CN-121526713-B - Attention weighting-based advertisement efficiency attribution method

CN121526713BCN 121526713 BCN121526713 BCN 121526713BCN-121526713-B

Abstract

The invention relates to the technical field of advertisement synergy, in particular to an advertisement synergy attribution method based on attention weighting, which comprises the following steps: firstly, based on family identification aggregation cross-screen user logs, multi-dimensional interaction characteristics such as visual duration, sound-painting state and the like in the advertisement exposure process are collected, and the attention score of single exposure is calculated. Next, attention scores are introduced as weights into the multi-contact attribution model, and preliminary contribution values of the contacts are calculated. Meanwhile, a calibration coefficient is constructed by using a real increment conversion value obtained by low-flow training, and a preliminary attribution result is corrected by a multidimensional barrel-dividing strategy. The method solves the problem that the traditional model underestimates the non-explicit interaction value.

Inventors

LAI YILING
ZHANG CHUNYAN
ZHU RONGKUN

Assignees

杭州华数智屏信息技术有限公司

Dates

Publication Date: 20260512
Application Date: 20260115

Claims (6)

1. The attention weighting-based advertisement efficiency attribution method is characterized by comprising the steps of aggregating user behavior logs from different terminals based on family identification dimension, collecting multidimensional interactive features associated with user exposure behavior in the advertisement exposure process based on the user behavior logs, vectorizing the multidimensional interactive features and generating attention feature vectors; analyzing attention feature vectors based on a logistic regression model, and calculating attention scores of single advertisement exposure, wherein the attention scores are used as weights and input into a multi-contact attribution (MTA) model, and preliminary attribution contribution values of all advertisement contacts are calculated; The method comprises the steps of triggering an increment parameter correction training process, performing sparse sampling on a designated advertisement campaign to construct a calibration sample set, obtaining a real increment conversion value of the calibration sample set in a specific sub-bucket, dividing a target flow into a training set flow and a reference set flow based on a preset sampling strategy, performing advertisement delivery logic on the training set flow, performing airdrop delivery logic on the reference set flow to form a calibration sample set with homogeneity, and calculating the difference between the conversion rate in the training set flow and the natural conversion rate in the reference set flow to obtain the real increment conversion value; Calculating a calibration coefficient based on the real increment conversion value and the predicted increment conversion value, and associating the calibration coefficient with a multidimensional barrel-dividing strategy as a correction parameter and storing the correction parameter in a calibration coefficient knowledge base, wherein the dimension according to the multidimensional barrel-dividing strategy comprises at least one of advertisement position type, marketing target, crowd label and industry category; Correcting the preliminary attribution contribution value by applying the calibration coefficient, and outputting incremental advertisement return of the appointed advertisement activity with evidence grade and timeliness mark and the incremental touch man number of the training set flow; the calculation mode of the incremental touch number comprises the steps of obtaining the total conversion number of the training set flow in the calibration sample set, calculating the natural conversion rate based on the reference set flow, calculating the natural conversion number in the training set flow, wherein the natural conversion number is equal to the total number of users of the training set flow multiplied by the natural conversion rate, and subtracting the natural conversion number from the total conversion number to obtain the incremental touch number.
2. The attention-weighting-based advertisement effectiveness attribution method according to claim 1, wherein the multi-dimensional interaction feature comprises at least one of a visual time length feature, a sound-picture state feature, a screen duty cycle feature and an active interaction feature, wherein the visual time length feature is a ratio of an actual playing time length of an advertisement in a screen to a total time length of the advertisement, the sound-picture state feature represents whether the device is in a mute state and the device is in a picture-in-picture playing state, the screen duty cycle feature is an area proportion of advertisement content occupying a display screen, and the active interaction feature represents active operation behavior generated by a user in a preset time window during and after exposure, and the active operation behavior comprises at least one of voice search, cross-screen interaction, fast forward operation, backward operation and pause operation.
3. The attention-weighted advertisement effectiveness attribution method based on the attention-weighted advertisement, as set forth in claim 1, is characterized in that the construction process of the logistic regression model comprises the steps of defining positive samples as events of preset conversion behaviors occurring on mobile terminal equipment associated with home identification within a preset time window after advertisement exposure, defining negative samples as exposure events of the preset conversion behaviors not occurring, training by adopting a supervised learning classification algorithm, learning a functional relation of attention feature vectors mapped to occurrence probabilities of the preset conversion behaviors, and taking the occurrence probabilities output by the model as attention scores of the single advertisement exposure.
4. The attention-weighted advertisement effectiveness attribution method as set forth in claim 1, wherein inputting the attention score as a weight into a multi-contact attribution MTA model, calculating preliminary attribution contribution values of each advertisement contact comprises constructing the MTA model using a shape algorithm, defining conversion path values which are aggregate values of the attention scores of all contacts on the path, and reconstructing a feature function in the shape algorithm Wherein In order for the channel alliance to be present, Is defined as all that include Based on the sum of the path values of the conversion paths of the medium channels A marginal contribution for each contact is calculated, the marginal contribution being in positive correlation with the attention score.
5. The attention-weighted advertisement effectiveness attribution method based on the attention weighting, as set forth in claim 1, wherein the generating logic of the evidence level and the timeliness mark comprises obtaining a generating time of a calibration coefficient matched by a current attribution request and calculating an interval duration between the generating time and the current time, marking as a first-level evidence level when the attribution request triggers an incremental parameter correction training flow for the advertisement activity, marking as a second-level evidence level when the attribution request does not trigger the training flow and simultaneously satisfies three conditions that the interval duration is smaller than a first preset time threshold, the calibration coefficient is in an effective lifetime, and a sub-bucket dimension of the calibration coefficient is completely matched with an advertisement activity attribute, marking as a third-level evidence level when any one of conditions is satisfied, marking as a first-level evidence that the sub-bucket dimension of the calibration coefficient is only partially matched with the advertisement activity attribute, marking as a second-level evidence level when the sub-bucket dimension of the calibration coefficient is completely matched with the advertisement activity attribute, determining as a third-level evidence level after degradation processing is completed, and outputting a criterion that the first-level of the calibration coefficient is completely matched with the advertisement activity attribute, and outputting a criterion when any one of the sub-level of the calibration coefficient is completely matched with the first-level evidence level.
6. The attention-weighting based advertisement enhancement attribution method as set forth in claim 1, further comprising a lifecycle maintenance step of calibrating the coefficient knowledge base for each calibration coefficient stored in the calibration coefficient knowledge base Setting effective life time, periodically scanning knowledge base when detecting a certain calibration coefficient And monitoring the evidence grade state of the specific sub-bucket, and automatically generating a new incremental parameter correction training request when the evidence grade is detected to be lower than a preset grade threshold value.

