融合评论文本与评分交互特征的推荐方法

Recommendation method combining interactive features of review and rating matrix

为缓解协同过滤推荐算法中评分数据稀疏问题对推荐结果的影响,提出一种融合文本评论和用户评分交互的推荐算法。通过将用户和商品评论各自潜在主题向量与用户、商品的潜在因子向量进行融合并各自进行评分,经过动态线性加权融合做出整体评分预测。在公开的多组数据集上,以推荐结果的均方根误差(RMSE)和平均绝对误差(MAE)为评估指标进行实验验证。实验结果表明,提出算法可以更好地刻画用户偏好和商品特征,有效缓解了评论数据稀疏性影响,提高推荐结果的准确性。

To alleviate the impact of sparse scoring data in the collaborative filtering recommendation algorithm on the recommendation results, a recommendation algorithm that combining text comments and user ratings interaction was proposed. The latent topics of user and product reviews were fused with the latent rating factors of user and product and scored respectively, the overall score prediction was obtained by dynamic linear weighted fusion. Root mean square error(RMSE) and mean absolute error(MAE) were used as evaluation indicators to experimentally verify the recommended results. Experimental results on open datasets show that the recommendation algorithm proposed can better characterize user preferences and item characteristics, alleviate the problem of sparsity of review data, and improve the accuracy of recommendation.