E-commerce advertisement image optimization method based on effect data

Abstract

The invention discloses an E-commerce advertisement image optimization method based on effect data, and particularly relates to the technical field of data processing. Advertisement putting effect data including click rate, conversion rate, residence time and user interaction behaviors are obtained from a plurality of advertisement putting platforms, normalization processing is performed on the data, core elements of advertisement images are decomposed based on a computer vision technology, and the influence of each element on the advertisement effect is analyzed. Training the normalized advertisement effect data by using a deep learning model so as to predict expected effects of different advertisement image combination schemes on each delivery platform; based on the prediction result, the system automatically adjusts the visual elements of the advertisement image, and performs iterative optimization in combination with the A / B test strategy, so that the advertisement image optimization scheme can be dynamically adjusted to improve the accurate advertisement putting effect, and the consistency and optimization effect of the advertisement strategy in different putting environments are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to an e-commerce advertising image optimization method based on effect data. Background Art

[0002] With the rapid development of the e-commerce industry, merchants are increasingly relying on online advertising to increase product exposure and sales conversion rates. As an important part of advertising, the design quality of e-commerce advertising images directly affects consumers' purchasing decisions. At present, the optimization of advertising images mainly relies on manual design and experience judgment, lacking systematic and data-driven optimization methods. In recent years, with the development of big data and artificial intelligence technologies, the technology of optimizing advertising images based on user interaction data (such as click-through rate, conversion rate, dwell time, etc.) has gradually attracted attention. By analyzing the advertising effect data, image elements (such as color, layout, product display method, etc.) can be automatically adjusted to improve advertising effectiveness.

[0003] The prior art has the following deficiencies: Since e-commerce ads are usually delivered on multiple platforms (such as social media, search engines, short video platforms, etc.), there are large differences in user behavior patterns, device environments, page layouts, etc. on different platforms, resulting in significant deviations in the performance data of the same advertising images on different platforms. For example, an image may have a high click-through rate on social media, but perform poorly in search engine ads, and even affect the optimization decisions of the overall advertising strategy. Existing optimization methods are often adjusted based on data from a single platform and lack a unified cross-platform optimization mechanism, resulting in inconsistent optimization results in different delivery environments, affecting the overall effectiveness of the advertisement. Summary of the invention

[0004] The purpose of the present invention is to provide an e-commerce advertising image optimization method based on effect data to address the deficiencies in the background technology.

[0005] In order to achieve the above object, the present invention provides the following technical solution: an e-commerce advertising image optimization method based on effect data, comprising: Acquire advertising effect data of advertising images from multiple advertising delivery platforms, wherein the advertising effect data includes click-through rate, conversion rate, dwell time and user interaction behavior, and normalize the advertising effect data of different platforms; Based on computer vision technology, decompose the elements in the advertising image and analyze the impact of each element on the advertising effect, including color, layout, font and product display method; Based on the deep learning model, the normalized advertising effect data is used to train the optimization model and predict the expected effects of different advertising image combinations on various delivery platforms; Automatically adjust the visual elements of the ad image based on the prediction results, and continuously iterate and optimize based on the A / B testing strategy; Based on the real-time optimization of the visual elements of advertising images and changes in users' personalized preferences, the optimization plan of advertising images is dynamically adjusted to improve the precise delivery of advertisements.

[0006] Preferably, the decomposing advertising image elements based on computer vision technology includes: Use K-Means clustering or principal component analysis to extract the color themes of advertising images and calculate color contrast, saturation, and brightness; Use object detection models to identify key visual elements in advertising images, including products, text, and background, and calculate their layout ratios in the image; Use OCR technology to extract text information from advertisements and analyze the impact of font type, color, and readability on ad click-through rate.

[0007] Preferably, the optimization model based on the deep learning model adopts a deep neural network, and its training process includes: Use pre-trained CNN to extract deep features of advertising images and convert them into vectors of fixed length; combine the characteristics of the advertising platform and user behavior characteristics for comprehensive modeling; When training the optimization model, the mean square error is used as the loss function, and the Adam optimizer is used for gradient update to improve the prediction accuracy.

