Research On Fusion Recommendation Algorithm Based On Embedding Vectors In Implicit Feedback

With the steady development of the social economy and the increasing popularity of Internet technology,the e-commerce industry has developed relatively maturely in China.This article mainly focuses on the massive user implicit feedback data like browsing click behavior recorded with timestamp.On the one hand,with the help of deep learning which can get effective implicit semantic information,the article uses the method of word2vec in natural language processing,which analogizes user behavior sequences with word sequences and time-order-co-occurrence relationship analogies with word context.The time-order-co-occurrence relationship of commodities in users'behavior sequence shows that the goods with high co-occurrence frequency may have higher similarity.The article uses Skip-gram model based on Negative Sampling(SGNS)to get commodities embedding vectors.On the other hand,users' implicit behaviors are firstly converted into a scoring matrix and used by the latent semantic model(LFM)to train the embedding vectors of users and commodities under latent factors.Then reusing the advantages of the GBDT model in continuous feature processing and the advantages of LR model in discrete feature processing,the user embedding vector and commodity embedding vector obtained by LFM are used as the joint input of the GBDT-LR model to train the fusion click rate estimation model.At last,the article fuses these two parts to produce more accurate recommendation results.As a recall part of the recommendation system,the article firstly gets the top-N rough recall commodity sets based on users' historical behaviors and the embedding vector obtained by the SGNS model which well describes the similarity of commodities,then use the LFM-GBDT-LR fusion model to predict and reorder the click-through rate of the crude recalled commodities set.After a series of experiments,the embeddingSGNS obtained in this paper can effectively mine the implicit features of commodities.The LFM-GBDT-LR fusion model performs better on the model evaluation indexes than the single model.The LFM-GBDT-LR click rate prediction model incorporating embeddingSGNS helpfully reduces the calculation time and also has a good recommendation effect.