时域交互网络中动态嵌入轨迹预测方法

Dynamic embedding trajectory prediction method in temporal interaction networks

在电子商务、社交网络和教育等领域对用户和物品之间的顺序交互关系进行建模至关重要。通过表示学习模拟用户和物品的动态演化具有广阔应用前景,其中每个用户和物品均可以嵌入到欧几里得空间中,并且可以通过该空间中的嵌入轨迹对其演化进行建模。然而,现有的动态嵌入方法仅在用户采取某一动作并且未对嵌入用户和物品的未来轨迹进行建模时才会生成嵌入向量。针对这一问题,提出一种用户-物品耦合循环神经网络模型UICRNN(user-item coupled recurrent neural network model),学习用户和物品的嵌入轨迹。该模型使用2个循环神经网络来更新用户和物品交互关系的嵌入。此外,UICRNN模型可以对用户和物品的未来嵌入轨迹进行建模,引入了一种新的投影算子,该算子学习并估计将来可能嵌入的用户,估计的嵌入用户用于预测未来的用户和物品之间的交互关系。在多个真实数据集进行大量实验,预测未来交互的准确率至少提高42.91%,预测用户状态变化的准确率平均提高15.79%。

It is essential to model sequential interactions between users and items in areas such as e-commerce, social networking, and education. Representation learning is of broad application values by learning and simulating the dynamical evolution of users and items, where each user and object can be embedded in a Euclidean space, and its evolution can be modeled through embedding trajectories in the space. However, existing dynamic embedding methods only generate embedding vectors when users take some action and do not model the future trajectories of the embedded users and items. Aiming to cope with this problem, auser-item coupled recurrent neural network model(UICRNN) is proposed to learn the embedding trajectories of users and items. The model uses two recurrent neural networks to update the embeddings of interactions across users and items. In addition, the UICRNN model can also model the future embedding trajectories of users and items as well. A new projection operator is introduced, which learns and estimates the users who may be embedded in the future. These estimated embedded users are used to predict future interactions between users and items. Experiments are conducted in multiple real datasets and the results show that the accuracy of predicting future interactions can be improved by at least 42.91%, and the accuracy of predicting the state transmission of users can be improved by an average of 15.79%.