Design style intelligent recommendation and generation method and system based on user preference

Abstract

The invention provides a design style intelligent recommendation and generation method and system based on user preferences, and relates to the technical field of design styles, and the method comprises the steps: constructing a user preference feature vector; analyzing a user preference evolution rule, and generating a user preference evolution curve; calculating the time sequence correlation degree between the icon design features and the user preference features, and constructing a resonance scoring matrix; designing an adaptive resonance threshold screening model based on the resonance scoring matrix; creating an initial icon design scheme by using a deep generative adversarial network; and optimizing the design parameters through transfer learning and deep reinforcement learning, and outputting a personalized icon design recommendation scheme. According to the method, intelligent recommendation and generation of icon design can be realized, and the design efficiency and the user satisfaction are improved.

Description

Technical Field

[0001] The present invention relates to the technology of design styles, and particularly to an intelligent recommendation and generation method and system for design styles based on user preferences. Background Art

[0002] Icon design is an important part of an enterprise's brand image and is of great significance for enhancing brand recognition and market competitiveness. With the rapid development of Internet and artificial intelligence technologies, intelligent icon design methods based on user preferences have gradually become a research hotspot. Currently, in the field of icon design, artificial intelligence technologies such as machine learning and deep learning have been applied to recommend personalized icon design solutions for users by analyzing user behavior data and design preferences.

[0003] However, there are still some deficiencies in the existing intelligent recommendation technologies for icon design: First, the capture and analysis of user preferences are not comprehensive and in-depth enough, making it difficult to accurately grasp the potential needs and aesthetic tendencies of users. Second, the dynamic adaptability of the recommendation algorithm is insufficient, and it cannot timely track the changing trends of user preferences, resulting in a deviation between the recommended results and the actual needs of users. Finally, the generation process of design solutions lacks full utilization of the collective wisdom of the user group, making it difficult to produce innovative and highly recognizable design results.

[0004] To address these problems, there is an urgent need to develop an intelligent recommendation system for icon design that can comprehensively analyze explicit and implicit user data, dynamically track preference evolution, integrate collective intelligence, and continuously optimize to meet the needs of users for personalized and high-quality icon design. Summary of the Invention

[0005] Embodiments of the present invention provide an intelligent recommendation and generation method and system for design styles based on user preferences, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiments of the present invention,

[0007] An intelligent recommendation and generation method for design styles based on user preferences is provided, including:

[0008] Collecting design preference selections, rating feedback, annotation information, and interactive review texts in explicit user interaction data, and simultaneously obtaining browsing trajectories, stay durations, and operation sequences in implicit user behavior data. Based on the explicit interaction data and implicit behavior data, a user preference feature vector is constructed; the user preference feature vector is input into a temporal deep learning network, and the evolution law of user preferences is analyzed through a periodic pattern recognition algorithm to generate a user preference evolution curve, and the importance weights of features in each dimension are calculated based on an attention mechanism, and a feature importance score is output;

[0009] According to the user preference evolution curve and the feature importance score, use a multi-head attention network to calculate the temporal correlation between the icon design features and the user preference features, and construct a resonance score matrix; based on the resonance score matrix, design an adaptive resonance threshold screening model, which tracks the dynamic changes of the resonance coefficient through a bidirectional long short-term memory network, and combines a swarm intelligence algorithm to optimize the resonance score at the group level, generating group resonance enhanced features, which are used to guide the dynamic adjustment of subsequent icon design schemes;

[0010] Input the resonance score matrix and the group resonance enhanced features into a deep generative adversarial network, and create an initial icon design scheme through a conditional multi-branch generator; conduct multi-scale authenticity evaluation and resonance degree calculation on the initial icon design scheme, and screen out high-resonance coefficient design features; map the high-resonance coefficient design features to the knowledge graph enhanced design space through a transfer learning module, and continuously optimize the design parameters in combination with deep reinforcement learning, and continuously adjust and output the final personalized icon design recommendation scheme according to the dynamic feedback of the user preference evolution curve.

[0011] Input the user preference feature vector into a temporal deep learning network, analyze the user preference evolution law through a periodic pattern recognition algorithm, generate a user preference evolution curve, and calculate the importance weight of each dimension feature based on the attention mechanism, and output the feature importance score including:

[0012] Input the user preference feature vector into a temporal deep learning network, which includes a feature extraction layer and a temporal analysis layer. The feature extraction layer uses a bidirectional gated recurrent unit to extract the temporal features of the user preference feature vector, and the temporal analysis layer uses a long short-term memory network to perform long-term dependence analysis on the temporal features to obtain a user preference temporal feature sequence;

[0013] Conduct periodic pattern analysis on the user preference temporal feature sequence, extract periodic features through Fourier transform, combine the sliding average method of an adaptive window to identify local periodic patterns, and use a dynamic programming algorithm to optimize the combination of the local periodic patterns to obtain the user preference evolution law;

[0014] Based on the user preference evolution law, use a piecewise polynomial regression model to fit the user preference change trend, determine the key change nodes through an adaptive breakpoint selection algorithm, and generate a user preference evolution curve in combination with spline smoothing technology;

[0015] Construct a multi-layer attention network to calculate the importance of each dimension feature in the user preference evolution curve. The multi-layer attention network captures the mutual relationship between feature dimensions through a self-attention mechanism, and analyzes the evolution characteristics of features over time using a cross-attention mechanism;

[0016] Based on the output results of the multi-layer attention network, the analytic hierarchy process is used to calculate the initial weights of the feature dimensions in the user preference curve, and the initial weights are dynamically adjusted by combining time series correlation analysis, and the feature importance scores are obtained through normalization processing.

