A Multi-modal Recommendation Network Based On EM Routing Algorithm

Benefiting from the rapid development of the Internet,we are able to access massive amounts of information at any time.However,how to extract from massive data of valuable information is a problem that we need to solve at present.Recommend system is considered as a solution of this problem in which recommend models are the most valuable part.Recommendation model utilizes personal characteristics and the historical behavior of the user to infer their possible interests' tendency.According to these possible tendencies,recommend model matches all the information in the database with similarity,and takes out the several objects with the highest matching score as the result of recommendation.Most of the current mainstream recommendation models extract all the user's behavior characteristics into a high-dimensional vector to express the interest tendency.This method is often unable to accurately summarize the user's behavior tendency in a vector when facing the common ultra-long behavior sequences.The recommendation model based on deep learning improves some of the problems in the traditional recommendation model,such as no longer designing features based on prior knowledge and allowing the model to explore possible features by fitting data,etc.,but its design ideas are still subject to Limitations of the traditional recommendation model.With the rapid development of big-data,sequence features are no longer the only choice for recommended models.How to use other types of data to enrich the content of high-level features has become a major research direction of recommendation models.Most of the current mainstream recommendation models abstract all the user's behavior characteristics into a high-dimensional vector to express the user's interest tendency.This method is often unable to accurately summarize the user's behavior tendency in a vector when facing the common ultra-long behavior sequences.In order to fully tap the user's high-dimensional characteristics,we combine the capsule network with a recommendation model based on deep learning,find the user's multiple interest tendencies through the S2 I algorithm,and make recommendations based on these different interest tendencies.Compared with the traditional recommendation model,which recommends based on a single interest tendency,our model has achieved 4.8%,2.1%,and 0.2 respectively on the Amazon Electron,Amazon Movies-TV,and Amazon Home-Kitchen datasets compared to the benchmark model.% Performance improvement.In addition,we also analyzed the relationship between the number of interest tendencies and the accuracy of the recommendation results on the movielens-2K dataset.Finally,through qualitative experiments,we proved that the S2 I algorithm can indeed cluster similar interest tendencies.In the final output of the algorithm,items with higher similarity are clustered in the same capsule.The current mainstream recommendation models are mostly limited to the analysis of the behavior sequence itself,instead of using other types of features to complete the content that the sequence features cannot express.In this article,we conduct research and experiments on how to better extract image features,and integrate the extracted image features into the recommendation model.In the course of the experiment,we found that the traditional CNN model cannot learn the rotation invariance of the picture,which leads to the fact that the image features extracted by the CNN network are not general and the migration ability is poor,while the capsule network can better extract the item features.The spatial relationship between the image features extracted by the capsule network is strong.However,the capsule network cannot handle pictures with complex backgrounds,so it cannot be directly applied to the extraction of item pictures.In order to solve this problem,we proposed a multi-channel capsule network(MLSCN).By improving the network structure and Squash function,our model can extract good image features on the basis of ensuring rotation invariance.In quantitative experiments,our model achieved 99.73% accuracy on MNIST,76.79%classification accuracy on CIFAR10,and 65.37% accuracy on MNIST-CB.Through multiple sets of ablation experiments,we have verified the effectiveness of the improved method.Through robustness experiments,we verify that MLSCN retains the powerful migration capabilities of the capsule network.Finally,we integrate the extracted features into the recommendation model to improve the accuracy of the recommendation results.