Method, device, recommendation server and storage medium for information recommendation

Abstract

The embodiment of the invention discloses an information recommendation method, device, recommendation server and storage medium. Wherein, the method includes: predicting the feature vector of the recommendation information through the recommendation model to obtain the click-through rate of the recommendation information, and removing the click-through rate of the recommendation information that has the same feature as the feature vector and whose frequency is less than or equal to the first threshold in the recommendation model. The model parameters corresponding to the features; according to the click through rate, send the recommendation information. In the technical solution provided by the embodiment of the present invention, in the training process of the recommended model, only the model parameters corresponding to the features whose frequency is greater than the first threshold are retained, which reduces the size of the recommended model and reduces the space occupied by parameters in the recommended model. Query the model parameters corresponding to each feature in the feature vector, which reduces the time spent in the query process. Only the features corresponding to the model parameters retained in the recommendation model are analyzed, which improves the accuracy of the information recommendation while ensuring the accuracy of the recommendation. efficiency.

Description

technical field [0001] Embodiments of the present invention relate to the technical field of machine learning, and in particular to an information recommendation method, device, recommendation server, and storage medium. Background technique [0002] With the rapid development of artificial intelligence interaction technology, pre-trained machine learning models are more and more widely used in big data scenarios of various businesses such as recommendation, search, and advertising. The background will obtain a large number of relevant recommendations for users' search needs. At this time, it is necessary to predict the click through rate (Click Through Rate, CTR) of multiple recommended information, so as to push the optimal information to the user according to the click through rate of each recommended information. [0003] At present, the business prediction model that is usually trained by the logistic regression algorithm to try to fit the user's click behavior on the r...

AI Technical Summary

Problems solved by technology

[0003] At present, the business prediction model that is usually trained by the logistic regression algorithm to try to fit the user's click behavior on the recommended information includes the model weight coefficient under each feature dimension in the recommended information. When predicting the click-through rate of each recommended information, It is necessary to query the feature values ​​of each feature dimension in the recommendation information and the corresponding model weight coefficients. At this time, there are a large number of high-dimensional sparse features in the recommendation information, such as the user identity (identity, ID) under the identity dimension in the recommendation information. , host ID or device ID, etc., can reach hundreds of millions of dimensions at most, so that the business prediction model will occupy a large amount of memory space when storing a large number of high-dimensional sparse features in the recommendation information, and faces the problem of insufficient memory; and for a large number of features For recommendation information, it takes a long time to query the model parameters and model weight coefficients under each feature dimension in the recommendation information, which makes the business prediction model too large, resulting in low prediction efficiency of recommendation information

Images