Research On One-class Collaborative Filtering Based On Confidence Weighting

One-class collaborative filtering(OCCF)refers to collaborative filtering algorithms that only use implicit feedback for recommendation.Although the traditional collaborative filtering algorithms show outstanding performance on the recommendation of explicit feedback such as scoring,they are still facing problems such as poor interpretability,sensitivity to noise,and poor scalability in one-class problem.Aiming at addressing the above problems,a new one-class collaborative filtering recommendation algorithm based on confidence weighting was proposed.In the construction phase of the model,a confidence function was proposed for learning users' behavioral confidence level,which mapped the users' selection tendency in the behavior frequency domain to the confidence probability,and the function could be used as an independent unit to learn the credibility of the users' behavior.The confidence function was embedded into the implicit feedback recommendation model(IFRM)to form an one-class collaborative filtering recommendation algorithm based on confidence weighting,which was named the Confidence-Weighted Implicit Feedback Recommendation Model(CWIFRM).In the optimization stage of the model,a heterogeneous confidence optimization algorithm based on stochastic gradient descent(SGD)with different optimization in different directions was proposed.In order to verify the effectiveness of the algorithm,the experiments were conducted on three different datasets for CWIFRM together with multiple algorithms.During the test phase,each algorithm was tested with streaming data simulation recommendations for TopN recommendation.Experiments show that CWIFRM can significantly improve the recommendation quality in terms of recommendation accuracy,and the application of the heterogeneous confidence optimization algorithm can further improve the recommendation quality.The results prove that the proposed confidence function helps to improve the recommendation quality and provides good interpretability for the recommendation.Heterogeneous confidence optimization algorithm further enhances the interpretability and the application scope of the model.The recommendation of the model is stable,and the data noise can be weakened to some extent.With the linear weighted form and a wide range of parameter customizations,the model ensures good scalability.