[0008] Preferably, the optimizing the advertisement image based on the A / B testing strategy includes: Generate multiple different versions of advertising image optimization solutions, and use the multi-armed bandit algorithm to dynamically allocate traffic and test the click-through rate and conversion rate of each version of the advertisement; Use t-test or Bayesian optimization method to analyze the test results to determine whether the optimized solution is statistically significantly better than the benchmark solution; Based on the A / B test results, the optimal advertising image version is automatically selected for delivery, and the advertising strategy is continuously optimized in combination with the reinforcement learning model.

[0009] Preferably, the user's interest attenuation factor and visual stimulus adaptation index are obtained and input into a machine learning model for comprehensive calculation to obtain the optimization effect value of the advertising strategy, and the optimization effect value is compared with a preset optimization effect standard threshold, and it is determined whether the optimization plan of the advertising image needs to be adjusted based on the comparison result.

[0010] Preferably, the method for obtaining the interest attenuation factor is: collecting the user's advertising interaction data in the past period of time, setting the user's interaction degree with the advertisement at time t as , set the time window T; the smoothing factor α determines the influence ratio of new data and historical data, and the interest decay factor is calculated using exponentially weighted moving average, expressed as: ;in: is the interest decay factor of the previous time step, and IDF is the interest decay factor.

[0011] Preferably, the method for obtaining the visual stimulus adaptation index is: inputting the advertisement image I into the pre-trained saliency model: ; S is a saliency heat map, which indicates the visual attractiveness of different regions of the image. It is a saliency prediction model that calculates the proportion of high saliency areas in the saliency heat map , the expression is: ; is the indicator function, N is the total number of pixels in the image, is the significance score of pixel (i, j) in the significance heat map, is the saliency threshold; calculate the global color contrast of the image , the expression is: ; is the brightness value of pixel i, is the global brightness mean; OCR is used to extract the advertisement text and calculate the text contrast , the expression is: ; is the average color contrast of the text area, is the average color contrast of the background area; The calculated global color contrast and text contrast are normalized, and the visual stimulus adaptation index is obtained by weighted average summation of the normalized global color contrast and text contrast.

[0012] Preferably, the interest attenuation factor and the visual stimulus adaptation index are converted into a comprehensive feature vector, and the comprehensive feature vector is used as the input of the machine learning model. The machine learning model uses each group of comprehensive feature vectors to predict the optimization effect value label of the advertising strategy as the prediction target, and takes minimizing the sum of prediction errors of the optimization effect value labels of all advertising strategies as the training target. The machine learning model is trained until the sum of prediction errors converges and the model training is stopped. The optimization effect value of the advertising strategy is determined according to the model output results, wherein the machine learning model is a polynomial regression model.

[0013] Preferably, the obtained optimization effect value is compared with a preset optimization effect standard threshold. If the optimization effect value is greater than or equal to the preset optimization effect standard threshold, it means that the advertising optimization strategy is effective and the current optimization plan continues to be used; if the optimization effect value is less than the preset optimization effect standard threshold, it means that the advertising optimization strategy is invalid and the advertising image design needs to be adjusted.

[0014] In the above technical solution, the technical effects and advantages provided by the present invention are: 1. The present invention realizes intelligent optimization of cross-platform advertising delivery effects through key technologies such as multi-platform advertising data normalization, computer vision analysis, deep learning prediction, A / B test optimization, and personalized adjustment. Compared with existing methods, the present invention breaks through the limitations of single platform optimization, effectively solves the problem of data differences between different advertising delivery platforms, and makes the optimization results consistent across platforms. In addition, the present invention predicts the expected effects of different advertising image combination schemes based on deep learning models, and dynamically adjusts the visual elements of advertising images in combination with reinforcement learning strategies, making advertising delivery more accurate and improving advertising click-through rate (CTR) and conversion rate (CVR).

