recommendation Diversity Improvement Method Based on Neural Factorization Machine

神经因子分解机模型很好的解决了数据稀疏场景下的点击率预测问题,但仅仅关注于预测准确性导致该模型的推荐多样性效果不佳。针对上述问题,提出一种基于神经因子分解机的推荐多样性提升方法。方法通过用户-项目的交互历史构建多样性输入矩阵,利用用户活跃度和项目流行度对多样性输入矩阵进行修正,并以不同方案将多样性输入矩阵融入神经因子分解机模型,以此作为附加信息源来增强神经因子分解机模型的多样性表达能力。实验结果表明,在MovieLens、Film Trust及Book-Crossing三种稀疏程度不同的数据集上,所提出的方法均能在推荐准确性小幅度损失的情况下,较大幅度的提升推荐列表的多样性。

The neural factorization machine model solves the problem of click-through-rate prediction in the scenario of sparse data,but only focusing on the prediction accuracy leads to the poor recomm endation diversity effect of this model.Aiming at this problem,this paper proposes a diversity enhancement method ba sed on neural factorization machine.This method constructs a diversity input matrix through the interaction history of users and items,and uses the user activity and item popularity to modify the diversity input matrix.The diversity input matrix is integrated into the neural factorization machine model in different schemes,which is used as an addition al information source to enhance the diversity expression ability of the neural factorization machine model.The experimental results show that on three datasets with different degrees of spa rsity,MovieLens,Film Trust and Book-Crossing,the proposed method can greatly improve the recommendation diversity with a small loss of recommendation accuracy.