Knowledge graph (KG) enhanced recommender systems (RSs), i.e., KGRSs, have received extensive attention recently. However, the modeling strategy in existing KGRSs tends to excessively depend on the feature data of KG component or the historical data of RS component, which affects the performance of KGRSs in two aspects. First, as the KG data is usually incomplete and contains noise, excessive dependency on KG data limits the expressive ability and damages model robustness. Second, excessive dependency on the historical data in RS also limits the expressive ability because historical data is not enough to model users’ complex and changeable behavioral preferences. In order to solve these issues, this paper proposes a robust yet effective twopart modeling (TPM) strategy to reduce the strong dependency on KG feature data and RS historical data while make full use of multi-source data from both the KG and the RS components. That is, both the users and items in KGRSs are modeled by two parts information, i.e., a context part and an auxiliary part. The TPM is a general strategy that can be simply implemented into existing KGRSs to further improve models’ expressive ability and robustness. This paper applies the proposed TPM strategy to a typical KGRS named multitask feature learning for KG enhanced recommendation (MKR), to obtain TMKR. Experimental results show that TMKR can not only obtain a great prediction accuracy and recommendation stability, but also has a significant time cost reduction compared to MKR.