[0015] 2. The present invention also combines the user interest decay factor (IDF) and the visual stimulation adaptation index (GAH) to construct a personalized optimization strategy to ensure that the advertising content can adapt to changes in user interests in real time and improve the attractiveness and long-term effect of the advertisement. The optimization effect value of the advertising strategy is calculated through a machine learning model, and the need to further adjust the advertising image is determined based on the optimization effect standard threshold, so that the optimization process has the characteristics of data-driven, real-time response, and adaptive optimization. Overall, the present invention realizes a precise, intelligent, and automated advertising optimization solution, which can greatly improve the effect of advertising delivery, reduce the cost of invalid delivery, and enhance the competitiveness of advertisers on different platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0017] Figure 1 The present invention is a flow chart of the method. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] For examples, see Figure 1 As shown, the e-commerce advertising image optimization method based on effect data described in this embodiment includes: Acquire advertising effect data of advertising images from multiple advertising delivery platforms, wherein the advertising effect data includes click-through rate, conversion rate, dwell time and user interaction behavior, and normalize the advertising effect data of different platforms; Based on computer vision technology, decompose the elements in the advertising image and analyze the impact of each element on the advertising effect, including color, layout, font and product display method; Based on the deep learning model, the normalized advertising effect data is used to train the optimization model and predict the expected effects of different advertising image combinations on various delivery platforms; Automatically adjust the visual elements of the ad image based on the prediction results, and continuously iterate and optimize based on the A / B testing strategy; Based on the real-time optimization of the visual elements of advertising images and changes in users' personalized preferences, the optimization plan of advertising images is dynamically adjusted to improve the precise delivery of advertisements.

[0020] The method for obtaining and normalizing advertising image delivery effect data from multiple advertising delivery platforms comprises: Select multiple advertising platforms (including but not limited to social media, search engines, short video platforms, etc.). Obtain advertising effectiveness data on each platform through the API interface, data crawling tools or advertising management background provided by the platform. The collected advertising effectiveness data includes but is not limited to: Click-Through Rate (CTR): the ratio of the number of times users click on an ad to the number of times the ad is exposed. Conversion Rate (CVR): the proportion of users who complete the target behavior (such as purchase, registration, etc.) after clicking on the ad. Dwell Time: the average length of time a user stays on an ad page or related link. User Interaction: user behavior data on ads, such as likes, shares, comments, swipe browsing, etc.

[0021] Preprocess and clean the data, delete duplicate records, and ensure data uniqueness. Identify and remove outliers, such as extreme click-through rates (too high or too low), invalid clicks (such as robot clicks), etc. Unify the data time format of different platforms to ensure data time sequence consistency. Use interpolation, mean filling, or machine learning methods (such as KNN filling) to fill in missing data to ensure data integrity.

[0022] Since the indicator range and calculation method of different platforms may be different, standardization is required to make the data comparable. Use the minimum-maximum normalization method to normalize all values ​​to the range of [0,1] or use Z-Score standardization to calculate the standard score of the data so that its mean is 0 and the standard deviation is 1. Establish a cross-platform data comparison table to ensure that similar indicators on different platforms can be matched accordingly, for example: the interaction rate of platform A ≈ the likes + sharing ratio of platform B.

[0023] Store processed normalized data for subsequent machine learning model training or ad image optimization. Use data visualization tools (such as Power BI, Tableau) to monitor cross-platform delivery effects and provide optimization strategy support. Regularly retrain normalized models to ensure that data processing methods adapt to the ever-changing ad delivery environment.

[0024] The method of decomposing advertising image elements and analyzing the influence of each element on the advertising effect based on computer vision technology includes: Obtain advertising images that have been delivered from multiple advertising delivery platforms (social media, search engines, short video platforms, etc.) and collect the corresponding delivery effect data (such as click-through rate, conversion rate, etc.). Convert the formats of advertising images on different platforms (such as JPEG and PNG to standard RGB format) to ensure the consistency of image data. Use existing manually annotated datasets or adopt semi-supervised learning methods to annotate the main visual elements in advertising images (such as background color, product proportion, text area, etc.).

