The personalized recommendation method is an efficient and effective solution to optimize educational resources, which can effectively extract learners’ preference characteristics and select suitable high-quality educational resources for them to improve teaching efficiency. In this paper, we construct a confidence-aware recommendation model for educational resources. In order to reduce the sparsity problem of rating data, a constraint method based on the implicit factor of user-object interaction is designed to construct the implicit factor of item-user interaction by using the information of a single review. Meanwhile, a confidence matrix is introduced to improve the accuracy of recommendations, and a loss function is constructed using the maximum a posteriori estimation theory. The loss function is optimized by introducing a small batch gradient descent algorithm at the end. Two datasets are used to verify the model’s performance, and the impact on students’ innovation ability is assessed in teaching practice. It is found that the model in this paper outperforms the other three classification models, and the test results on the two datasets of online_shopping_10_cats and Course Classify achieve an accuracy of more than 0.95. The experimental class’s innovation ability increased by 37.6% from 46.25 to 63.63, while the control class’s pre- and post-scores of 46.36 and 47.88 were not significantly different. It can be seen that the experimental class has progressed differently in the four dimensions after applying the model of this paper. This study proposes new ideas and feasible paths for combining the optimal allocation of educational resources with cutting-edge information technology, empowering students’ innovation ability through big data technology.