Research On Hybrid Recommendation Algorithm Enhancement By Stacked Denosing Autoencoder And Users' Labels

With the fast development of computer technology and information communications,the issue of information overload becomes more and more critical.It is difficult for users to find what they rally want quickly and accurately in massive data ocean.Therefore,providing personalized recommendation to users has become a hot research topic.The traditional recommendation algorithm only uses the structured data such as scores to generate the recommendation results,without using the unstructured UGC text information(such as comments,labels,text-content,etc.),users' interest cannot be identified accurately,and it's easy to overfit especially when dataset is sparse.In order to solve these problems,an improved recommendation algorithm based on deep learning algorithm "stacked denoising autoencoder" has been proposed,which extracts features from users' massive free-text tags,and combines with collaborative filtering algorithm to provide more accurate and personalized recommendation services for Internet users.The main work of this paper includes four points:(1)Introducing free-text label informationThe traditional collaborative filtering recommendation algorithm only utilizes the user's rating information,and the recommendation result is difficult to match the user's interest focus.Therefore,it is necessary to introduce additional auxiliary recommendation information to improve the accuracy of the recommendation result.The free text label marked by the user can reflect the user's interest preference for the item.This paper introduces the label information as an auxiliary recommendation basis to enhance the interest matching degree of the recommendation algorithm.(2)Establish a label-feature extraction modelSince the user tags are unstructured free texts,the number is large and the distribution is sparse.For the sparse distribution,the model over-fitting problem will be caused.This paper uses the tag set expansion method to reduce the impact of data sparsity on the performance of the algorithm;in the massive data set scenario The dictionary of tag phrases will be very large.Using traditional one-hot codes to encode tags will cause the computational dimension to explode.This paper uses the word vector compression algorithm word2 vec to encode and compress the tags,which can effectively reduce the computational complexity of the recommended algorithm training.For text feature extraction,this paper uses Stacked Denoising Autoencoders(SDAE)model to extract valid interest expression features from tag data.(3)Dynamic label weightsThe traditional algorithm considers the tags to be independent,and does not consider the relationship between tags and user interest preferences.This paper proposes dynamic tag weights based on user sentiment expression and scoring feedback.User tags are divided into positive and negative categories.Extract and then use the Factorization Machine(FM)to perform feature normalization to improve the interest expression of the tag features.(4)Improved collaborative filtering modelThe user tag feature is introduced on the basis of the collaborative filtering recommendation.When the "user-interest" attention matrix and the "item-interest" quality matrix are generated,the constraint of the tag feature is applied,and the score feedback is used to improve the interest matching ability of the recommendation algorithm.The algorithm of this paper is experimentally verified on the large open source dataset "MovieLens".The results show that the proposed algorithm can improve the accuracy of recommendation algorithm and the coverage of recommendation results.Comparing to other “deep learning” based algorithm models,there is no significant difference on complexity and training performance.