融合BiGRU和记忆网络的会话推荐算法

Session-based recommendation algorithm combined with BiGRU and memory network

为消除噪声、充分使用邻域信息并考虑用户动态兴趣,提出一种融合BiGRU和记忆网络的会话推荐模型。使用BiGRU捕获会话总体特征,使用另一个加入缩放点积自注意力的BiGRU消除噪声项的干扰并捕获细粒度的用户兴趣,使用记忆网络通过邻域会话信息预测当前的会话意图,改进融合选通门进行特征融合并计算每个候选项的推荐分数。通过在两个数据集上的实验,验证了该模型能够准确预测用户意图,提高推荐效果。

To eliminate noise, make full use of neighborhood information and consider users’ dynamic interests, a session-based recommendation model combined with BiGRU and memory network was proposed. BiGRU was used to capture the overall cha-racteristics of the session. Another BiGRU with scaled dot-product self-attention was used to eliminate the interference of noise items and capture fine-grained user interests. The memory network was used to predict the current conversation intention through the neighborhood conversation information, and the fusion gate was improved for feature fusion, and the recommendation score of each candidate was calculated. Experimental results on two data sets show that the model can accurately predict user intention and improve the recommendation effect.