CN-121981779-A - Advertisement material AB test method and system based on multidimensional dynamic weight distribution

CN121981779ACN 121981779 ACN121981779 ACN 121981779ACN-121981779-A

Abstract

The invention relates to the field of advertisement putting and recommending, and discloses an advertisement material AB test method and system based on multidimensional dynamic weight distribution. And calculating the estimated click rate variance of the dynamic optimal hierarchy in real time, and dynamically regulating and controlling the flow proportion sent to the orthogonal analysis hierarchy by using a flow respiration formula and a flow damping formula. Experimental analysis is performed at the orthogonal level, updating the independent effect values and generating the confidence flag bits. And further, generating a controlled characteristic gating coefficient according to the confidence zone bit, and calculating the final throwing weight in the dynamic optimization hierarchy by utilizing the adaptive optimization formula in combination with the real-time predicted value. And finally, distributing and updating a closed loop according to the flow proportion and the put weight. The invention can realize the dynamic balance of exploration and utilization, effectively calibrate the model pre-estimated deviation and improve the test convergence speed and global benefit.

Inventors

HE GUANGDE
LI HONGJIE
LIU MIAO
WU SHUBIN
KONG XIAOMING

Assignees

广州随手播网络科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260211

Claims (10)

1. The advertisement material AB test method based on multidimensional dynamic weight distribution is characterized by comprising the following steps of: Constructing a hierarchical flow architecture comprising an orthogonal analysis level and a dynamic optimization level, extracting multidimensional feature vectors of newly accessed advertisement materials, searching a historical feature effect library which is accumulated and generated in the past operation of the orthogonal analysis level, and calculating an initial pre-evaluation value of the advertisement materials; Acquiring a predicted click rate set of all the in-process candidate materials in the dynamic optimal hierarchy in real time, calculating variance, and regulating and controlling the final flow proportion issued to the orthogonal analysis hierarchy by utilizing a flow respiration formula and a flow damping formula based on the variance; Receiving flow according to the final flow proportion at the orthogonal analysis level, executing orthogonal experimental analysis, updating independent effect values in the history characteristic effect library, and generating a confidence zone bit according to the statistical significance of the independent effect values; acquiring a real-time predicted value of the advertisement material, generating a characteristic gating coefficient according to the confidence zone bit, and calculating the final throwing weight of the advertisement material in the dynamic preferential hierarchy by utilizing a self-adaptive preferential formula in combination with the real-time predicted value; And executing flow distribution according to the final flow proportion and the final delivery weight, and performing closed-loop updating by using user feedback data until convergence conditions are met to generate a test conclusion.
2. The method for AB testing advertisement materials based on dynamic weight distribution of multiple dimensions according to claim 1, wherein constructing a hierarchical traffic architecture comprising orthogonal analysis levels and dynamic preference levels, extracting multi-dimensional feature vectors of newly accessed advertisement materials, retrieving a historical feature effect library accumulated and generated in past operation of the orthogonal analysis levels, and calculating an initial pre-evaluation value of the advertisement materials comprises: Logically dividing the accessed total flow into the orthogonal analysis level and the dynamic optimization level with independent functions by utilizing a flow distribution unit; When a new advertisement material is accessed, an image recognition and natural language processing technology is called to extract the multidimensional feature vector contained in the advertisement material, and the multidimensional feature vector is used as a retrieval key value; Searching and acquiring the independent effect value of each characteristic dimension in the historical characteristic effect library based on the search key value, and calculating the initial estimated value of the advertisement material at the initial moment by using an initial estimated formula; The initial pre-determined value is input as an initial state of the advertising material in the dynamic preference hierarchy.
3. The method for testing the advertisement material AB based on multidimensional dynamic weight distribution according to claim 1, wherein the step of adjusting the final flow ratio issued to the orthogonal analysis level by using a flow respiration formula and a flow damping formula based on the variance specifically comprises: Calculating the target flow proportion which the orthogonal analysis level should bear at the current moment by utilizing the flow respiration formula based on the calculated variance; the flow respiration formula is calculated according to the variance, the lower flow proportion limit of the orthogonal analysis level, the upper flow proportion limit of the orthogonal analysis level, the sensitivity coefficient and the variance inflection point threshold; The variance inflection point threshold is preset according to the average variance of the historical operation data, and the sensitivity coefficient is preset according to the slope requirement of flow regulation on uncertainty change.
4. The method of claim 3, wherein the step of adjusting the final flow ratio delivered to the orthogonal analysis level using a flow respiration formula and a flow damping formula based on the variance further comprises: And obtaining the final flow ratio calculated at the previous moment as the last moment ratio, substituting the final flow ratio and the target flow ratio into the flow damping formula to perform low-pass filtering processing, and calculating to obtain the final flow ratio issued to the orthogonal analysis level at the current moment.
5. The method of claim 1, wherein the steps of receiving traffic at the orthogonal analysis level according to the final traffic proportion and performing orthogonal experimental analysis, and updating the independent effect values in the historical characteristic effect library specifically comprise: Receiving traffic at the orthogonal analysis level according to the final traffic proportion, mapping the characteristic dimension of the advertisement material into an experimental factor, mapping the specific value of the characteristic into an experimental level, and displaying the advertisement material according to a combination specified by an orthogonal table strategy; the performing orthogonal experimental analysis includes: and respectively counting average click rates of a characteristic-containing sample set and a non-characteristic sample set, and updating the independent effect value in the history characteristic effect library by calculating the difference between the average click rate of the characteristic-containing sample set and the average click rate of the non-characteristic sample set.
