Data model training method and system based on display scheme

Abstract

The invention discloses a data model training method and system based on a display scheme. The method comprises the following steps: S1, constructing a multi-source situation data set; s2, preprocessing the multi-source situation data set, and integrating the multi-source situation data set into a single situation vector; s3, generating a user preference portrait, and fusing the situation vector and the user preference portrait to generate a dynamic situation preference portrait; s4, inputting the dynamic situation preference portrait into a conditional generative adversarial network to generate a display scheme; s5, embedding an adaptive neural regulation and control module, and optimizing a display scheme generation effect; s6, introducing a hierarchical situation control mechanism; s7, iteratively updating the user preference portrait based on historical interaction feedback; and S8, analyzing the similarity between the current situation and the historical situation, judging the applicability of the display scheme by applying a meta-learning algorithm, and migrating the effective display scheme in the historical situation to the current situation. According to the method, a multi-dimensional context awareness and dynamic feedback optimization model training method is adopted, and context self-adaptive generation of the display scheme is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of data model training, and in particular to a display scheme-based data model training method and system. Background Art

[0002] In the current field of display technology, personalized recommendations and intelligent display have become important means to improve user experience. Traditional display systems mainly rely on static rules or display algorithms based on simple user preferences, and usually use basic user information (such as age, gender, browsing history, etc.) for simple content recommendations and display layouts. However, this static and relatively single display method has many shortcomings in user experience, especially in multi-situational and highly interactive scenarios. It is difficult to fully adapt to the diverse needs of users. With the development of smart devices and Internet of Things technology, user interaction behavior data, device information and environmental parameters can be collected in real time, which provides a basis for further optimization of intelligent display systems. However, the existing technology still has many shortcomings in making full use of multi-dimensional situational data, dynamic feedback and multi-level control.

[0003] The shortcomings of traditional display systems are mainly reflected in contextual adaptation and personalized needs. Display systems usually find it difficult to perceive and adapt to changes in user contexts in real time. Existing display systems usually only consider users' historical data and ignore users' immediate needs in different contexts. For example, the preferences of the same user may vary during the day and night, indoors and outdoors, but existing display systems find it difficult to flexibly capture and adapt to these changes. In addition, users' interest preferences may change over time, and the recommendation algorithms of traditional systems often lack the recognition and adaptive update mechanism for user interest fluctuations, making it difficult to adjust recommended content in a timely manner. Therefore, after users use the same system for a long time, the matching degree of the displayed content may gradually decrease, resulting in reduced user experience and user satisfaction.

[0004] In the field of intelligent display, the application of generative adversarial networks (GAN) and deep learning technologies has been gradually introduced to generate diverse display content. However, the application of existing GANs also has limitations, mainly the lack of dynamic adaptability to the context. Most GAN models rely on fixed training data sets and model parameters when generating content, and it is difficult to adjust the characteristics of the generated content in real time according to the user's current context. In actual scenarios, users' interactive behaviors (such as clicks, dwell time, sliding behaviors, etc.) can provide rich feedback information, which can guide the generator to adjust the content generation strategy to better meet the user's immediate needs. However, the existing GAN generation technology has not fully utilized the user's interactive feedback to dynamically update the generation model, which makes the generated content insufficient in matching the user context in real time.

[0005] In addition, traditional display systems usually adopt a single-layer situational control method, which makes it difficult to achieve hierarchical control in complex situations. User needs often have multi-level characteristics, and the needs for the theme, style, and details of the display content may be different in different situations. For example, users need concise and clear information display in information situations, but may prefer visual product displays in shopping situations; users have high requirements for brightness in daytime environments, but may prefer dark themes when browsing at night. Existing display systems can usually only perform single control based on the user's main needs, and cannot perform layer-by-layer matching on multiple levels such as display themes, styles, and details, resulting in insufficient adaptability of display content in specific scenarios, and failure to maximize user experience and interactive effects.

[0006] In terms of feedback optimization, existing technologies usually lack self-optimization mechanisms, making it difficult to iteratively update the display content generation process based on real-time user behavior feedback. Most display systems lack a complete feedback loop and adaptive optimization mechanism, and usually only provide display content passively, but are unable to optimize recommendation strategies based on user behavior feedback. For example, when a user's click-through rate or dwell time decreases, the system cannot make real-time adjustments to the display content generation strategy, and thus cannot promptly improve the relevance of the content and the user's experience. In addition, existing optimization mechanisms are mainly based on preset rules or simple statistical data, making it difficult to perform in-depth optimization based on complex user behavior patterns. Therefore, the improvement in user experience is usually intermittent, and it is difficult to form continuous adaptive optimization.

[0007] In terms of cross-scenario migration and expansion, the scalability of existing display systems is relatively limited, and it is difficult to quickly adapt to new needs in different scenarios. Each change in the scenario may require the system to retrain and adjust the display model, which consumes a lot of time and resources, making it difficult for the display solution to quickly adapt in dynamic scenarios. The introduction of meta-learning and transfer learning technologies can improve the adaptability of the model to a certain extent, but the application of these technologies in existing systems is relatively preliminary and cannot achieve rapid migration and expansion of display solutions. For example, in similar scenarios, the current display system cannot quickly migrate historical display strategies to new scenarios, resulting in low adaptability of the system in new scenarios. This not only increases the training overhead of the system in multiple scenarios, but also limits the display system's ability to respond quickly under multiple scenarios and multiple demand conditions.

[0008] Therefore, how to provide a data model training method and system based on a display scheme is an urgent problem that those skilled in the art need to solve. Summary of the invention

[0009] One purpose of the present invention is to propose a method and system for training a display scheme data model. The present invention achieves a high degree of matching between display content and user context through multi-dimensional context perception, adaptive generation and feedback optimization methods. It adopts a combination of hierarchical context control and conditional generative adversarial networks to adjust the generated content in real time according to the user's click rate, dwell time and sliding behavior, achieving a personalized and adaptive display effect. At the same time, combined with meta-learning and transfer learning technology, the system can quickly respond to user needs in multiple contexts, with the advantages of high matching, fast response speed and strong scalability, providing an innovative solution for intelligent display.

