CN-121981782-A - Dynamic user portrait modeling and short video advertisement real-time matching method based on incremental learning

CN121981782ACN 121981782 ACN121981782 ACN 121981782ACN-121981782-A

Abstract

The invention provides a dynamic user portrait modeling and short video advertisement real-time matching method based on incremental learning, which comprises the steps of collecting user real-time behavior data, contextual information and advertisement interaction logs, processing the user real-time behavior data, injecting the user real-time behavior data and the contextual information and advertisement interaction logs into a message queue to form an original real-time data stream, consuming the real-time data stream through a stream processing frame, complementing user static attributes and aggregating short-term dynamic behaviors to obtain a dynamic interest vector with a user short term, performing incremental training on an online model based on the historical user short-term dynamic interest vector and corresponding user feedback as samples to generate a dynamic portrait model capable of predicting user global interest preferences in real time, calculating a mixed characterization vector of the dynamic user portrait based on the dynamic portrait model, and recalling a matched candidate advertisement set from an advertisement library based on the mixed characterization vector. The invention realizes the real-time performance and dynamic evolution of the user portrait and greatly improves the accuracy of advertisement touch.

Inventors

ZHANG YIQI
GAO WANCHAO
SONG GUOHANG
WANG YINGZHAN

Assignees

聚告(上海)网络科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260128

Claims (8)

1. A dynamic user portrait modeling and short video advertisement real-time matching method based on incremental learning is characterized by comprising the following steps: collecting user real-time behavior data, context information and advertisement interaction log, processing and then injecting into a message queue, forming an original real-time data stream; Consuming real-time data flow through a flow processing frame, complementing user static attribute and aggregating short-term dynamic behavior to obtain a short-term dynamic interest vector with the user; performing incremental training on the online model based on the historical user short-term dynamic interest vector and corresponding user feedback as samples to generate a dynamic portrait model capable of predicting the user global interest preference in real time; Calculating a mixed characterization vector of the dynamic user portrait in real time based on a dynamic portrait model, and recalling a matched candidate advertisement set from an advertisement library in a multi-way mode based on the mixed characterization vector; sending the candidate advertisement set into a multi-target pre-estimated model updated in an increment mode, predicting each index and sequencing to generate an optimal advertisement sequence; And re-collecting the optimal advertisement sequence and the user real-time feedback data brought by the optimal advertisement sequence, and injecting the optimal advertisement sequence into a data stream to provide samples for continuous increment updating of the portrait and the estimated model.
2. The method for real-time matching of dynamic user portrayal modeling and short video advertisement based on incremental learning according to claim 1, wherein the steps of collecting user real-time behavior data, context information and advertisement interaction log, processing the collected user real-time behavior data, context information and advertisement interaction log, and then injecting the processed user real-time behavior data, context information and advertisement interaction log into a message queue to form an original real-time data stream comprise: Carrying out standardized encapsulation on all acquired real-time data to form a unified data message; and writing the data message into a high-throughput distributed message queue in real time, and completing decoupling and buffering of the data to obtain a continuous original real-time data stream.
3. The method for real-time matching of dynamic user portrayal modeling and short video advertisement based on incremental learning according to claim 1, wherein the specific process of complementing user static attribute and aggregating short-term dynamic behavior to obtain the short-term dynamic interest vector with user comprises: In the stream processing process, the static attribute of the user is associated and complemented according to the user ID by inquiring the user center or the user portrait database; counting the behavior frequency of a user on different videos, advertisements, topics and creators in real time based on a window function of a stream processing frame in near real time; Calculating short-term interest scores of the users on all labels, categories and entities in real time according to the aggregated behavior data and the weight of the behaviors to form short-term dynamic interest vectors; The calculation formula of the interest score is as follows: ; Wherein, the Representing the user's interest score in label/category i at the t-th time window, Representing a set of behaviors within a current time window, The weight of the behavior j is represented, Representing a collection of content associated with a label/category i, Representing an indication function, behavior j belonging to And 1 if not, and 0 if not.
4. The method for real-time matching of dynamic user portrayal modeling and short video advertisement based on incremental learning according to claim 3, wherein the generating the dynamic portrayal model capable of predicting the global interest preference of the user in real time by performing incremental training on the online model based on the historical user short-term dynamic interest vector and the corresponding user feedback as samples specifically comprises: pre-training the online deep learning model through the historical user short-term dynamic interest vector and corresponding user feedback to obtain a basic model; And updating the basic model by real-time user short-term dynamic interest vector sign, user static attribute, context information and real-time feedback generated by the user later to obtain an updated online incremental model, thereby obtaining the dynamic portrait model.
5. The method for real-time matching of dynamic user portrayal modeling and short video advertisement based on incremental learning according to claim 4, wherein the specific process of updating the basic model comprises: based on the generated real-time user short-term dynamic interest vector sign user static properties, context information and the like are used as model input features; taking real-time feedback generated by a user later as a sample tag; The model continuously receives the sample pairs in a streaming mode, and immediately performs incremental update to adjust model parameters.
6. The method for real-time matching of dynamic user portrayal modeling and short video advertisements based on incremental learning of claim 1, wherein the calculating of the hybrid token vector of dynamic user portrayal based on dynamic portrayal model in real time, and the multi-recall matching candidate advertisement set from advertisement library based on the hybrid token vector comprises: when receiving an advertisement matching request, the dynamic portrait model calculates a comprehensive hybrid representation vector representing the dynamic user portraits of the current global interest preference of the user in real time according to the latest short-term dynamic interest vector, the static attribute and the current context of the user; And triggering a multi-path recall strategy based on the mixed characterization vector in parallel, wherein the multi-path recall strategy recalls a plurality of candidate advertisements respectively, and combining and de-duplicating the candidate advertisements to form a primary screening candidate advertisement set.
7. The method for real-time matching of dynamic user portrayal modeling and short video advertisements based on incremental learning of claim 1, wherein the specific flow of calculating the hybrid token vector of the dynamic user portrayal comprises: predicting the preference probability distribution of the user on the full label/category/advertisement in real time by utilizing the updated incremental model; The calculation formula of the model prediction is as follows: ; Wherein, the Representing a predicted probability of feedback from the user to a content or advertisement, Representing input feature vectors, including short-term interest vectors, static attributes, contextual features, etc., Representing a multi-layer nonlinear transformation of the neural network, Expressing the transpose of the weight matrix, representing the bias term, Representing an activation function; the prediction result is weighted and fused with the short-term dynamic interest vector and the long-term static portrait to generate a comprehensive dynamic user portrait with real-time evolution, and the dynamic user portrait is expressed as a mixed characterization vector.
8. The method for real-time matching of dynamic user portrayal modeling and short video advertisement based on incremental learning according to claim 1, wherein the specific flow of generating the optimal advertisement sequence comprises: feature stitching is performed for all advertisements in the candidate advertisement set; The characteristics comprise user side characteristics, advertisement side characteristics, context characteristics and cross characteristics; Sending the spliced candidate advertisement feature set into a multi-target prediction model which is also used for incremental learning, and predicting a plurality of key indexes of each candidate advertisement in real time; according to a preset target, a plurality of key indexes predicted by the model are fused into a final score through a comprehensive sorting score function; The comprehensive sorting function is as follows: ; Wherein, the Representing the final ranking score for ad a, Representing the estimated click-through rate of advertisement a, Representing the estimated conversion rate of advertisement a, Representing the estimated complete sowing rate of the advertisement a, wherein alpha, beta and gamma represent weights; and sorting the candidate advertisements in real time according to the final score, and selecting Top-N to form an optimal advertisement sequence finally displayed to the user.

