| The main method of click-through rate prediction is to predict the probability of a user clicking on an item via models based on information such as user personal information,user historical click records,item information,etc.Whether in the field of computational advertising or recommender systems,CTR prediction models play a very important role.At present,feature interaction learning and sequence modeling are the two main research directions of CTR prediction models based on deep learning.The existing models have achieved good results,but there are still some problems to be solved:(1)In terms of feature interaction learning,many models simply use the inner product and Hadamard product to capture interactive features,while ignoring the impact of different interactive features on the prediction results;(2)In terms of sequence modeling,the user’s short-term interest and general interest will have different effects on the user’s next click behavior.In terms of capturing general interest,existing models pay more attention to the relationship between item Embedding vectors(union-level)while ignoring the relationship between elements in the item Embedding vector(point-level).The attention mechanism can capture useful information from the input features through weight distribution.Improved models are proposed in this paper respectively for the above problems by combining the attention mechanism and an online product recommender system is built by combing these proposed models.The specific work of this paper is as follows:(1)For the problem of ignoring the importance of different feature interaction information in terms of feature interaction learning,SEFM(Squeeze-and-Excitation Factorization Machines)model is proposed in this paper,which improves SENet which is a type of attention mechanism with factorization machines to assign different weights to interactive features.We build a CTR prediction model based on SEFM.The AUC values of the SEFM model on the Movie&TV,Video Game and Movie Lens are 0.8612,0.7880 and 0.9203,proving the effectiveness of the model.(2)For the problem of ignoring the relationship between elements in the item Embedding vector while capturing general interest in terms of sequence modeling,the Lambda layer based Convolutional Sequence Embedding(LCSE)model proposed in this paper uses horizontal filters and vertical filters to capture user general preferences from union-level and point-level on the basis of using the Lambda layer which is a kind of linear attention to learn user general interest from user’s click history through weight distribution.The AUC values of the LCSE model on the Electronic,Movie&TV and Movie Lens are 0.8707,0.9036 and 0.9467,proving the effectiveness of the model.In addition,due to the additional information of the time dimension,the LCSE model can often achieve better prediction results than the common click-through rate prediction models based on feature interaction learning.(3)The click-through rate prediction model is often applied in the ranking module in the recommender system to obtain a more accurate ranking of the recommendation results.We build a distributed online product recommender system around the SEFM model and LCSE model proposed in this paper.The recommendation module consists of a recall module and a sorting module.The SEFM model proposed in this paper pays more attention to the mining of user and target item feature interaction information and the LCSE model pays more attention to the capture of user historical preference information.Therefore,the sorting module uses a multi-way sorting method that integrates the SEFM model and the LCSE model by weighting.After coding,testing and deployment,the realized online product recommender system can meet the expected requirements. |