[0010] A method for training a display scheme data model according to an embodiment of the present invention includes the following steps: S1. Collect users’ micro-context data through IoT devices, extract macro-context data from user device information, and construct a multi-source context data set; S2. Preprocess the multi-source context dataset, weight the context features from different sources through a multi-head attention mechanism, and integrate the multi-source context features into a single context vector; S3, extract preference features from the user's historical interaction data, generate a user preference portrait, and fuse the context vector with the user preference portrait to generate a dynamic context preference portrait; S4, inputting the generated dynamic situation preference portrait into the conditional generative adversarial network as a condition to generate a display plan that meets the current situation; S5. Embed an adaptive neural control module, use the context feature feedback loop to dynamically adjust the generation strategy of the conditional generative adversarial network, adjust the weight of the generation model in real time according to the user's click rate, dwell time and sliding behavior, and optimize the display solution generation effect; S6. Introducing a hierarchical situation control mechanism, wherein the hierarchical situation control mechanism includes macro situation control, meso situation control and micro situation control; S7, collecting feedback data of user interactions, inputting the feedback data into the dynamic situational preference profile and adaptive neural control module, adjusting the weight parameters of the conditional generative adversarial network, and iteratively updating the user preference profile based on historical interaction feedback; S8. Analyze the similarity between the current situation and the historical situation through situation similarity detection, apply the meta-learning algorithm to determine the applicability of the display plan, and use transfer learning to transfer the effective display plan in the historical situation to the current situation.

[0011] Optionally, the S4 specifically includes: S41. Obtaining dynamic situational preference portraits ,in Table ith feature dimension value in the dynamic situation preference portrait; S42. The dynamic situation preference portrait V is input into the generator of the conditional generative adversarial network. The generator generates a preliminary display content vector C by mapping V into a hidden layer vector Z and combining it with the generated noise vector N: ; Among them, tanh represents the hyperbolic tangent activation function, represents the adaptive adjustment parameter, The weight matrix representing the dynamic situational preference profile, represents the weight matrix of the hidden layer vector, represents the weight matrix of the noise vector, b represents the bias term, n represents the number of feature dimensions of the dynamic situational preference portrait, m represents the number of dimensions of the hidden layer vector, and p represents the number of dimensions of the noise vector. represents the smoothing term, , and Represent the weight coefficients in the weight matrix of dynamic situation preference portrait, hidden layer vector and noise vector respectively, Represents the jth feature dimension value in the hidden layer vector, Represents the kth feature dimension value in the noise vector; S43, outputting a preliminary display scheme according to the processing of the preliminary display content vector C, wherein the preliminary display scheme includes layout elements, color matching elements and content elements, wherein the layout elements include button positions and image display area size, the color matching elements include background color and text color, and the content elements include type, style and text content of the display content; S44. The output of the generator is evaluated using a discriminator. The input conditions are the dynamic situation preference portrait V and the preliminary display content vector C. The discriminator determines the generation scheme according to the matching degree of the display content and outputs the matching degree score: ; Among them, D(V,C) represents the matching score, exp represents the exponential function, and U represents the weight matrix in the discriminator. represents the context vector weight in the discriminator, represents the bias term, f(V,C) represents the combined feature vector of the dynamic situation preference portrait V and the preliminary display content vector C, and f(V) represents the feature vector of the dynamic situation preference portrait V. Represents each weight item in the context vector weight, Represents the weight items in the preliminary display content vector C, represents each feature value in the preliminary display content vector C, a represents the number of feature dimensions of the context vector, and b represents the number of feature dimensions of the preliminary display content vector C; S45. If the matching score D(V,C) does not reach the set matching threshold, the weight matrix and bias value of the generator are optimized by directional propagation to generate display content that is more in line with the current situation; S46. Through multiple iterations, in the alternating training process of the generator and the discriminator, the display scheme generated by the generator gradually conforms to the dynamic situation preference portrait V, and finally outputs a display scheme that conforms to the current situation.

[0012] S46. Through multiple iterations, in the alternating training process of the generator and the discriminator, the display scheme generated by the generator gradually conforms to the dynamic situation preference portrait V, and finally outputs a display scheme that conforms to the current situation.

[0013] Optionally, the S5 specifically includes: S51. Collect real-time user interaction data, including click-through rate , Dwell time and sliding behavior , the click rate Indicates the user's interest in the displayed content, the dwell time Reflects the user's attention, the sliding behavior Indicates the user's tendency to explore displayed content and interaction area preference; S52. Click-through rate , Dwell time and sliding behavior Standardize and convert to a unified feedback score: ; Among them, F represents the unified feedback score, Represents the weight coefficient of click rate, represents the weight coefficient of the residence time, Represents the weight coefficient of sliding behavior, represents the smoothing coefficient, represents the weight parameter of the k-th layer weight in the conditional generative adversarial network, represents the feature value of the kth layer in the conditional generative adversarial network, and N represents the number of layers of the generator in the conditional generative adversarial network; S53, inputting the unified feedback score F into the adaptive neural control module, and using the unified feedback score F to adjust the weight matrix of the conditional generative adversarial network generator; S54, using a gradient descent method to update in real time the generator weight matrix, the hidden layer vector weight matrix, the noise vector weight matrix and the bias value in the conditional generative adversarial network; S55, introduce the control parameters of the adaptive feedback loop , and ,in Indicates the response sensitivity of the click rate. represents the response sensitivity of the residence time, Indicates the response sensitivity of the sliding behavior; S56. After updating and iterating the weights of the feedback loop, the display solution generation effect is optimized. When the click rate, dwell time and sliding behavior reach a predetermined threshold, it is determined that the generated display solution has met the current situation requirements.