Description

Dynamic user portrait modeling and short video advertisement real-time matching method based on incremental learning Technical Field The invention relates to the technical field of commercial data processing of advertisements, in particular to a dynamic user portrait modeling and short video advertisement real-time matching method based on incremental learning. Background With the rapid development of the mobile internet and the short video industry, the short video platform has become an important place for digital advertisement delivery. Conventional user portrayal modeling and advertisement matching methods are typically based on offline, batch mode, with significant hysteresis, specifically, conventional user portrayal is heavily dependent on historical long-term behavioral data (e.g., past 30 day behaviors) and is periodically updated through t+1 offline jobs. The method can not capture real-time transient interest focuses of users (for example, the users just search once or generate a series of interactions with certain types of videos), so that the image refreshing frequency is low, the current intention of the users can not be truly reflected, and therefore the optimal advertisement putting time is missed, the existing advertisement click rate (CTR) and conversion rate (CVR) estimated model also adopts an offline training and periodical deployment mode, the latest behavior feedback of the users can not be learned in real time by the model, and the model can not be rapidly adapted to the change of global flow distribution or the emerging hot spot trend, so that the effectiveness of advertisement ordering strategies is attenuated along with the time, and the long-term benefits of a platform are influenced. Therefore, there is an urgent need in the art for an advertisement matching method capable of sensing user interest changes in real time, dynamically updating user images, and realizing real-time cooperation of a model and a business closed loop based on incremental learning, so as to solve the above technical problems and improve the accuracy and business efficiency of advertisement delivery. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a dynamic user portrait modeling and short video advertisement real-time matching method based on incremental learning, which aims to solve the problems in the background art. The invention aims at realizing the technical scheme that the method for real-time matching of the dynamic user portrait modeling and the short video advertisement based on the incremental learning comprises the following steps: collecting user real-time behavior data, context information and advertisement interaction log, processing and then injecting into a message queue, forming an original real-time data stream; Consuming real-time data flow through a flow processing frame, complementing user static attribute and aggregating short-term dynamic behavior to obtain a short-term dynamic interest vector with the user; performing incremental training on the online model based on the historical user short-term dynamic interest vector and corresponding user feedback as samples to generate a dynamic portrait model capable of predicting the user global interest preference in real time; Calculating a mixed characterization vector of the dynamic user portrait in real time based on a dynamic portrait model, and recalling a matched candidate advertisement set from an advertisement library in a multi-way mode based on the mixed characterization vector; sending the candidate advertisement set into a multi-target pre-estimated model updated in an increment mode, predicting each index and sequencing to generate an optimal advertisement sequence; And re-collecting the optimal advertisement sequence and the user real-time feedback data brought by the optimal advertisement sequence, and injecting the optimal advertisement sequence into a data stream to provide samples for continuous increment updating of the portrait and the estimated model. As a further preferred aspect, the method collects user real-time behavior data, context information and advertisement interaction log, processes them, and then injects them into the message queue, the forming of the original real-time data stream specifically comprises: Carrying out standardized encapsulation on all acquired real-time data to form a unified data message; and writing the data message into a high-throughput distributed message queue in real time, and completing decoupling and buffering of the data to obtain a continuous original real-time data stream. As a further preferred aspect, the specific process of supplementing the user static attribute and aggregating the short-term dynamic behavior to obtain the short-term dynamic interest vector with the user includes: In the stream processing process, the static attribute of the user is associated and complemented according to the user ID by inquiring the user center or