Search

CN-121981777-A - Digital marketing effect prediction evaluation system based on multi-mode data

CN121981777ACN 121981777 ACN121981777 ACN 121981777ACN-121981777-A

Abstract

The invention relates to the technical field of internet advertisement service and discloses a digital marketing effect prediction evaluation system based on multi-modal data, which comprises a characteristic extraction unit, a characteristic decoupling unit and a characteristic analysis unit, wherein the characteristic extraction unit maps the multi-modal data into an initial characteristic vector, and the characteristic decoupling unit separates an inherent characteristic vector and an environment characteristic vector by utilizing an orthogonal constraint term containing an adjustable weight coefficient; the invention realizes the self-adaptive dynamic calibration of the characteristic decoupling boundary by constructing a closed-loop control mechanism based on residual feedback, and solves the problem of characteristic attribute drift caused by market environment rheology.

Inventors

  • LI JIAYING

Assignees

  • 深圳万象博联文化传媒有限公司

Dates

Publication Date
20260505
Application Date
20260104

Claims (10)

  1. 1. A digital marketing effect prediction assessment system based on multimodal data, comprising: The feature extraction unit is used for acquiring visual data, text data and attribute data of the object to be put in, and mapping each data into initial feature vectors with uniform dimensions respectively; The characteristic decoupling unit is used for receiving an initial characteristic vector, and respectively outputting a first characteristic component with causal relation of characterization and conversion results and a second characteristic component with statistical correlation of characterization and conversion results through a first mapping channel and a second mapping channel which are arranged in parallel; The residual analysis unit is used for acquiring an externally input real conversion label, calculating a prediction residual between the real conversion label and a prediction value generated based on the first characteristic component, and mapping the prediction residual into a residual state vector; The PID feedback control logic is configured to output negative adjustment increment when the correlation coefficient rises so as to reduce the weight coefficient, and a part of the second characteristic component, which is related to the prediction residual, is reserved in the first characteristic component, so that the constraint strength of the orthogonal constraint item is dynamically adjusted; and the bidding instruction generation unit is used for receiving the first characteristic component which is output by the characteristic decoupling unit and regulated by the weight coefficient, calculating the reference conversion probability, and generating the digital marketing delivery instruction containing the target bid amount by combining the real-time bidding strategy.
  2. 2. The system of claim 1, wherein the residual analysis unit performs arithmetic logic to calculate a normalized dot product of the residual state vector and the second feature component to obtain the correlation coefficient when calculating the correlation coefficient The operation expression is as follows: , wherein, As a residual state vector of the image data, As a second characteristic component of the image, The modulus length of the expression vector is a preset numerical stability constant for preventing the denominator from being zero; The value is used to quantify the amount of information contained in the second feature component that has an explanatory power on the prediction residual.
  3. 3. The system of claim 1, wherein the orthogonal constraint adjustment unit comprises a discrete PID controller, the controller calculates a weighted sum of a proportional term, an integral term, and a differential term as an adjustment increment, with a deviation of the correlation coefficient from a preset threshold as input, and the orthogonal constraint adjustment unit updates the weight coefficient according to the adjustment increment in a continuous time window.
  4. 4. The system of claim 1, further comprising a sensitivity analysis unit for applying normalized numerical increments to each feature dimension of the first feature component to generate a plurality of test feature vectors, respectively, before the bidding instruction generation unit generates the instruction, the sensitivity analysis unit calculating a change rate of the corresponding predictive conversion probability for each test feature vector and taking the change rate as a real-time weight factor of each feature dimension, and the bidding instruction generation unit being configured to weight the first feature component according to the real-time weight factor and generate the digital marketing delivery instruction based on the weighted feature vectors.
  5. 5. The digital marketing effect prediction evaluation system based on multi-mode data according to claim 1, further comprising a bidding environment analysis unit for acquiring bidding log data in a historical time window, wherein the bidding log data comprises a bidding record and a bidding failure record, the bidding environment analysis unit maps the bidding log data to a preset flow characteristic space to construct a competition heat data structure for representing the competition intensity of different flow areas, and the bidding instruction generation unit is further used for inquiring local competition gradient of a coordinate area where an object to be put is located in the competition heat data structure and adjusting a target bidding amount according to the local competition gradient.
  6. 6. The multi-modal data based digital marketing effect prediction assessment system of claim 1, wherein the first mapping channel is comprised of a multi-layer perceptron network for extracting the first characteristic component, the second mapping channel is comprised of a self-encoder network having a bottleneck layer structure for extracting the second characteristic component, and the orthogonal constraint term is configured to minimize an inner product square of the first mapping channel output vector and the second mapping channel output vector.
  7. 7. The system of claim 5, wherein the bid environment analysis unit maps the highest bid in the bid failure record to a high resistance reference point in the competitive heat data structure and maps the bid in the bid record to a reference point in the competitive heat data structure, and wherein the bid instruction generation unit executes numerical optimizing logic avoiding the high resistance reference point to locate a low competitive strength traffic region when generating the target bid amount.
  8. 8. The system of claim 4, wherein the sensitivity analysis unit applies a numerical increment only for a dimension of the first feature component whose numerical type is a continuous value, and wherein the sensitivity analysis unit performs a nearest neighbor substitution operation based on a vector space distance for a dimension of the numerical type that is a discrete class to generate the test feature vector.
  9. 9. The system of claim 1, wherein the bidding command generation unit further comprises an abnormal fusing module for monitoring a value of a weight coefficient, and when the weight coefficient is lower than a preset minimum value, the abnormal fusing module determines that the first feature component and the second feature component are collinearly entangled, and triggers a model resetting command or a bidding stopping command for an object to be put.
  10. 10. The multi-modal data-based digital marketing effect prediction assessment system of claim 1, wherein the feature extraction unit, the feature decoupling unit, the residual analysis unit, the orthogonal constraint adjustment unit, and the bid instruction generation unit are all disposed in a real-time computing server of the advertising exchange platform, and the system is configured to complete a data processing process from feature extraction to generation of the digital marketing delivery instruction within a preset time threshold after receiving an advertisement request.