[0014] Optionally, the S6 specifically includes: S61, identify the user's current main demand context, determine the theme and overall direction of the displayed content based on the user's behavior data and preference portrait, and analyze the user's contextual needs through the context weights. Adjust the theme of the displayed content to match the user's needs and situations; S62: Adapt the content style based on the user device information and environmental factors. The meso-scenario control module applies the corresponding style template based on different device and environmental characteristics, and uses the meso-scenario style weight coefficient Controlling style transfer; S63. When the user interacts with the displayed content, the user's click, slide and stay behaviors are detected in real time, and the details of the displayed content are adjusted according to the micro-interaction behaviors. The micro-context control module dynamically updates the displayed content according to the user's real-time feedback, including adjusting the button color, content typesetting and layout, and controlling the weight coefficient through the micro-context Control the adjustment range of display details; S64. Based on the hierarchical context control mechanism, a multi-level fitness score of the displayed content is defined to measure the adaptability of the displayed content at different context levels: ; Among them, A represents the multi-level fitness score of the displayed content, represents the macro-scenario control weight coefficient, represents the macro-scenario fitness, represents the micro-situational fitness, represents the weight parameter of the kth feature in macro context control, represents the kth feature in the macro-context control, H represents the number of features in the macro-context control, represents the weight parameter of the jth feature in the meso-situational control, represents the jth feature in the meso-situational control, J represents the number of features in the meso-situational control, represents the weight parameter of the i-th feature in micro-context control, represents the i-th feature in micro-contextual control, Q represents the number of features in micro-contextual control; S65, when the fitness score A reaches the preset threshold, it is considered that the displayed content has achieved the expected context matching effect at the macro, meso and micro context levels, and it is determined that the displayed content meets the multi-level context needs of the user; S66. If the fitness score A does not reach the preset threshold, the weight coefficients of the macro context control module, the meso context control module and the micro context control module are updated through the back propagation mechanism to gradually optimize the adaptability of the displayed content.

[0015] Optionally, the S7 specifically includes: S71. Collect comprehensive interactive data of users in displayed content, including click-through rate, browsing time, sliding frequency and hot spots, to form multi-dimensional interactive data; S72, converting the collected multi-dimensional interaction data into a context feedback vector, and integrating it with the user's historical portrait data after data cleaning and preprocessing, and generating a personalized vector with time and context specificity by analyzing the user's preference fluctuations in different contexts; S73, the dynamic situation preference portrait module generates a multi-stage user portrait update strategy based on the input of the situation feedback vector, wherein the update strategy separates short-term preferences from long-term preferences based on the user's interaction frequency, time interval, and changes in situational needs, and constructs a preference feature vector that meets the user's recent interests; S74, the adaptive neural control module receives the updated situation preference portrait information and automatically detects the deviation area of ​​the user's interest characteristics; S75. Through the multi-level feedback control mechanism of the adaptive neural control module, preference adjustment weights are further introduced to enable the generator to automatically adjust the display scheme of the displayed content. The preference adjustment weights are based on the user context feedback frequency and context change trend to dynamically adjust the style, color matching and structural display characteristics of the generated content. S76. In multiple rounds of iterative optimization, by recording the generated effect after each user interaction, gradually build a historical preference map in the user portrait, accumulate the user's preference change trajectory, identify the long-term pattern of user preference based on the historical preference map generator, and give priority to the long-term pattern of user preference when generating content.

[0016] Optionally, the S8 specifically includes: S81. Collect current user context data, including the user's real-time behavior characteristics, device information, and environmental parameters, to form a current context feature vector ; S82. Calculate the current situation feature vector based on the situation similarity detection model and the historical context feature vector set The similarity between them is calculated based on the cosine similarity or Euclidean distance measurement method to select the feature vector that is most similar to the current situation; S83. Apply the meta-learning algorithm to determine the adaptability of the display scheme in the historical context to the current context: ; in, represents the applicability score, H represents the number of historical context features, represents the weight of the i-th feature, represents the weight of the jth feature, represents the i-th eigenvalue of the current situation, represents the ith eigenvalue of the historical context, exp represents the exponential function, represents the exponential decay coefficient, represents the smoothing coefficient, represents the nonlinear adjustment coefficient, represents the kth eigenvalue of the current situation, represents the kth eigenvalue of the historical context; S84. For the display scheme determined to be applicable, the transfer learning module transfers the generation model parameters in the historical context to the current context, and the transfer process includes the selection of key parameters, the conversion of feature mapping and the optimization of content features; S85. For the migrated display scheme, the generated model is fine-tuned. The fine-tuning process is based on the differences between the current context and the historical context, and the display strategy, color matching style and interaction details in the model are adjusted; S86. After completing the migration and fine-tuning, the generation effect and user feedback data in the current situation are recorded and saved as new historical situation features.

[0017] A display scheme data model training system according to an embodiment of the present invention includes the following modules: The data collection module is used to collect the user's micro-context data through the IoT device and extract the macro-context data from the user's device information to generate a context data set containing multi-source context features; The data preprocessing module is used to clean, standardize and denoise the context data set, and to perform weighted fusion of context features from different sources through a multi-head attention mechanism to generate a single context vector; The user portrait generation module is used to extract preference features from the user's historical interaction data to generate a user preference portrait, and fuse the context vector with the user preference portrait to generate a dynamic context preference portrait; The conditional generative adversarial network module includes a generator and a discriminator, which is used to receive the dynamic situation preference portrait as a conditional input and generate a display plan that conforms to the current situation; Adaptive neural control module, which is used to dynamically adjust the generation strategy of the conditional generative adversarial network based on the real-time feedback of the user's click rate, dwell time and sliding behavior, and optimize the generation effect of the display solution through the feedback loop; Hierarchical context control module, including macro context control module, meso context control module and micro context control module, used to adjust the display scheme layer by layer according to user needs, device information and real-time interactive behavior; Feedback data processing module, used to collect and process user interaction feedback data, input the feedback data into the dynamic situation preference profile and adaptive neural control module, and iteratively update the user preference profile based on historical interaction feedback; The context similarity detection module is used to analyze the similarity between the current context and the historical context based on the context similarity detection model, determine the applicability of the historical display scheme in combination with the meta-learning algorithm, and transfer the highly applicable historical display scheme to the current context through the transfer learning module; The storage module is used to record and save the generation effect and user feedback data in the current situation after completing the migration and fine-tuning of the display plan, and update the situation feature database.