[0025] Computer vision technology is used to decompose the advertising images and extract the following key elements: Color feature extraction: Use color histogram analysis to count the main color distribution of advertising images. Use K-means clustering or principal component analysis (PCA) to extract the main color themes and classify them into categories such as "warm tones", "cold tones", and "high contrast". Calculate image parameters such as color contrast, saturation, and brightness to analyze the impact of different color combinations on advertising appeal.

[0026] Layout analysis: Detect the spatial distribution of different elements in the image (such as products, text, and background) through object detection algorithms (such as YOLO and Faster R-CNN). Calculate the visual hierarchy in the image (such as golden ratio layout, symmetry, center of gravity offset, etc.) and analyze the impact of different layout methods on user attention distribution. Use saliency detection technology to identify the areas that users are most likely to pay attention to, and perform correlation analysis with advertising CTR data.

[0027] Font analysis: Use OCR (optical character recognition) technology to extract text content from advertising images and identify font type, size, color, etc. Use text sentiment analysis technology to evaluate the impact of different font styles (such as bold, handwritten, modern style, etc.) on the emotional communication of advertisements. Count the CTR and conversion rate of advertisements with different font styles and analyze their correlation with user click behavior.

[0028] Analysis of product display methods: Use instance segmentation technology (such as Mask R-CNN) to separate the product and background in the image, and calculate the proportion of the product in the advertising image. Use pose estimation to analyze the actions and postures of the human model in the advertising image to determine their impact on user emotions and purchase intentions. Calculate factors such as product shadows, perspective angles, and lighting effects to analyze their impact on user perception.

[0029] Construct an advertising image feature dataset: Match the extracted advertising image elements (color, layout, font, product display method) with its advertising effect data (click-through rate, conversion rate, dwell time, etc.) to form a structured dataset.

[0030] Feature Importance Analysis: Use machine learning models (such as decision trees, XGBoost, and deep neural networks) to evaluate the contribution of different advertising elements to the effectiveness of advertising. Calculate the weights of color, layout, font, and product display to find out which elements have the greatest impact on CTR or CVR. Use SHAP (SHapley Additive exPlanations) to analyze model output and explain the specific impact of advertising elements on user click behavior. Cluster Analysis and Pattern Discovery: Use K-Means or DBSCAN cluster analysis to analyze different visual styles of advertising images and study the differences in the effectiveness of each style. Find out which advertising image patterns (such as "high contrast + centered layout + large fonts") are more popular with users, and use the results to optimize design strategies.

[0031] Based on the deep learning model, the normalized advertising effect data is used to train the optimization model and predict the expected effects of different advertising image combinations on various delivery platforms, including: DNN is suitable for processing multi-dimensional input data, and can learn the complex features of advertising images and predict their delivery effects. The model consists of multiple fully connected layers and can be used for regression tasks (such as click-through rate prediction and conversion rate prediction).

[0032] Obtain advertising images and their delivery effect data from multiple advertising delivery platforms (social media, search engines, short video platforms, etc.).

[0033] The normalized advertising effect data includes: advertising image features (color theme, layout structure, font style, product display method), delivery platform features (delivery platform type, user group, terminal device), user behavior features (click-through rate CTR, conversion rate CVR, dwell time Dwell Time, interaction rate Interaction Rate); Numerical features (such as click-through rate and conversion rate) are normalized to [0, 1] using Min-Max. Categorical features (such as delivery platform and ad type) are converted using one-hot encoding or embedding vectors. Image data uses pre-trained CNNs (such as ResNet) to extract deep features and convert them into fixed-length feature vectors.

