Advertisement click rate estimation method based on user preferences

Abstract

The invention discloses an advertisement click rate estimation method based on user preferences. The method comprises the following steps: acquiring data to obtain an advertisement click sequence of each user; preprocessing the advertisement click sequence to obtain a sequence training data set; constructing a click rate prediction model based on a user preference network; carrying out iterative training on the click rate prediction model based on the user preference network by adopting the training data set obtained by preprocessing; and inputting a complete advertisement click sequence of the user, learning user preference characteristics by utilizing the click rate prediction model based on the user preference network obtained by training, and obtaining a click rate prediction result. According to the invention, a novel attention mechanism is introduced, implicit long-term and short-term preferences in a user sequence are considered at the same time, user drift preferences and fusion preferences are further learned, the defects of an existing method in user preference diversity multi-dimensional and drift modeling can be overcome, and the advertisement click rate estimation accuracy is effectively improved from the perspective of user preference learning.

Description

technical field [0001] The present invention relates to the field of recommendation systems, and more specifically, to a method for estimating the click-through rate of advertisements based on user preferences. Background technique [0002] Click-through rate estimation is a key link in advertising delivery and recommendation. In the field of online advertising, click-through rate estimation is the ratio of advertising clicks to delivery volume, which measures the probability of an advertisement being clicked by a user. Its accuracy is directly related to platform user experience and online advertising platform revenue, and has extremely important research value. The preference information reflected in a series of behaviors such as clicking to buy greatly affects whether a user will click on an advertised product. [0003] At present, there are roughly two types of advertising click-through rate estimation methods. One is based on traditional machine learning methods for cl...

AI Technical Summary

Problems solved by technology

[0004] These methods focus on improving the ability of the model to learn low-level and high-level features and their interaction relationships, ignoring the effect of improving click-through rate prediction from the perspective of mining user preferences, and it is difficult to make full use of the hidden user preference information in the behavior sequence

The few methods proposed from the perspective of preference learning, such as DIN and DIEN, have the shortcomings of only considering local short-term preferences or the structure of preference learning is too primitive. Dig deep and exploit

Images