[0018] The beneficial effects of the present invention are: First, the present invention can collect the user's micro-context data and macro-context data in real time, integrate multi-source context features into a single context vector through a multi-head attention mechanism, and generate a dynamic context preference portrait based on the user's historical interaction data, so as to more accurately understand the user's preferences and current context needs. Through the conditional generative adversarial network, the system can dynamically consider the user's context when generating display content, so that the content is highly matched with the user's needs, which effectively makes up for the shortcomings of traditional display systems in context adaptability.

[0019] Secondly, the present invention introduces a feedback loop mechanism in the conditional generative adversarial network by embedding an adaptive neural control module, which can dynamically adjust the weight of the generation model according to the user's real-time feedback such as click rate, dwell time, and sliding behavior, thereby realizing self-optimization of the display scheme. The introduction of this feedback mechanism enables the system to quickly respond to changes in user behavior and optimize the generation strategy of display content based on interactive feedback, avoiding the drawbacks of relying on fixed rules in traditional systems, achieving continuous adaptive updates, and improving the overall user experience.

[0020] In addition, the application of the hierarchical context control mechanism enables the system to control content at the macro, meso and micro levels, and adjust the theme, style and details of the displayed content layer by layer to meet the diverse needs of users at different context levels. Unlike traditional display systems that usually only focus on a single level of context control, the hierarchical context control mechanism of the present invention greatly improves the adaptability of the displayed content through multi-level regulation at the macro, meso and micro levels, allowing users to obtain display solutions that are more in line with their preferences in different demand scenarios. This multi-level context control breaks the limitations of existing technologies and brings users a smoother and more natural interactive experience.

[0021] In terms of cross-context migration, the present invention uses a context similarity detection model and a meta-learning algorithm to quickly identify the similarity between the current context and the historical context, and migrates the applicable historical display scheme to the new context, thereby achieving rapid generation and adaptation of display content. The application of transfer learning not only reduces the need to retrain the display model in a new context, but also greatly improves the response speed and scalability of the system in multiple contexts. Regardless of how the user context changes, the system can quickly adapt to new display scenarios by migrating historical effective solutions, overcoming the shortcomings of traditional display systems in multi-context adaptability. This flexible adaptability enables the system to always maintain a high degree of matching of display content, while significantly reducing resource overhead and improving the intelligence and efficiency of the display system. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is an overall flow chart of a method for training a display scheme data model proposed by the present invention; Figure 2 This is a structural schematic diagram of a display scheme data model training system proposed by the present invention. DETAILED DESCRIPTION

[0023] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0024] refer to Figure 1 , a method for training a display scheme data model, comprising the following steps: S1. Collect users’ micro-context data through IoT devices, extract macro-context data from user device information, and construct a multi-source context data set; S2. Preprocess the multi-source context dataset, weight the context features from different sources through a multi-head attention mechanism, and integrate the multi-source context features into a single context vector; S3, extract preference features from the user's historical interaction data, generate a user preference portrait, and fuse the context vector with the user preference portrait to generate a dynamic context preference portrait; S4, inputting the generated dynamic situation preference portrait into the conditional generative adversarial network as a condition to generate a display plan that meets the current situation; S5. Embed an adaptive neural control module, use the context feature feedback loop to dynamically adjust the generation strategy of the conditional generative adversarial network, adjust the weight of the generation model in real time according to the user's click rate, dwell time and sliding behavior, and optimize the display solution generation effect; S6. Introducing a hierarchical situation control mechanism, wherein the hierarchical situation control mechanism includes macro situation control, meso situation control and micro situation control; S7, collecting feedback data of user interactions, inputting the feedback data into the dynamic situational preference profile and adaptive neural control module, adjusting the weight parameters of the conditional generative adversarial network, and iteratively updating the user preference profile based on historical interaction feedback; S8. Analyze the similarity between the current situation and the historical situation through situation similarity detection, apply the meta-learning algorithm to determine the applicability of the display plan, and use transfer learning to transfer the effective display plan in the historical situation to the current situation.

[0025] In this implementation, S4 specifically includes: S41. Obtaining dynamic situational preference portraits ,in Table ith feature dimension value in the dynamic situation preference portrait; S42. The dynamic situation preference portrait V is input into the generator of the conditional generative adversarial network. The generator generates a preliminary display content vector C by mapping V into a hidden layer vector Z and combining it with the generated noise vector N: ; Among them, tanh represents the hyperbolic tangent activation function, represents the adaptive adjustment parameter, The weight matrix representing the dynamic situational preference profile, represents the weight matrix of the hidden layer vector, represents the weight matrix of the noise vector, b represents the bias term, n represents the number of feature dimensions of the dynamic situational preference portrait, m represents the number of dimensions of the hidden layer vector, and p represents the number of dimensions of the noise vector. represents the smoothing term, , and Represent the weight coefficients in the weight matrix of dynamic situation preference portrait, hidden layer vector and noise vector respectively, Represents the jth feature dimension value in the hidden layer vector, Represents the kth feature dimension value in the noise vector; S43, outputting a preliminary display scheme according to the processing of the preliminary display content vector C, wherein the preliminary display scheme includes layout elements, color matching elements and content elements, wherein the layout elements include button positions and image display area size, the color matching elements include background color and text color, and the content elements include type, style and text content of the display content; S44. The output of the generator is evaluated using a discriminator. The input conditions are the dynamic situation preference portrait V and the preliminary display content vector C. The discriminator determines the generation scheme according to the matching degree of the display content and outputs the matching degree score: ; Among them, D(V,C) represents the matching score, exp represents the exponential function, and U represents the weight matrix in the discriminator. represents the context vector weight in the discriminator, represents the bias term, f(V,C) represents the combined feature vector of the dynamic situation preference portrait V and the preliminary display content vector C, and f(V) represents the feature vector of the dynamic situation preference portrait V. Represents each weight item in the context vector weight, Represents the weight items in the preliminary display content vector C, represents each feature value in the preliminary display content vector C, a represents the number of feature dimensions of the context vector, and b represents the number of feature dimensions of the preliminary display content vector C; S45. If the matching score D(V,C) does not reach the set matching threshold, the weight matrix and bias value of the generator are optimized by directional propagation to generate display content that is more in line with the current situation; S46. Through multiple iterations, in the alternating training process of the generator and the discriminator, the display scheme generated by the generator gradually conforms to the dynamic situation preference portrait V, and finally outputs a display scheme that conforms to the current situation.