[0034] Build a deep neural network model, including: Input features include: normalized numerical features (such as CTR, CVR); one-hot encoded or embedded categorical features (such as delivery platform type); Ad image features extracted by pre-trained CNN, including: Hidden layer: The first hidden layer (128 neurons, ReLU activation) is responsible for learning the complex relationship between the ad image and the delivery platform. The second hidden layer (64 neurons, ReLU activation) further extracts the impact of ad elements on click-through rate and conversion rate. The third hidden layer (32 neurons, ReLU activation + Dropout 0.3) prevents overfitting and improves generalization ability.

[0035] Output layer: Use linear activation function for regression prediction and output key indicators such as expected click-through rate (CTR) and conversion rate (CVR) of advertising images on different delivery platforms.

[0036] Loss function: The mean square error (MSE) is used as the loss function: ;in, is the real click-through rate / conversion rate, is the model prediction value.

[0037] Optimizer: Adam is used for gradient optimization to improve training stability.

[0038] Training set / validation set division: Use 80% of the data as the training set and 20% as the validation set.

[0039] Batch training: Use Mini-Batch SGD, training 128 samples at a time to accelerate convergence.

[0040] Iterations: The number of training rounds is controlled according to the validation set error, usually set to 50~100 rounds.

[0041] Enter different advertising image features (color, layout, font, product display method) and delivery platform information as input for the new advertising plan.

[0042] The DNN model predicts the expected click-through rate (CTR) and conversion rate (CVR) of this advertising image combination on different delivery platforms.

[0043] Select the advertising combination with the highest predicted CTR and CVR for delivery.

[0044] Use reinforcement learning strategies to continuously update the optimal advertising plan.

[0045] Based on the prediction results of the deep learning model, the visual elements of the advertising image are automatically adjusted, and the A / B testing strategy is used for continuous iterative optimization to improve the overall effect of the advertising.

[0046] Input the advertising image combination plan and use deep learning models (such as DNN, CNN, etc.) to predict key indicators such as the expected click-through rate (CTR), conversion rate (CVR) and other key indicators of the advertising images on different delivery platforms.

[0047] If the predicted CTR or CVR is lower than the historical average, you need to optimize the image elements. If the effect of a certain delivery platform is low (for example, social media CTR is high, search advertising CTR is low), you need to optimize the visual design of the ad for that platform. Combined with user portrait data, determine whether personalized optimization is needed (such as color preferences of different groups of people).

[0048] Based on model analysis, the following visual elements are automatically optimized: Color optimization: Analyze the main color of the current ad image through computer vision techniques (such as color histogram and K-Means clustering). Refer to the best practices in the advertising industry (such as "warm colors increase click-through rate") to adjust the main color and contrast. Use generative adversarial networks (GAN) to generate different color versions for subsequent testing.

[0049] Layout optimization: Use object detection (YOLO, Faster R-CNN) to identify the position of products, text, background and other elements. Adjust the placement of products or texts based on user visual focus analysis (such as saliency map) to enhance appeal. Refer to the layout mode of high CTR ads (such as "center layout vs. side layout") to automatically adjust the layout of elements.

[0050] Font optimization: Use OCR (optical character recognition) to identify existing advertising texts and analyze their fonts, sizes, and colors. Combined with advertising samples with high CTR, it is recommended to change to a more attractive font style (such as bold vs. thin, sans serif vs. handwriting). Use NLP (natural language processing) technology to analyze the emotional tendency of advertising sentences and optimize the content of advertising copy.

[0051] Optimize product display: Use instance segmentation (Mask R-CNN) to separate products and backgrounds to adjust the proportion of products in advertising images. If historical data shows that "panoramic display" has a low conversion rate, while "close-up display" has a higher conversion rate, the product display method will be automatically switched. GAN is used to generate multiple product display angles for subsequent testing to select the best solution.

[0052] Personalized optimization (for different user groups): Combine user profiles (such as age, gender, and interests) to generate multiple personalized ad versions. For example, young users prefer dynamic colors and high-contrast designs, while older users prefer simple designs.

