Abstract: [Objective/Significance] Ripple Net recommendation model is optimized by modifying noise and changing preference diffusion rules to improve recommendation accuracy. [Method/process] To identify and correct the noise in the scoring data set by referring to the idea of "the users participating in the evaluation at the same level and the evaluated items at the same level"; A preference diffusion rule is developed to preferentially select knowledge graph nodes connected by semantic relation of high word frequency and a user preference data set is constructed. [Result/Conclusion] The average accuracy of different K values of the optimized model is 5% higher than that of the original Ripple Net model in the Top-K scenario. The average AUC value of the test set increased from 90.7% to 92.3% under the CTR click rate prediction scenario. The performance of Ripple recommendation model can be improved by modifying the noise in data set and improving the preference diffusion rule.

Keywords: knowledge map; RippleNet recommendation model; Preference diffusion; Noise filtering;