[0026] In this implementation manner, S5 specifically includes: S51. Collect real-time user interaction data, including click-through rate , Dwell time and sliding behavior , the click rate Indicates the user's interest in the displayed content, the dwell time Reflects the user's attention, the sliding behavior Indicates the user's tendency to explore displayed content and interaction area preference; S52. Click-through rate , Dwell time and sliding behavior Standardize and convert to a unified feedback score: ; Among them, F represents the unified feedback score, Represents the weight coefficient of click rate, represents the weight coefficient of the residence time, Represents the weight coefficient of sliding behavior, represents the smoothing coefficient, represents the weight parameter of the k-th layer weight in the conditional generative adversarial network, represents the feature value of the kth layer in the conditional generative adversarial network, and N represents the number of layers of the generator in the conditional generative adversarial network; S53, inputting the unified feedback score F into the adaptive neural control module, and using the unified feedback score F to adjust the weight matrix of the conditional generative adversarial network generator; S54, using a gradient descent method to update in real time the generator weight matrix, the hidden layer vector weight matrix, the noise vector weight matrix and the bias value in the conditional generative adversarial network; S55, introduce the control parameters of the adaptive feedback loop , and ,in Indicates the response sensitivity of the click rate. represents the response sensitivity of the residence time, Indicates the response sensitivity of the sliding behavior; S56. After updating and iterating the weights of the feedback loop, the display solution generation effect is optimized. When the click rate, dwell time and sliding behavior reach a predetermined threshold, it is determined that the generated display solution has met the current situation requirements.

[0027] In this implementation manner, S6 specifically includes: S61, identify the user's current main demand context, determine the theme and overall direction of the displayed content based on the user's behavior data and preference portrait, and analyze the user's contextual needs through the context weights. Adjust the theme of the displayed content to match the user's needs and situations; S62: Adapt the content style based on the user device information and environmental factors. The meso-scenario control module applies the corresponding style template based on different device and environmental characteristics, and uses the meso-scenario style weight coefficient Controlling style transfer; S63. When the user interacts with the displayed content, the user's click, slide and stay behaviors are detected in real time, and the details of the displayed content are adjusted according to the micro-interaction behaviors. The micro-context control module dynamically updates the displayed content according to the user's real-time feedback, including adjusting the button color, content typesetting and layout, and controlling the weight coefficient through the micro-context Control the adjustment range of display details; S64. Based on the hierarchical context control mechanism, a multi-level fitness score of the displayed content is defined to measure the adaptability of the displayed content at different context levels: ; Among them, A represents the multi-level fitness score of the displayed content, represents the macro-scenario control weight coefficient, represents the macro-scenario fitness, represents the micro-situational fitness, represents the weight parameter of the kth feature in macro context control, represents the kth feature in the macro-context control, H represents the number of features in the macro-context control, represents the weight parameter of the jth feature in the meso-situational control, represents the jth feature in the meso-situational control, J represents the number of features in the meso-situational control, represents the weight parameter of the i-th feature in micro-context control, represents the i-th feature in micro-contextual control, Q represents the number of features in micro-contextual control; S65, when the fitness score A reaches the preset threshold, it is considered that the displayed content has achieved the expected context matching effect at the macro, meso and micro context levels, and it is determined that the displayed content meets the multi-level context needs of the user; S66. If the fitness score A does not reach the preset threshold, the weight coefficients of the macro context control module, the meso context control module and the micro context control module are updated through the back propagation mechanism to gradually optimize the adaptability of the displayed content.

[0028] In this implementation manner, the S7 specifically includes: S71. Collect comprehensive interactive data of users in displayed content, including click-through rate, browsing time, sliding frequency and hot spots, to form multi-dimensional interactive data; S72, converting the collected multi-dimensional interaction data into a context feedback vector, and integrating it with the user's historical portrait data after data cleaning and preprocessing, and generating a personalized vector with time and context specificity by analyzing the user's preference fluctuations in different contexts; S73, the dynamic situation preference portrait module generates a multi-stage user portrait update strategy based on the input of the situation feedback vector, wherein the update strategy separates short-term preferences from long-term preferences based on the user's interaction frequency, time interval, and changes in situational needs, and constructs a preference feature vector that meets the user's recent interests; S74, the adaptive neural control module receives the updated situation preference portrait information and automatically detects the deviation area of ​​the user's interest characteristics; S75. Through the multi-level feedback control mechanism of the adaptive neural control module, preference adjustment weights are further introduced to enable the generator to automatically adjust the display scheme of the displayed content. The preference adjustment weights are based on the user context feedback frequency and context change trend to dynamically adjust the style, color matching and structural display characteristics of the generated content. S76. In multiple rounds of iterative optimization, by recording the generated effect after each user interaction, gradually build a historical preference map in the user portrait, accumulate the user's preference change trajectory, identify the long-term pattern of user preference based on the historical preference map generator, and give priority to the long-term pattern of user preference when generating content.

[0029] In this implementation manner, S8 specifically includes: S81. Collect current user context data, including the user's real-time behavior characteristics, device information, and environmental parameters, to form a current context feature vector ; S82. Calculate the current situation feature vector based on the situation similarity detection model and the historical context feature vector set The similarity between them is calculated based on the cosine similarity or Euclidean distance measurement method to select the feature vector that is most similar to the current situation; S83. Apply the meta-learning algorithm to determine the adaptability of the display scheme in the historical context to the current context: ; in, represents the applicability score, H represents the number of historical context features, represents the weight of the i-th feature, represents the weight of the jth feature, represents the i-th eigenvalue of the current situation, represents the ith eigenvalue of the historical context, exp represents the exponential function, represents the exponential decay coefficient, represents the smoothing coefficient, represents the nonlinear adjustment coefficient, represents the kth eigenvalue of the current situation, represents the kth eigenvalue of the historical context; S84. For the display scheme determined to be applicable, the transfer learning module transfers the generation model parameters in the historical context to the current context, and the transfer process includes the selection of key parameters, the conversion of feature mapping and the optimization of content features; S85. For the migrated display scheme, the generated model is fine-tuned. The fine-tuning process is based on the differences between the current context and the historical context, and the display strategy, color matching style and interaction details in the model are adjusted; S86. After completing the migration and fine-tuning, the generation effect and user feedback data in the current situation are recorded and saved as new historical situation features.