6. The method for AB testing of advertising materials based on multi-dimensional dynamic weight distribution of claim 1, wherein generating a confidence flag bit based on statistical significance of said independent effect values comprises: Monitoring the sample accumulation amount and standard error of the orthogonal analysis level in the current time window in real time, and calculating a confidence interval width based on the standard error to judge the statistical significance of the independent effect value; Generating the confidence flag bit indicating an untrusted state when a condition that the sample accumulation amount is less than a minimum sample threshold or the confidence interval width is greater than a confidence interval width threshold is satisfied; when the condition is not met, generating the confidence zone bit indicating the trusted state; The confidence interval width threshold is preset according to the tolerance of the business to the misjudgment risk, and the minimum sample threshold is preset according to the minimum observation times required by the convergence of the statistical result ensured by the law of large numbers.
7. The method for AB testing of advertising materials based on multi-dimensional dynamic weight distribution of claim 1, wherein generating feature gating coefficients from said confidence flag bits comprises: if the confidence flag bit indicates an unreliable state, triggering a locking mechanism, and forcedly setting the characteristic gating coefficient to be 1; if the confidence flag bit indicates a trusted state, checking whether the independent effect value of the advertisement material conflicts with the real-time predicted value; If the independent effect value is negative, but the real-time predicted value is higher than the sum of the global reference value and the tolerance threshold, judging that the independent effect value is in conflict, generating a punishment gating coefficient with the value smaller than 1 as the characteristic gating coefficient, and if the independent effect value is not in conflict, generating the characteristic gating coefficient with the value equal to 1; The global reference value is obtained according to statistics of average click rate of all advertisement materials in a historical time window, and the tolerance threshold is preset according to tolerance of business to model overestimate risks.
8. The method for AB testing of advertising materials based on multi-dimensional dynamic weight distribution of claim 1, wherein the step of calculating final placement weights of said advertising materials in said dynamic preference hierarchy using an adaptive preference formula in combination with said real-time pre-estimate values comprises: combining a numerical scaling factor, an inverse temperature coefficient and the characteristic gating coefficient, carrying out weighted calculation and normalization processing on the real-time predicted value by utilizing the self-adaptive optimization formula, and calculating the final throwing weight of the advertisement material in the dynamic optimization hierarchy; The numerical scaling factor is preset for amplifying small differences of click rate values, and the inverse temperature coefficient is a parameter which increases in a nonlinear way along with experimental time to control the sharpness of probability distribution.
9. The method for testing advertisement materials AB based on multidimensional dynamic weight distribution according to claim 1, wherein the step of performing flow distribution according to the final flow ratio and the final delivery weight, and performing closed-loop update by using user feedback data until convergence conditions are satisfied, comprises the steps of: performing real-time total flow distribution on the orthogonal analysis level and the dynamic optimization level according to the final flow proportion by using a flow distribution unit; The advertisement materials are randomly sampled and displayed in a weighted mode in the dynamic optimization hierarchy according to the final throwing weight; Collecting the generated user feedback data through a stream processing architecture, transmitting the user feedback data back to the dynamic preference level to update the real-time predicted value and variance, and transmitting the user feedback data back to the orthogonal analysis level to update the independent effect value and the confidence zone bit; Terminating the test and generating the test conclusion when the variance of the dynamic preference hierarchy is below a variance convergence threshold; The generating the test conclusion includes: Calculating posterior conversion rate based on the real click quantity and exposure quantity accumulated by the dynamic optimization hierarchy in the test life cycle to generate a conversion effect ranking containing the quality ordering of each advertisement material; extracting the final converged independent effect value from the history characteristic effect library, and generating an independent effect value report for quantitatively revealing the specific contribution degree of each characteristic dimension to the conversion effect; The variance convergence threshold is preset according to the empirical variance level of the model after entering a steady state.
10. The advertisement material AB test system based on multi-dimensional dynamic weight distribution is characterized by being applied to the advertisement material AB test method based on multi-dimensional dynamic weight distribution as claimed in any one of claims 1-9, and comprising the following steps: An initialization module (10) for constructing a hierarchical traffic architecture including an orthogonal analysis hierarchy and a dynamic optimization hierarchy, extracting a multidimensional feature vector of a newly accessed advertisement material, retrieving a historical feature effect library accumulated and generated in the past operation of the orthogonal analysis hierarchy, and calculating an initial pre-evaluation value of the advertisement material; The flow regulation and control module (20) is used for acquiring the estimated click rate set of all the in-process candidate materials in the dynamic optimal hierarchy in real time, calculating variance, and regulating and controlling the final flow proportion issued to the orthogonal analysis hierarchy by utilizing a flow respiration formula and a flow damping formula based on the variance; The orthogonal analysis module (30) is used for receiving the flow according to the final flow proportion at the orthogonal analysis level and executing orthogonal experimental analysis, updating the independent effect values in the history characteristic effect library and generating a confidence zone bit according to the statistical significance of the independent effect values; The weight calculation module (40) is used for acquiring a real-time predicted value of the advertisement material, generating a characteristic gating coefficient according to the confidence zone bit, and calculating the final throwing weight of the advertisement material in the dynamic preference hierarchy by utilizing an adaptive preference formula in combination with the real-time predicted value; and the execution feedback module (50) is used for executing flow distribution according to the final flow proportion and the final delivery weight, and performing closed-loop updating by using user feedback data until a convergence condition is met to generate a test conclusion.

