A Study of Multi-Task Recommendation Models Incorporating Preference Propagation

To address the problem that knowledge graph can effectively reduce the triadic relationships of entities from multi-source heterogeneous data, but is not conducive to recommendation tasks and it is difficult to explore the potential association relationships of data using single-task learning, a multi-task recommendation model with fused preference propagation (MAPKR) is proposed. Firstly, the user’s preference feature set is extracted from the knowledge graph using ripple propagation; secondly, the potential features are shared based on the similar nearest neighbor structure, and the higher-order feature representations of items and entities are extracted by cross-compression units; finally, the recommendation module and the knowledge graph embedding module are trained alternately with multi-task learning, and the extracted feature vectors are predicted and recommended after normalized inner product operation. Experiments are conducted on three publicly available datasets and compared with five baseline models. Compared with MKR and RippleNet, the AUC improved by 2.53% and 2.66% on average on the three datasets; ACC improved by 3.07% and 3.46% on average. The results show that the model presented in this paper has good recommended performance.