Research On Recommendation Algorithm Based On Bilateral Conditional Variational Autoencoder With Multi-label Combinations

With the rapid development of Internet technology and the expansion of the popularity of the Internet,more and more people have begun to use the Internet for information search and sharing.The amount of information stored on the Internet has increased exponentially.Traditional collaborative filtering recommendation algorithms rely on users’ historical score data to make recommendations,but such data is usually very sparse,and the algorithm recommendation effect is not good.To change this situation,researchers have introduced deep learning techniques to improve collaborative filtering recommendation algorithms,including variational autoencoders(VAEs).Compared with the neural network model,the VAE model can learn the distribution of the potential feature representation of the data,which has a good effect on alleviating the data sparsity problem in the recommendation algorithm.However,the VAE model also has the following problems: 1)The model does not match the bidirectional nature of preference binary data.2)To improve the accuracy of the recommendation results,it is necessary to consider the auxiliary information of users or items(multi-condition label combinations),but the aggregation of conditional labels will lead to a significant increase in the dimension of conditional features,and training the model as an external condition will lead to the problem of data sparseness.3)Posterior collapse is prone to occur during the training process.Aiming at the above three problems,this paper researches and improves the application of the VAE model in the recommendation algorithm.The main work is as follows:1.Introduce the idea of bilateral variational autoencoder(Bi-VAE)to improve the conditional variational autoencoder for collaborative filtering(CVAEs),and obtain the bilateral conditional variational autoencoder for collaborative filtering(BICVAE).The BICVAE model splits the user-project preference matrix into user scoring vectors and project-scoring vectors by rows and columns.The implicit representations of users and items are extracted from user rating vectors,user side information,item rating vectors,and item side information,respectively,by using BICVAE models.Dot product the implicit representations of users and items to generate a new preference matrix.Finally,the problem that the CVAEs model does not match the bidirectional nature of the preference data is improved.2.A recommendation algorithm(Split-Merge BICVAE)based on a multi-label combination bilateral conditional variational autoencoder is proposed.Since the BICVAE model needs to use the auxiliary information of users and items at the same time,after encoding this information,the label combination of users and items will be obtained.Using these label combinations to train the model will lead to extremely severe data sparsity.To alleviate this problem,this paper proposes a split merging framework for multi-label combinations and applies it to BICVAE models.The framework utilizes the generative nature of the BICVAE model and the idea of bagging algorithms.First,the conditional labels in the user/project auxiliary information are separated by attributes and used to train multiple BICVAE models,each of which can understand the task from a different perspective and generate prediction results.Then,the combined strategy of weighted summation is used to combine these prediction results to obtain the final Top-N recommendation list.For the weight distribution problem involved in the result merging,the quantum particle swarm optimization algorithm with constraints is used to solve the optimal weight combination.The results show that the Split-Merge BICVAE algorithm has better recommendation accuracy than the baseline model.3.Split-Merge Skip-BICVAE algorithm is proposed on the basis of Split-Merge BICVAE algorithm combined with Skip generation model.The BICVAE model may encounter a posterior collapse during training,which affects the model’s learning of useful user and item representations.Therefore,to avoid this phenomenon,a generative jump model is introduced to improve the BICVAE model,and a skip BICVAE model(Skip-BICVAE)is obtained.The Skip-BICVAE model is trained using the split-merge framework to generate a Top-N recommendation list.The experimental results show that the recommendation accuracy of the Split-Merge Skip-BICVAE algorithm is further slightly improved on the basis of the Split-Merge BICVAE algorithm,and the occurrence of posterior collapse can be effectively avoided.