[0030] refer to Figure 2 , a display scheme data model training system, including the following modules: The data collection module is used to collect the user's micro-context data through the IoT device and extract the macro-context data from the user's device information to generate a context data set containing multi-source context features; The data preprocessing module is used to clean, standardize and denoise the context data set, and to perform weighted fusion of context features from different sources through a multi-head attention mechanism to generate a single context vector; The user portrait generation module is used to extract preference features from the user's historical interaction data to generate a user preference portrait, and fuse the context vector with the user preference portrait to generate a dynamic context preference portrait; The conditional generative adversarial network module includes a generator and a discriminator, which is used to receive the dynamic situation preference portrait as a conditional input and generate a display plan that conforms to the current situation; Adaptive neural control module, which is used to dynamically adjust the generation strategy of the conditional generative adversarial network based on the real-time feedback of the user's click rate, dwell time and sliding behavior, and optimize the generation effect of the display solution through the feedback loop; Hierarchical context control module, including macro context control module, meso context control module and micro context control module, used to adjust the display scheme layer by layer according to user needs, device information and real-time interactive behavior; Feedback data processing module, used to collect and process user interaction feedback data, input the feedback data into the dynamic situation preference profile and adaptive neural control module, and iteratively update the user preference profile based on historical interaction feedback; The context similarity detection module is used to analyze the similarity between the current context and the historical context based on the context similarity detection model, determine the applicability of the historical display scheme in combination with the meta-learning algorithm, and transfer the highly applicable historical display scheme to the current context through the transfer learning module; The storage module is used to record and save the generation effect and user feedback data in the current situation after completing the migration and fine-tuning of the display plan, and update the situation feature database.

[0031] Embodiment 1: In order to verify the feasibility of the present invention in implementation, the present invention is applied to the personalized shopping display of a large e-commerce platform. The platform has a large number of users and significant personalized needs. Different users expect to obtain different shopping content and display methods in multiple scenarios such as time, place, and equipment. The traditional display system of the e-commerce platform is difficult to adapt to the diverse needs of users. It usually displays fixed content without considering the real-time context of the user, resulting in a low match between the displayed content and user needs, affecting the user's browsing experience, dwell time, and purchase conversion rate. In order to improve the user experience and interaction rate, the platform introduced a context-adaptive display system based on display solution data model training to meet the high demand of users for personalized display.

[0032] In this embodiment, the system first collects the user's contextual data in real time through the Internet of Things device, including the temperature, light, device type and environmental information of the user's location. The contextual features of each dimension are weighted and fused through a multi-head attention mechanism, and this information is integrated into a single contextual vector. Suppose a user opens an e-commerce platform application through a mobile device on a weekday afternoon. The system records the user's device type, location (office area), light conditions (brighter), and recent browsing history. The system generates and updates the user's contextual preference portrait, which shows that the user prefers to browse smart home and office supplies during office hours, and prefers a bright and simple display style.

[0033] When generating a display plan, the system uses the conditional generative adversarial network as the input condition to generate a preliminary display plan that matches the user's current situation. The generator generates display content that includes product recommendations, display layout, and color style, and evaluates the fitness of the generated content through the discriminator to ensure that the display content meets the user's real-time needs. Assume that the generated preliminary display plan includes a set of product recommendations for office supplies and smart homes, with a simple and bright layout, mainly displayed on a white background and light text. During the generation process, the system monitors the user's interactive behavior data in real time, including click-through rate, dwell time, and sliding frequency. Through the adaptive neural control module, the generator weights are continuously adjusted according to real-time feedback to optimize the display content generation effect and ensure that the system can quickly respond to changes in user needs.

[0034] When users swipe on the display page and stay on it for a long time, the system will adjust the display layout to highlight specific products, while appropriately optimizing the display order of the content. In the hierarchical situational control, the system first identifies the user's current demand theme as "daily necessities in office scenes" from a macro perspective, and further refines it to the meso-level display style adaptation (such as bright color style during the day); at the micro level, it adjusts details such as button position and product image size in real time to ensure that it meets the user's visual preferences and interaction needs. For example, when the user's click-through rate increases significantly when selecting an office desk, the system will automatically move the recommended content of this category to the front of the page to enhance the recommendation effect.

[0035] To verify the beneficial effects of the system of the present invention, the e-commerce platform conducted system comparison experiments in different scenarios. The comparison experiments were conducted in different locations and situations to evaluate the advantages of the situational adaptive display system in personalized recommendations. In this experiment, the system collected situational features at different locations and time periods of users. The collected data included temperature, light, device type, and user stay time, etc., to analyze the system's performance in terms of real-time adaptability, matching degree, and click-through conversion rate.

[0036] The system was experimented in offices, cafes, and homes at different times in the morning and afternoon, and recorded the user interaction data when using traditional static display and context-adaptive display systems. The following is part of the collected data and analysis results.

[0037] Table 1 Comparison of user behavior data in different usage scenarios ; The experimental results show that with the support of the context-adaptive display system, the user's stay time, click-through rate and conversion rate in various situations have been significantly improved. In the morning and afternoon office scenes, the average stay time of users using the context-adaptive system increased by about 30% compared with the traditional system, the click-through rate increased by more than 40%, and the conversion rate increased significantly; in cafes and home scenes, whether in bright or dark environments, the context-adaptive system can adjust the display content according to the current situational characteristics, thereby improving user experience and interactive performance.

[0038] For example, in the dim light conditions of home scenes, the system will give priority to displaying dark-themed display interfaces to reduce the interference of bright colors. In the evening cafe scene, the system automatically optimizes the color tone and layout of the displayed content to match the user's current viewing comfort. These adjustments make the user's browsing experience in the context-adaptive system smoother, further improving the matching degree between the displayed content and user needs.