Description

Advertisement material AB test method and system based on multidimensional dynamic weight distribution Technical Field The invention relates to the field of advertisement delivery and recommendation, in particular to an advertisement material AB test method and system based on multidimensional dynamic weight distribution. Background In an online advertisement delivery system, AB testing is a common means of evaluating the effectiveness of advertisement materials and optimizing delivery strategies. When new materials are accessed, the conventional advertisement material testing method generally adopts a strategy of random initialization or global mean value assignment, and history characteristic data cannot be effectively utilized for priori estimation. This approach results in new materials lacking reference bases at the initial stage of testing, requiring a longer sample accumulation period to converge, reducing the exploration efficiency at the cold start stage. Meanwhile, the traditional test flow distribution strategy mostly adopts fixed segmentation proportion, and real-time adjustment is difficult to carry out according to uncertainty of model judgment. When the model uncertainty is high, the fixed flow limits the exploration speed, which easily causes the local optimum, and when the model uncertainty is low, the excessive exploration flow causes the loss of conversion benefits, so that the effective balance between unknown exploration and known utilization cannot be obtained. In addition, the existing click rate estimation mainly depends on a deep learning model, and a single model is easily influenced by sample deviation to generate an overfitting or overestimation phenomenon. Because of lacking a checking mechanism based on statistical significance, poor-quality materials with estimated deviation occupy exposure resources for a long time, so that the conversion effect of integral throwing is reduced. Disclosure of Invention Aiming at the defects of the prior art, the invention provides an advertisement material AB test method and system based on multidimensional dynamic weight distribution, which solves the problems that the cold start exploration efficiency of new materials in the prior art is low, the dynamic balance exploration and utilization flow according to uncertainty is difficult, and a statistical verification mechanism aiming at model estimated deviation is lacking. In order to achieve the above purpose, the invention is realized by the following technical scheme: The invention provides an advertisement material AB test method based on multidimensional dynamic weight distribution, which comprises the following steps: Constructing a hierarchical flow architecture comprising an orthogonal analysis level and a dynamic optimization level, extracting multidimensional feature vectors of newly accessed advertisement materials, searching a historical feature effect library which is accumulated and generated in the past operation of the orthogonal analysis level, and calculating an initial pre-evaluation value of the advertisement materials; Acquiring a predicted click rate set of all the in-process candidate materials in the dynamic optimal hierarchy in real time, calculating variance, and regulating and controlling the final flow proportion issued to the orthogonal analysis hierarchy by utilizing a flow respiration formula and a flow damping formula based on the variance; Receiving flow according to the final flow proportion at the orthogonal analysis level, executing orthogonal experimental analysis, updating independent effect values in the historical characteristic effect library, and generating a confidence zone bit indicating the credible state of data according to the statistical significance of the independent effect values; Acquiring a real-time predicted value of the advertisement material, generating a controlled characteristic gating coefficient according to the confidence zone bit, and calculating the final throwing weight of the advertisement material in the dynamic preferential hierarchy by utilizing a self-adaptive preferential formula in combination with the real-time predicted value; And executing flow distribution according to the final flow proportion and the final delivery weight, and performing closed-loop updating by using user feedback data until convergence conditions are met to generate a test conclusion. Further, the step of constructing a hierarchical flow architecture including an orthogonal analysis level and a dynamic optimization level, extracting a multidimensional feature vector of a newly accessed advertisement material, searching a historical feature effect library accumulated and generated in the past operation of the orthogonal analysis level, and calculating an initial pre-evaluation value of the advertisement material specifically comprises the following steps: Logically dividing the accessed total flow into the orthogonal analysis level and the dynamic op