Research On Recommendation Algorithm Based On Collaborative Filtering And Deep Learning

As an information discovery tool in the era of big data,recommendation systems are widely used in various fields of production and life.Recommendation algorithms are the soul of recommendation systems,which can intelligently calculate expected results using existing data.The collaborative filtering algorithm is a commonly used algorithm in the recall phase of the recommendation system.However,due to sparse data and serious lack of scoring information,the similarity calculation results have a large deviation,and the prediction accuracy of the system is low.The user’s rating information cannot fully represent the user’s wishes.The collaborative filtering algorithm makes full use of the fixed rating information to predict the recommendation results,and its generalization ability is poor.Recommendation algorithms based on deep learning are used in the ranking stage of recommendation systems.By collecting a large amount of feature information and deeply mining the features,potential features are discovered and used for recommendation result prediction to improve the accuracy of the recommendation system.The feature representation and processing methods of existing deep learning recommendation algorithms are single and cannot accurately predict recommendation results.This thesis studies the collaborative filtering algorithm in the recall phase and the deep learning recommendation algorithm in the sorting phase of the recommendation system,and proposes two recommendation algorithms.(1)Fuzzy Genetic Two Step Collaborative Filtering(FGTSCF)based on fuzzy genetic algorithm.For the problem of data sparsity,the FGTSCF model utilizes trust networks to obtain the shortest path length and user centrality between users,and integrates the shortest path length,user centrality distance,and user rating differences to serve as the similarity between users.To solve the problem of weak generalization ability of collaborative filtering,fuzzy logic is used to fuzzy process the scoring information to obtain fuzzy vectors,and cosine similarity and Pearson correlation coefficient are combined to supplement user similarity to expand the search scope of the model.To improve the accuracy of model prediction,genetic algorithm is used to adjust the similarity between users,search for the optimal solution of similarity between users,and find the predicted score value that is closest to the target score.Conduct offline experiments on the open-source datasets Epinions and Ciao DVD to verify the impact of different parameters on the model.By adjusting the model with the optimal parameters and conducting comparative experiments,the overall performance of the FGTSCF model is better than that of other typical recommendation models.(2)Feature Refinement Depth Field aware Factorization Machine(FRDFFM)for feature reconstruction.The FRDFFM model excavates implicit feature information from shallow and deep layers,crosses artificial features from both feature and domain aspects,and uses deep learning technology to discover deep representations of features.Using feature reconstruction technology to construct supplementary features of the original features,introducing contextual information to adapt to different environments.Offline experiments were conducted on open-source datasets Criteo and Avazu,and the experimental results verified the feasibility and practicality of the model.According to comparative experiments,the proposed FRDFFM model has higher prediction accuracy compared to the current mainstream deep recommendation models.