The task of session-based recommendation is predicting the next recommendation item when available information only includes the anonymous behavior sequence. Previous methods of session-based recommendation usually integrate the general interest, dynamic interest, and current interest to promote recommendation performance. However, most existing methods ignore the non-monotone feature interactions when building user’s dynamic interest and model item-item transitions through a linear way when building user’s current interest, which reduces the performance of model. In this paper, we design a novel method for session-based recommendation with recurrent and convolutional neural network. Specifically, The Gated Recurrent Unit with item-level attention mechanism learns the user’s general interest, while the convolutional operation with horizontal filter and vertical filter search for user’s current interest and dynamic interest. Moreover, the outputs of recurrent operation and convolutional operation are concatenated to generate the recommendation. Furthermore, we evaluate the proposed model on three real-world datasets which come from e-commerce and music API, respectively. The experimental results show that our model outperforms the state-of-the-art methods on session-based recommendation.

当可用信息仅包含匿名行为序列时，基于会话的推荐任务将预测下一个推荐项目。基于会话的推荐的先前方法通常将总体兴趣，动态兴趣和当前兴趣相结合以促进推荐性能。但是，大多数现有方法在建立用户的动态兴趣时会忽略非单调特征交互，而在建立用户的当前兴趣时会通过线性方式进行模型项的转换，这会降低模型的性能。在本文中，我们设计了一种新的基于递归和卷积神经网络的基于会话的推荐方法。具体来说，具有项目级关注机制的门控循环单元可以学习用户的总体兴趣，而水平滤波器和垂直滤波器的卷积运算则搜索用户的当前兴趣和动态兴趣。此外，将递归运算和卷积运算的输出连接起来以生成推荐。此外，我们分别在来自电子商务和音乐API的三个真实数据集上评估了提出的模型。实验结果表明，我们的模型优于基于会话的推荐的最新方法。