Since most recommendation systems are based on the calculation of items' or users' similarity, the results can't give consideration to both the complementarity and similarity of recommened objects. An algorthm framework for calculating the joint match degree of items for recommendation systems was proposed. In the framework, combining with the informations of users' interaction feedback, items' textual knowledge and structural knowledge, the joint match degrees of the target item and those candidate items were calculated respectively. Integrating the match degrees by using logistic regression, a list of items matched with the target item was obtained. Through the experiments on a Taobao real data set, it is indicated that the model significantly improve the performance of recommendation collocation compared to the recommendation algorithm based on similarity only. Moreover, in the situation of fewer users' interaction record, the model can also have better accuracy.