Recommendation system is an significant study goal in the realm of information filtering system. The recommender system predicts the items that users are interested in depended on the user’s past operational data. Dynamic behavior sequence can be extracted by self-attention mechanism, and most models assume that the interaction history is regarded as an ordered sequence without considering the time interval between each interaction. Our input entities and items are both interconnected and highly correlated in the knowledge graph and recommendation modules respectively. This paper fuses them to recommend to users, and proposes a multi-task feature learning recommendation model that fuses time interval and knowledge graph, explicitly modeling interaction timestamps within a sequential modeling framework, and fused with knowledge graph embedding (KGE) to assist with the recommended task. The experimental results show that on the real dataset MovieLens1M, the AUC, ACC, Precision and Recall indicators are used for evaluation, and the proposed model is better than the mainstream benchmark models.