联合训练下融合编解码器的序列推荐算法

Sequence recommendation algorithm for encoder-decoder fusion under joint training

现有基于Transformer的推荐算法通常仅考虑使用编码器来进行推荐预测,缺乏利用解码器去“解码”用户行为序列的能力,不能较为准确预测用户下一次的交互行为。为解决此问题,基于阿里巴巴电子商务推荐的行为序列模型(BST)提出联合训练下融合编解码器的序列推荐算法模型BSTEAD。通过采用联合训练机制,设置Transformer预测任务和BST预测任务。将两条预测任务的损失进行加权求和,得到最终的损失函数。在MovieLens和Goodbooks两个公共数据集上的实验结果表明,BSTEAD推荐算法与5个对比模型相比性能具有显著提升,验证了联合训练机制下解码器对推荐任务的有效性。

The existing Transformer based recommendation algorithms usually only consider using encoders for recommendation prediction,lacking the ability to decode user behavior sequences using decoders,and cannot accurately predict the next user interaction behavior.To address this issue,a sequence recommendation algorithm model BSTEAD was proposed based on the behavior sequence model(BST)of Alibaba’s e-commerce recommendation,which integrated encoder-decoder under joint trai-ning.By adopting a joint training mechanism,Transformer prediction tasks and BST prediction tasks were set up.The losses of the two prediction tasks were weighted and summed to get the final Loss function.Experimental results on two common datasets,MovieLens and Goodbooks,show that the proposed BSTEAD recommendation algorithm significantly improves the performance compared to five comparative models,verifying the effectiveness of the decoder for recommendation tasks under the joint training mechanism.