Description

Attention weighting-based advertisement efficiency attribution method Technical Field The invention relates to the technical field of advertisement synergy, in particular to an advertisement synergy attribution method based on attention weighting. Background Along with the evolution of terminal technology, user behavior logs are generally dispersed in a plurality of heterogeneous devices such as intelligent televisions, mobile terminals, PCs and the like, so that a complex cross-screen data transmission and processing link is formed. In the technical field of attribution calculation based on big data, the existing data processing method has the following technical defects: First, the integrity and granularity of the interactive data acquisition is insufficient. Existing computer systems typically employ a trigger mechanism based on explicit discrete events such as "clicks" to log user behavior. However, in the scenes of OTT large screen startup, video streaming media playing, and the like, the interaction between the user and the terminal often appears as a continuous unstructured state such as a visual duration, a sound-picture state (e.g., silence, picture-in-picture), and the like. Due to the lack of effective capturing and vectorizing processing means for the non-explicit state characteristics, the data dimension of the system input layer is lost, so that the information of the cross-screen user behavior path is broken, and a computer cannot restore the real human-computer interaction process. Second, large-scale data processing architectures present a technical bottleneck between computing accuracy and system resource consumption. To improve the accuracy of the fitting of a multi-contact attribution (MTA) model, it is often necessary to introduce causal inference or complex machine learning algorithms. However, in the face of massive advertisement exposure logs (usually up to hundred million data sizes), performing a high-complexity calibration operation on the full data generates huge computational load and memory I/O overhead, resulting in excessive system processing delay, while adopting a simple statistical model results in low logic confidence of the output result due to failure to reject data noise. At present, the industry lacks a data processing architecture capable of realizing high-efficiency throughput of large-scale data and dynamic correction of model parameters through sparse data feedback under extremely low system overhead. For this purpose, attention weighting-based advertisement effectiveness attribution methods are proposed. Disclosure of Invention The invention aims to provide an advertisement efficiency attribution method based on attention weighting, which comprises the steps of firstly collecting multidimensional interactive features such as visual duration, sound and painting states and the like in the advertisement exposure process based on household identification aggregation cross-screen user logs, and calculating the attention score of single exposure. Next, attention scores are introduced as weights into the multi-contact attribution model, and preliminary contribution values of the contacts are calculated. Meanwhile, a calibration coefficient is constructed by using a real increment conversion value obtained by low-flow training, and a preliminary attribution result is corrected by a multidimensional barrel-dividing strategy. The method solves the problem that the traditional model underestimates the non-explicit interaction value. In order to achieve the above purpose, the present invention provides the following technical solutions: the attention weighting-based advertisement synergy attribution method comprises the steps of aggregating user behavior logs from different terminals based on family identification dimension, collecting multidimensional interactive features associated with user exposure behavior in the advertisement exposure process based on the user behavior logs, vectorizing the multidimensional interactive features, and generating attention feature vectors; analyzing attention feature vectors based on a logistic regression model, and calculating attention scores of single advertisement exposure, wherein the attention scores are used as weights and input into a multi-contact attribution (MTA) model, and preliminary attribution contribution values of all advertisement contacts are calculated; Triggering an increment parameter correction training process, namely performing sparse sampling on a designated advertisement campaign to construct a calibration sample set, acquiring a real increment conversion value of the calibration sample set in a specific sub-bucket, taking the real increment conversion value as an anchor point, calculating a predicted increment conversion value based on a multi-contact attribution (MTA) model aiming at a sample user in the calibration sample set, calculating a calibration coefficient based on the real increment conversion value