Research And Implementation Of A Hybird Recommendation System Based On Auto Encoder And Canopy-Kmeans Algorithm

Collaborative filtering is currently the most commonly used recommendation algorithm.Although the collaborative filtering recommendation algorithm has been widely used,there are still data sparsity problem and cold start problem.The problem of data sparsity is that users can only score very few products in a large number of commodities,the sparseness of the score data will be very high,and the sparsity of the score vector is also very high.The cold start problem is that the system doesn't have enough data of target users,which makes the recommendation algorithm unable to accurately analyze the interests and hobbies of the target users.These two problems will ultimately affect the effect of the recommendation algorithm.Starting from these two problems,this thesis completes the following work.1.This thesis proposes a user feature reduction algorithm based on autoencoder,which can alleviate the data sparsity problem and cold start problem in the recommendation algorithm.The autoencoder algorithm is often used to reduce the dimensionality of sparse features in deep learning.The autoencoder is introduced into the recommendation algorithm to reduce the dimensionality of the scoring vector,so that the reduced-dimensional vector has stronger expression ability and alleviates the data sparsity problems.The coding of the user's personal information is spliced with the scoring vector,and the spliced vector is dimension reduced using autoencoder to obtain a vector containing key information to solve the nearest neighbor set of the target user.The cold start problem is alleviated by using the original scores of other users in the nearest neighbor set to score the target user's unrated items.2.A hybrid recommendation algorithm based on autoencoder user feature reduction algorithm and Canopy-Kmeans clustering algorithm.The Canopy-Kmeans clustering algorithm is introduced to reduce the similarity calculation times when solving the nearest neighbor set by using the high similarity within the cluster and the low similarity between clusters,thus improving the efficiency of solving the nearest neighbor set of the target user.This thesis compares with other recommended algorithms on the Movielens dataset.The experimental results show that the proposed hybrid recommendation algorithm can effectively alleviate the data sparsity problem and cold start problem,and the hybrid recommendation algorithm has better effect than other recommended algorithms in the experiment.3.Based on autoencoder and Canopy-Kmeans hybrid recommendation algorithm,a movie recommendation system is designed and implemented with Django framework.The thesis analyzes requirements of system,and designs the architecture of the system.The entire system is divided into presentation layer,business logic layer,and data storage layer.The presentation layer consists of foreground system and management system.The front-end system includes personalized recommendations,querying movies by category,and querying movies by name.The back-end system includes movie data management functions and user data management functions.The business logic layer is composed of data processing module,recommendation engine module,and other functional modules.The data processing module processes the data to meet the corresponding format requirements.The recommendation engine module is the core module of the system and takes care of training the model and generating recommended results.The other function modules correspond to the functions of the presentation layer page.The data storage layer takes care of storing system data.This thesis implements various functional modules of the system based on Django.Test the various functions of the system using test cases with Movielens data set.The results show that the movie recommendation system can be applied practically and meet the requirements determined in the design phase.