基于评论情感向量化表示的双向跨域推荐

Dual Cross Domain Recommendation Based on Sentiment Vectorization of Reviews

摘要: 多数跨域推荐模型在利用评论文本进行跨域推荐时，没有考虑到评论文本中的情感信息。而对于考虑到评论情感信息的跨域模型多数都是进行单向跨域推荐。因此，本文提出一种基于评论情感向量化的双向跨域推荐模型。该模型对不同领域的评论文本利用bert + Transformer方法进行评论情感分类，得到隐含用户情感信息的情感向量化表示，再从中计算相对应的用户偏好向量和商品特征向量，然后根据生成的特征向量进行跨域推荐。对于跨域学习，本文在神经因子分解机(NFM)模型的基础上利用潜在正交映射函数学习两个域内的用户偏好以及跨域学习到的用户偏好，从而进行双向推荐。通过在四组数据集上进行对比实验，实验结果表明，该模型能有效缓解了数据稀疏的问题。

Abstract: When most cross-domain recommendation models use review text for cross-domain recommendation, they do not consider the emotional information in the review text. For cross-domain models that take into account the emotional information of comments, most of them are one-way cross-domain recommendations. Therefore, this paper proposes a dual cross-domain recommendation model based on comment sentiment vectorization. The model uses the bert + Transformer method to classify the review emotions of review texts in different fields, and obtains the emotion vectorized representation of the implicit user emotion information, and then calculates the corresponding user preference vector and product feature vector from it, and then cross domain recommendation is carried out according to the generated feature vectors. For cross-domain learning, based on the neural factorization machine (NFM) model, this paper uses the latent orthogonal mapping function to learn the user preferences in the two domains and the user preferences learned across domains, so as to perform two-way recommendation. Through comparative experiments on four datasets, the experimental results show that the model can effectively alleviate the problem of data sparseness.