[0039] Through comparative experiments, it is verified that the context-adaptive display system provided by the present invention has significant advantages in personalized display. Its context feature collection and feedback mechanism can quickly respond to user needs in diverse contexts, achieve higher user stay time and click-through conversion rate, and prove that the present invention can effectively solve the problem of insufficient personalized recommendation in traditional systems.

[0040] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for training a display scheme data model, characterized in that: The steps include: S1. Collect users’ micro-context data through IoT devices, extract macro-context data from user device information, and construct a multi-source context data set; S2. Preprocess the multi-source context dataset, weight the context features from different sources through a multi-head attention mechanism, and integrate the multi-source context features into a single context vector; S3, extract preference features from the user's historical interaction data, generate a user preference portrait, and fuse the context vector with the user preference portrait to generate a dynamic context preference portrait; S4, inputting the generated dynamic situation preference portrait into the conditional generative adversarial network as a condition to generate a display plan that meets the current situation; S5. Embed an adaptive neural control module, use the context feature feedback loop to dynamically adjust the generation strategy of the conditional generative adversarial network, adjust the weight of the generation model in real time according to the user's click rate, dwell time and sliding behavior, and optimize the display solution generation effect; S6. Introducing a hierarchical situation control mechanism, wherein the hierarchical situation control mechanism includes macro situation control, meso situation control and micro situation control; S7, collecting feedback data of user interactions, inputting the feedback data into the dynamic situational preference profile and adaptive neural control module, adjusting the weight parameters of the conditional generative adversarial network, and iteratively updating the user preference profile based on historical interaction feedback; S8. Analyze the similarity between the current situation and the historical situation through situation similarity detection, apply the meta-learning algorithm to determine the applicability of the display plan, and use transfer learning to transfer the effective display plan in the historical situation to the current situation.

2. A method for training a display scheme data model according to claim 1, characterized in that: The S4 specifically includes: S41. Obtaining dynamic situational preference portraits ,in Table ith feature dimension value in the dynamic situation preference portrait; S42. The dynamic situation preference portrait V is input into the generator of the conditional generative adversarial network. The generator generates a preliminary display content vector C by mapping V into a hidden layer vector Z and combining it with the generated noise vector N: ; Among them, tanh represents the hyperbolic tangent activation function, represents the adaptive adjustment parameter, The weight matrix representing the dynamic situational preference profile, represents the weight matrix of the hidden layer vector, represents the weight matrix of the noise vector, b represents the bias term, n represents the number of feature dimensions of the dynamic situational preference portrait, m represents the number of dimensions of the hidden layer vector, and p represents the number of dimensions of the noise vector. represents the smoothing term, , and Represent the weight coefficients in the weight matrix of dynamic situation preference portrait, hidden layer vector and noise vector respectively, Represents the jth feature dimension value in the hidden layer vector, Represents the kth feature dimension value in the noise vector; S43, outputting a preliminary display scheme according to the processing of the preliminary display content vector C, wherein the preliminary display scheme includes layout elements, color matching elements and content elements, wherein the layout elements include button positions and image display area size, the color matching elements include background color and text color, and the content elements include type, style and text content of the display content; S44. The output of the generator is evaluated using a discriminator. The input conditions are the dynamic situation preference portrait V and the preliminary display content vector C. The discriminator determines the generation scheme according to the matching degree of the display content and outputs the matching degree score: ; Among them, D(V,C) represents the matching score, exp represents the exponential function, and U represents the weight matrix in the discriminator. represents the context vector weight in the discriminator, represents the bias term, f(V,C) represents the combined feature vector of the dynamic situation preference portrait V and the preliminary display content vector C, and f(V) represents the feature vector of the dynamic situation preference portrait V. Represents each weight item in the context vector weight, Represents the weight items in the preliminary display content vector C, represents each feature value in the preliminary display content vector C, a represents the number of feature dimensions of the context vector, and b represents the number of feature dimensions of the preliminary display content vector C; S45. If the matching score D(V,C) does not reach the set matching threshold, the weight matrix and bias value of the generator are optimized by directional propagation to generate display content that is more in line with the current situation; S46. Through multiple iterations, in the alternating training process of the generator and the discriminator, the display scheme generated by the generator gradually conforms to the dynamic situation preference portrait V, and finally outputs a display scheme that conforms to the current situation.

3. The method for training a display scheme data model according to claim 1, characterized in that: The S5 specifically includes: S51. Collect real-time user interaction data, including click-through rate , Dwell time and sliding behavior , the click rate Indicates the user's interest in the displayed content, the dwell time Reflects the user's attention, the sliding behavior Indicates the user's tendency to explore displayed content and interaction area preference; S52. Click-through rate , Dwell time and sliding behavior Standardize and convert to a unified feedback score: ; Among them, F represents the unified feedback score, Represents the weight coefficient of click rate, represents the weight coefficient of the residence time, Represents the weight coefficient of sliding behavior, represents the smoothing coefficient, represents the weight parameter of the k-th layer weight in the conditional generative adversarial network, represents the feature value of the kth layer in the conditional generative adversarial network, and N represents the number of layers of the generator in the conditional generative adversarial network; S53, inputting the unified feedback score F into the adaptive neural control module, and using the unified feedback score F to adjust the weight matrix of the conditional generative adversarial network generator; S54, using a gradient descent method to update in real time the generator weight matrix, the hidden layer vector weight matrix, the noise vector weight matrix and the bias value in the conditional generative adversarial network; S55, introduce the control parameters of the adaptive feedback loop , and ,in Indicates the response sensitivity of the click rate. represents the response sensitivity of the residence time, Indicates the response sensitivity of the sliding behavior; S56. After updating and iterating the weights of the feedback loop, the display solution generation effect is optimized. When the click rate, dwell time and sliding behavior reach a predetermined threshold, it is determined that the generated display solution has met the current situation requirements.