[0053] Create an A / B test group: Generate multiple optimized ad versions (version A: original ad, version B: optimized ad). Divide user traffic to ensure that the audiences of the A / B groups are similar and the test is fair. 50% of users see version A (unoptimized) and 50% see version B (optimized).

[0054] Place A / B group ads on various advertising platforms and collect key data such as CTR, CVR, user interaction rate, etc. Monitor user behavior data, such as click area (heat map analysis), dwell time, conversion path, and evaluate the effect of advertising optimization.

[0055] Calculate the improvement of version B compared to version A (e.g., CTR increased by 15%). Use hypothesis tests (e.g., t-tests) to determine whether the test results are statistically significant (p<0.05). If the optimized version (version B) performs significantly better than the original version (version A), use it as the new benchmark ad.

[0056] Feed the latest data from A / B testing to the deep learning model to optimize the ad image prediction algorithm. Use reinforcement learning (such as DQN) to continuously optimize ad design decisions.

[0057] If the A / B test is successful, the ad delivery weight will be automatically adjusted to increase the delivery ratio of the optimized version. The multi-armed bandit algorithm is used to dynamically allocate budget to the best performing ad version to maximize ROI (return on investment).

[0058] Compare A / B test results between different delivery platforms and adjust the advertising design strategies for different platforms. For example: Social media CTR is high → add interactive elements (such as like and share buttons); search ads CTR is low → simplify the design and improve product readability.

[0059] Based on the real-time optimization of the visual elements of advertising images and changes in users' personalized preferences, the optimization plan of advertising images is dynamically adjusted to improve the precise delivery of advertisements.

[0060] During the advertising process, the optimization of advertising images needs to be real-time (quickly adapt to changes in the market and user behavior) and adapt to changes in user personalized preferences over time. Therefore, by obtaining the user's interest attenuation factor and visual stimulation adaptation index, and inputting them into the machine learning model for comprehensive calculation, the optimization effect value of the advertising strategy is obtained, and the optimization effect value is compared with the preset optimization effect standard threshold, and it is determined whether the optimization plan of the advertising image needs to be adjusted based on the comparison results.

[0061] The method for obtaining the interest decay factor is to collect the user's advertising interaction data in the past period of time, and set the user's interaction degree with the advertisement at time t (such as click-through rate CTR, dwell time, interaction rate, etc.) as , set the time window T (such as the last 7 days or 30 days) to calculate the interest decay factor.

[0062] The smoothing factor α determines the influence ratio of new data to historical data, 0<α≤1. The larger the α (such as 0.5~0.9), the greater the influence of recent data, which is suitable for short-term interest changes. The smaller the α (such as 0.1~0.3), the greater the influence of historical data, which is suitable for long-term interest trend analysis.

[0063] The interest attenuation factor is calculated using exponentially weighted moving average, and the expression is: ;in: is the interest decay factor of the previous time step, and IDF is the interest decay factor.

[0064] The method for obtaining the visual stimulus adaptation index is as follows: input the advertisement image I into the pre-trained saliency model: ; S is the saliency heat map, which indicates the visual attractiveness of different regions of the image (pixel level). It is a saliency prediction model, such as DeepGaze II (based on CNN+LSTM, simulating human visual attention). SAM-ResNet (based on ResNet, optimizing saliency prediction).

[0065] Calculate the percentage of high-significance areas in the saliency heatmap , the expression is: ; is an indicator function that determines whether the pixel (i, j) is above the significance threshold (such as the top 20% most significant area), N is the total number of pixels in the image, is the significance score of pixel (i, j) in the significance heat map, is the significance threshold, which indicates the judgment standard of high significance (for example, the threshold point of the top 20% maximum value can be taken).

[0066] Calculate the global color contrast of an image , the expression is: ; is the brightness value of pixel i, is the global brightness mean.

[0067] Use OCR (Optical Character Recognition) to extract advertisement text and calculate text contrast , the expression is: ; is the average color contrast of the text area, is the average color contrast of the background area.

