Research On Recommendation Algorithm Based On Comment Text

As the carrier of users' questions,suggestions and attitudes,the comment text of the recommendation system is of great value to discover users' preferences.Deep collaborative neural network DeepCoNN,transform network TransNets and neural rating regression tip generation model NRT all use comment text to predict user preferences.TransNets and NRT also use comments to provide explainable text to users.However,the above model does not distinguish the importance of comments,resulting in the "noise" of unimportant comments,which affects the accuracy of user preference prediction and the quality of explainable text.This thesis proposes a multi-task algorithm,ERR,which is based on the comment text.Aiming at the comment "noise" problem,different weights are assigned according to the importance of the comment,so as to reduce the interference of the comment "noise".Aiming at the problem of recommendation interpretability,a sequence-to-sequence recommendation interpretation generation model is constructed according to the user's comments on different projects and the comments received from different users.Meanwhile,the comment denoising method proposed in this thesis is used to improve the quality of the generated explainable text.Comparing TransNets and DeepCoNN with ERR to predict user preferences in datasets in three different fields,the experimental results show that ERR's prediction accuracy is improved by an average of 2.5% to 5% compared with both,and the improvement is larger on more sparse datasets.In the comparison between TransNets and NRT and ERR in generating explainable text,the generated text similarity of ERR has an average improvement of about 20%,and the generated explainable text examples are of good generality and readability.