4. The method for training a display scheme data model according to claim 1, characterized in that: The S6 specifically includes: S61, identify the user's current main demand context, determine the theme and overall direction of the display content based on user behavior data and preference portraits, and analyze the user's contextual needs through the context rights module. Re-adjust the theme of the displayed content to match the user's needs and situations; S62: Adapt the content style based on the user device information and environmental factors. The meso-scenario control module applies the corresponding style template based on different device and environmental characteristics, and uses the meso-scenario style weight coefficient Controlling style transfer; S63. When the user interacts with the displayed content, the user's click, slide and stay behaviors are detected in real time, and the details of the displayed content are adjusted according to the micro-interaction behaviors. The micro-context control module dynamically updates the displayed content according to the user's real-time feedback, including adjusting the button color, content typesetting and layout, and controlling the weight coefficient through the micro-context Control the adjustment range of display details; S64. Based on the hierarchical context control mechanism, a multi-level fitness score of the displayed content is defined to measure the adaptability of the displayed content at different context levels: ； Among them, A represents the multi-level fitness score of the displayed content, represents the macro-scenario control weight coefficient, represents the macro-scenario fitness, represents the micro-situational fitness, represents the weight parameter of the kth feature in macro context control, represents the kth feature in the macro-context control, H represents the number of features in the macro-context control, represents the weight parameter of the jth feature in the meso-situational control, represents the jth feature in the meso-situational control, J represents the number of features in the meso-situational control, represents the weight parameter of the i-th feature in micro-context control, represents the i-th feature in micro-contextual control, Q represents the number of features in micro-contextual control; S65, when the fitness score A reaches the preset threshold, it is considered that the displayed content has achieved the expected context matching effect at the macro, meso and micro context levels, and it is determined that the displayed content meets the multi-level context needs of the user; S66. If the fitness score A does not reach the preset threshold, the weight coefficients of the macro context control module, the meso context control module and the micro context control module are updated through the back propagation mechanism to gradually optimize the adaptability of the displayed content.

5. The method for training a display scheme data model according to claim 1, characterized in that: The S7 specifically includes: S71. Collect comprehensive interactive data of users in displayed content, including click-through rate, browsing time, sliding frequency and hot spots, to form multi-dimensional interactive data; S72, converting the collected multi-dimensional interaction data into a context feedback vector, and integrating it with the user's historical portrait data after data cleaning and preprocessing, and generating a personalized vector with time and context specificity by analyzing the user's preference fluctuations in different contexts; S73, the dynamic situation preference portrait module generates a multi-stage user portrait update strategy based on the input of the situation feedback vector, wherein the update strategy separates short-term preferences from long-term preferences based on the user's interaction frequency, time interval, and changes in situational needs, and constructs a preference feature vector that meets the user's recent interests; S74, the adaptive neural control module receives the updated situation preference portrait information and automatically detects the deviation area of ​​the user's interest characteristics; S75. Through the multi-level feedback control mechanism of the adaptive neural control module, preference adjustment weights are further introduced to enable the generator to automatically adjust the display scheme of the displayed content. The preference adjustment weights are based on the user context feedback frequency and context change trend to dynamically adjust the style, color matching and structural display characteristics of the generated content. S76. In multiple rounds of iterative optimization, by recording the generated effect after each user interaction, gradually build a historical preference map in the user portrait, accumulate the user's preference change trajectory, identify the long-term pattern of user preference based on the historical preference map generator, and give priority to the long-term pattern of user preference when generating content.

6. The method for training a display scheme data model according to claim 1, characterized in that: The S8 specifically includes: S81. Collect current user context data, including the user's real-time behavior characteristics, device information, and environmental parameters, to form a current context feature vector ; S82. Calculate the current situation feature vector based on the situation similarity detection model and the historical context feature vector set The similarity between them is calculated based on the cosine similarity or Euclidean distance measurement method to select the feature vector that is most similar to the current situation; S83. Apply the meta-learning algorithm to determine the adaptability of the display scheme in the historical context to the current context: ; in, represents the applicability score, H represents the number of historical context features, represents the weight of the i-th feature, represents the weight of the jth feature, represents the i-th eigenvalue of the current situation, represents the ith eigenvalue of the historical context, exp represents the exponential function, represents the exponential decay coefficient, represents the smoothing coefficient, represents the nonlinear adjustment coefficient, represents the kth eigenvalue of the current situation, represents the kth eigenvalue of the historical context; S84. For the display scheme determined to be applicable, the transfer learning module transfers the generation model parameters in the historical context to the current context, and the transfer process includes the selection of key parameters, the conversion of feature mapping and the optimization of content features; S85. For the migrated display scheme, the generated model is fine-tuned. The fine-tuning process is based on the differences between the current context and the historical context, and the display strategy, color matching style and interaction details in the model are adjusted; S86. After completing the migration and fine-tuning, the generation effect and user feedback data in the current situation are recorded and saved as new historical situation features.

7. A display scheme data model training system, which executes the display scheme data model training method according to any one of claims 1 to 6, characterized in that: Includes the following modules: The data collection module is used to collect the user's micro-context data through the IoT device and extract the macro-context data from the user's device information to generate a context data set containing multi-source context features; The data preprocessing module is used to clean, standardize and denoise the context data set, and to perform weighted fusion of context features from different sources through a multi-head attention mechanism to generate a single context vector; The user portrait generation module is used to extract preference features from the user's historical interaction data to generate a user preference portrait, and fuse the context vector with the user preference portrait to generate a dynamic context preference portrait; The conditional generative adversarial network module includes a generator and a discriminator, which is used to receive the dynamic situation preference portrait as a conditional input and generate a display plan that conforms to the current situation; Adaptive neural control module, which is used to dynamically adjust the generation strategy of the conditional generative adversarial network based on real-time feedback from users’ click rate, dwell time, and sliding behavior, and optimize the generation effect of the display solution through feedback loop; Hierarchical context control module, including macro context control module, meso context control module and micro context control module, used to adjust the display scheme layer by layer according to user needs, device information and real-time interactive behavior; Feedback data processing module, used to collect and process user interaction feedback data, input the feedback data into the dynamic situation preference profile and adaptive neural control module, and iteratively update the user preference profile based on historical interaction feedback; The context similarity detection module is used to analyze the similarity between the current context and the historical context based on the context similarity detection model, determine the applicability of the historical display scheme in combination with the meta-learning algorithm, and transfer the highly applicable historical display scheme to the current context through the transfer learning module; The storage module is used to record and save the generation effect and user feedback data in the current situation after completing the migration and fine-tuning of the display plan, and update the situation feature database.

Images