[0068] The calculated global color contrast and text contrast are normalized, and the visual stimulus adaptation index is obtained by weighted average summation of the normalized global color contrast and text contrast.

[0069] The interest attenuation factor and the visual stimulation adaptation index are converted into a comprehensive feature vector, and the comprehensive feature vector is used as the input of the machine learning model. The machine learning model predicts the optimization effect value label of the advertising strategy for each group of comprehensive feature vectors as the prediction target, and minimizes the sum of prediction errors of the optimization effect value labels of all advertising strategies as the training target. The machine learning model is trained until the sum of prediction errors converges and the model training is stopped. The optimization effect value of the advertising strategy is determined according to the model output results, wherein the machine learning model is a polynomial regression model.

[0070] The method for obtaining the optimization effect value of the advertising strategy is to obtain the corresponding function expression from the comprehensive feature vector training data of the trained machine learning model: ; In the formula, is the output function of the model, IDF is the interest attenuation factor, GAH is the visual stimulus adaptation index, The optimization effect value of the advertising strategy.

[0071] Compare the obtained optimization effect value with the preset optimization effect standard threshold. If the optimization effect value is greater than or equal to the preset optimization effect standard threshold, it means that the advertising optimization strategy is effective and the current optimization plan continues to be used. If the optimization effect value is less than the preset optimization effect standard threshold, it means that the advertising optimization strategy is invalid and the advertising design needs to be adjusted, including optimizing color, layout, and text readability. Adjust the delivery strategy (change the target population, optimize the delivery time). Retrain the model and update the optimization strategy.

[0072] Dynamically adjust the optimization plan for advertising images to improve the accuracy of advertising delivery, specifically: If IDF decreases but GAH remains stable, it means that users are interested in the ad content but need something new. Simply adjust the color or font.

[0073] If both IDF and GAH decrease, it means that the advertisement needs more substantial adjustments, such as revising the layout and adjusting the product display method.

[0074] If IDF remains stable but GAH decreases, it means that the visual design is outdated but user interest still exists. You can adjust the contrast, saturation, lighting, shadows, etc. to enhance visual stimulation.

[0075] Dynamically optimize ad images: Generative adversarial networks (GANs) are used to generate multiple visual style optimization solutions. Combined with A / B testing, the optimal adjustment solution is determined. Reinforcement learning (such as DQN) is used to continuously update ad optimization strategies to adapt to market changes.

[0076] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0077] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0078] It should be understood that the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there may be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. In addition, the character " / " in this article generally indicates that the objects associated with each other are in an "or" relationship, but it may also indicate an "and / or" relationship, which can be understood by referring to the context. A person of ordinary skill in the art can appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this article can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0079] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. An e-commerce advertising image optimization method based on effect data, characterized in that: include: Acquire advertising effect data of advertising images from multiple advertising delivery platforms, wherein the advertising effect data includes click-through rate, conversion rate, dwell time and user interaction behavior, and normalize the advertising effect data of different platforms; Based on computer vision technology, decompose the elements in the advertising image and analyze the impact of each element on the advertising effect, including color, layout, font and product display method; Based on the deep learning model, the normalized advertising effect data is used to train the optimization model and predict the expected effects of different advertising image combinations on various delivery platforms; Automatically adjust the visual elements of the ad image based on the prediction results, and continuously iterate and optimize based on the A / B testing strategy; Based on the real-time optimization of the visual elements of advertising images and changes in users' personalized preferences, the optimization plan of advertising images is dynamically adjusted to improve the precise delivery of advertisements.

2. The e-commerce advertising image optimization method based on effect data according to claim 1, characterized in that: The decomposition of advertising image elements based on computer vision technology includes: Use K-Means clustering or principal component analysis to extract the color themes of advertising images and calculate color contrast, saturation, and brightness; Use object detection models to identify key visual elements in advertising images, including products, text, and background, and calculate their layout ratios in the image; Use OCR technology to extract text information from advertisements and analyze the impact of font type, color, and readability on ad click-through rate.

