基于交互注意力的可解释性推荐方法

AN INTERPRETABLE RECOMMENDATION BASED ON INTERACTIVE ATTENTION

目前基于评论的推荐算法大多采用静态独立的方法提取用户和物品评论的潜在特征表示,将用户的偏好表示为静态特征向量,而用户在与不同的物品交互时通常表现出不同的偏好。因此,提出一种基于交互注意力的可解释性推荐方法,利用交互注意力来研究用户评论和物品评论之间的相关性。另外,受LSTM中门控制的启发,模型还增加门控层来自适应地合并两边网络提取出的特征向量,利用注意力因子分解机进一步对高阶特征交互进行建模,实现评分预测。通过注意力权重衡量评论信息,提高预测评分的可解释性。实验结果表明,推荐结果准确性进一步提高,验证了该方法的有效性。

At present,most of the review-based recommendation algorithms use static and independent methods to extract potential feature representations of user and item reviews,and represent user preferences as static feature vectors.Users often show different preferences when interacting with different items.Therefore,this paper proposes an interpretable recommendation method based on interactive attention,which uses interactive attention to study the correlation between user reviews and item reviews.In addition,inspired by the gate control in LSTM,the model also added a gate control layer to adaptively merge the feature vectors extracted from both sides of the network,and used the attention factorization machine to further model high-order feature interactions to achieve score prediction.Review information was measured by attention weight to improve the interpretability of prediction scores.The experimental results show that the accuracy of the recommendation results is further improved,which verifies the effectiveness of the recommendation method.