Description

Digital marketing effect prediction evaluation system based on multi-mode data Technical Field The invention relates to a digital marketing effect prediction evaluation system based on multi-mode data, and belongs to the technical field of internet advertisement services. Background In the current internet advertisement putting system, the click rate or conversion rate is predicted by utilizing multi-mode data, which is the basis of flow accurate distribution and resource dynamic configuration, advertisement material vision, text and attribute data are mapped into high-dimensional feature vectors by utilizing a feature extraction network in the prior art, and the intrinsic value main feature of a representative product and the background modification environment feature are separated by setting orthogonal constraint or attention weight, so that the interference of non-core background noise on commodity value evaluation is reduced. However, when the conventional scheme faces complex and changeable market dynamics, the feature processing logic reveals the rigidifying defect, for example, the Chinese patent with publication number of CN118691347A discloses a multi-mode data based digital marketing effect prediction evaluation method, which is used for acquiring multi-mode data including texts, images and videos, constructing a training set training evaluation prediction model technical path by combining on-line praise comments and off-line electroencephalogram questionnaire feedback data, and introducing bioelectricity and other biofeedback data to widen evaluation dimensions, but the technical essence still belongs to a static feature engineering and off-line supervision learning mode, specifically, the mapping weight between the features and results after model training is cured, the default feature and causal attributes are kept constant on a time axis, real-time rheology of the market environment in the follow-up launch process cannot be perceived, in actual service scenes such as cross-cycle launch or new material cold start, the feature and causal attributes are not constant and unchanged, part of conventional time periods are defined as interference background elements, the specific market trend or marketing nodes are converted into original driving user conversion core elements, the prior art has the feature separation boundary parameters in the model trend training stage, the model trend dynamic change cannot be perceived, and the external environment is enabled to change, when the static environment is converted into the static environment change, the static environment is enabled to be high, the prediction model has high value, and the important information is lost, and the system is caused by the prediction model is lost. Therefore, how to construct and keep shielding the conventional noise, simultaneously, automatically sense the change of the market environment and adaptively adjust the characteristic decoupling boundary, and accurately evaluate the technical scheme of the commercial value of the multi-mode data in the dynamic market becomes the technical problem to be solved by the invention. Disclosure of Invention In order to solve the problems in the background technology, the technical scheme of the invention is as follows, a digital marketing effect prediction evaluation system based on multi-mode data comprises: The feature extraction unit is used for acquiring visual data, text data and attribute data of the object to be put in, and mapping each data into initial feature vectors with uniform dimensions respectively; The characteristic decoupling unit is used for receiving an initial characteristic vector, and respectively outputting a first characteristic component with causal relation of characterization and conversion results and a second characteristic component with statistical correlation of characterization and conversion results through a first mapping channel and a second mapping channel which are arranged in parallel; The residual analysis unit is used for acquiring an externally input real conversion label, calculating a prediction residual between the real conversion label and a prediction value generated based on the first characteristic component, and mapping the prediction residual into a residual state vector; The PID feedback control logic is configured to output negative adjustment increment when the correlation coefficient rises so as to reduce the weight coefficient, and a part of the second characteristic component, which is related to the prediction residual, is reserved in the first characteristic component, so that the constraint strength of the orthogonal constraint item is dynamically adjusted; and the bidding instruction generation unit is used for receiving the first characteristic component which is output by the characteristic decoupling unit and regulated by the weight coefficient, calculating the reference conversion probability, and generating the digital m