3. The e-commerce advertising image optimization method based on effect data according to claim 2 is characterized by: The optimization model based on the deep learning model adopts a deep neural network, and its training process includes: Use pre-trained CNN to extract deep features of advertising images and convert them into vectors of fixed length; combine the characteristics of the advertising platform and user behavior characteristics for comprehensive modeling; When training the optimization model, the mean square error is used as the loss function, and the Adam optimizer is used for gradient update to improve the prediction accuracy.

4. The e-commerce advertising image optimization method based on effect data according to claim 1, characterized in that: The optimization of the advertisement image based on the A / B testing strategy includes: Generate multiple different versions of advertising image optimization solutions, and use the multi-armed bandit algorithm to dynamically allocate traffic and test the click-through rate and conversion rate of each version of the advertisement; Use t-test or Bayesian optimization method to analyze the test results to determine whether the optimized solution is statistically significantly better than the benchmark solution; Based on the A / B test results, the optimal advertising image version is automatically selected for delivery, and the advertising strategy is continuously optimized in combination with the reinforcement learning model.

5. The e-commerce advertising image optimization method based on effect data according to claim 4 is characterized by: By obtaining the user's interest attenuation factor and visual stimulation adaptation index and inputting them into the machine learning model for comprehensive calculation, the optimization effect value of the advertising strategy is obtained, and the optimization effect value is compared with the preset optimization effect standard threshold. According to the comparison result, it is determined whether the optimization plan of the advertising image needs to be adjusted.

6. The e-commerce advertising image optimization method based on effect data according to claim 5 is characterized by: The method for obtaining the interest decay factor is to collect the user's advertising interaction data in the past period of time, and set the user's interaction degree with the advertisement at time t as , set the time window T; the smoothing factor α determines the impact ratio of new data and historical data, and the interest decay factor is calculated using exponentially weighted moving average, expressed as: ;in: is the interest decay factor of the previous time step, and IDF is the interest decay factor.

7. The e-commerce advertising image optimization method based on effect data according to claim 6 is characterized by: The method for obtaining the visual stimulus adaptation index is as follows: input the advertisement image I into the pre-trained saliency model: ; S is a saliency heat map, which indicates the visual attractiveness of different regions of the image. It is a saliency prediction model that calculates the proportion of high saliency areas in the saliency heat map , the expression is: ; is the indicator function, N is the total number of pixels in the image, is the significance score of pixel (i, j) in the significance heat map, is the significance threshold; Calculate the global color contrast of an image , the expression is: ; is the brightness value of pixel i, is the global brightness mean; OCR is used to extract the advertisement text and calculate the text contrast , the expression is: ; is the average color contrast of the text area, is the average color contrast of the background area; The calculated global color contrast and text contrast are normalized, and the visual stimulus adaptation index is obtained by weighted average summation of the normalized global color contrast and text contrast.

8. The e-commerce advertising image optimization method based on effect data according to claim 7 is characterized by: The interest attenuation factor and the visual stimulation adaptation index are converted into a comprehensive feature vector, and the comprehensive feature vector is used as the input of the machine learning model. The machine learning model predicts the optimization effect value label of the advertising strategy for each group of comprehensive feature vectors as the prediction target, and minimizes the sum of prediction errors of the optimization effect value labels of all advertising strategies as the training target. The machine learning model is trained until the sum of prediction errors converges and the model training is stopped. The optimization effect value of the advertising strategy is determined according to the model output results, wherein the machine learning model is a polynomial regression model.

9. The e-commerce advertising image optimization method based on effect data according to claim 8, characterized in that: The obtained optimization effect value is compared with the preset optimization effect standard threshold. If the optimization effect value is greater than or equal to the preset optimization effect standard threshold, it means that the advertising optimization strategy is effective and the current optimization plan continues to be used; if the optimization effect value is less than the preset optimization effect standard threshold, it means that the advertising optimization strategy is invalid and the advertising image design needs to be adjusted.

Images