Research On Recommendation Method Based On Image-Text Data Fusion

The recommendation system can use data of products themselves or combine user behaviors to recommend products to users,and provide information and suggestions.It is the main technical means of current e-commerce system to achieve precise services and improve"stickiness".Most commodities in e-commerce system have two types of data:text and image.The characteristic information contained in these two types of data is often complementary.Effective fusion of the two types of information can greatly improve the accuracy and interpretability of the recommendation results.But it is very difficult and is currently a hot research topic.The main goal of the thesis is to explore the fusion method of image and text to achieve effective recommendation,and focus on solving the accuracy of feature extraction of different categories of data and the fusion of two features.For the image data of commodities,based on the SqueezeNet network,by reducing the size of the convolution kernel of Squeeze and Expand structures in Fire Module,and increasing the number of pooling layers,the utilization of category features was improved.For the text data of commodities,the data after the word segmentation and stop words removal were input into the network to obtain the feature word vector,and then the cosine similarity calculation method was used to obtain the similarity between texts.Eventually,the fusion of the features of two modalities was adopted by decision fusion.Through experiments on the number of network layers,the size of convolution kernel and the number of pooling layer,the improved SqueezeNet network in this thesis achieved a recognition accuracy of 98%.Through the PCA dimensionality reduction algorithm verification,when the number of hidden layer units was 10,Word2vec had a good distinction between text features.Finally,compared with using only text and using only image data,the fusion algorithm proposed in this thesis performed best in accuracy,recall and F value.Experiments and analysis indicated that the improved image feature extraction algorithm in the thesis improved the utilization rate of key features,and the reasonable text feature extraction algorithm played a key role in supplementing image features.After the image and text data were fused,cascading recommendation method was finally used.Recommendation solved the problem that the difference between different modal features at the feature level was large and difficult to deal with in high dimensions,and effectively completed the recommendation task.