[0017] Based on the resonance scoring matrix, an adaptive resonance threshold screening model is designed. The adaptive resonance threshold screening model tracks the dynamic changes of the resonance coefficient through a bidirectional long short-term memory network, and combines a swarm intelligence algorithm to optimize the resonance scores at the group level, generating group resonance enhanced features. The group resonance enhanced features are used to guide the dynamic adjustment of the subsequent icon design scheme, including:

[0018] An adaptive resonance threshold screening model is constructed based on the resonance scoring matrix. The adaptive resonance threshold screening model uses a sliding window to segment the resonance scoring matrix, extracts the scoring distribution features through kernel density estimation, dynamically calculates the resonance threshold by combining the distribution kurtosis and skewness, and obtains the threshold screening result;

[0019] The threshold screening result is input into a bidirectional long short-term memory network. The bidirectional long short-term memory network includes a gating adjustment layer and a feature fusion layer. The gating adjustment layer uses a gating mechanism to adjust the information retention ratio, and the feature fusion layer introduces residual connections and layer normalization processing, and fuses bidirectional features through an attention mechanism to track the dynamic changes of the resonance coefficient;

[0020] Based on the dynamic changes of the resonance coefficient, a swarm intelligence optimization algorithm is constructed. The swarm intelligence optimization algorithm takes the scoring weight and threshold parameter as the optimization objectives, enhances the search ability by using local optimal solution items, dynamically adjusts the inertia weight by combining a non-linear decreasing strategy, and optimizes the resonance scores at the group level to obtain the group level optimization result;

[0021] The group level optimization result is input into a feature fusion network. The feature fusion network adjusts the historical features through a time decay factor, analyzes the feature correlation by using a spatial attention mechanism, and integrates the group opinions based on a confidence weighted method to generate group resonance enhanced features. The group resonance enhanced features are used to guide the dynamic adjustment of the icon design scheme.

[0022] Based on the dynamic changes of the resonance coefficient, a swarm intelligence optimization algorithm is constructed. The swarm intelligence optimization algorithm takes the scoring weight and threshold parameter as the optimization objectives, enhances the search ability by using local optimal solution items, dynamically adjusts the inertia weight by combining a non-linear decreasing strategy, and optimizes the resonance scores at the group level to obtain the group level optimization result, including:

[0023] Construct a swarm intelligence optimization algorithm based on the dynamic change of the resonance coefficient. The swarm intelligence optimization algorithm takes the scoring weight and the threshold parameter as the optimization objectives. The scoring weight is used to quantify the importance of each dimension feature in the resonance score, and the threshold parameter is used to control the screening criteria of the resonance score;

[0024] Initialize the population of the swarm intelligence optimization algorithm, calculate the fitness values of each individual in the initial population, and screen high-quality solutions according to the fitness values to construct a local optimal solution set, which is used to enhance the local search ability of the algorithm;

[0025] In the iterative optimization process, according to the dynamic change characteristics of the resonance coefficient, select high-quality solutions from the local optimal solution set for local search, dynamically adjust the range and accuracy of the local search, and enhance the search ability of the algorithm through the local optimal solutions;

[0026] Adopt a non-linear decreasing strategy to dynamically adjust the inertia weight, determine the weight decreasing coefficient according to the dynamic change degree of the resonance coefficient, and adaptively update the inertia weight by using the weight decreasing coefficient. The inertia weight is used to regulate the swarm search behavior;

[0027] Iteratively optimize the population based on the updated inertia weight, calculate the optimized swarm fitness value, update the local optimal solution set according to the swarm fitness value, and continue to iterate until convergence, and output the resonance score result optimized at the swarm level.

[0028] Input the resonance score matrix and the swarm resonance enhancement feature into a deep generative adversarial network, and create an initial icon design scheme through a conditional multi-branch generator; conduct multi-scale authenticity evaluation and resonance degree calculation on the initial icon design scheme, and the screened high resonance coefficient design features include:

[0029] Input the resonance score matrix and the swarm resonance enhancement feature into a feature fusion network. The feature fusion network uses a multi-head attention mechanism to calculate the feature weights, performs principal component analysis and dimensionality reduction processing on the resonance score matrix, and adaptively fuses the dimensionality-reduced features with the swarm resonance enhancement feature to generate a conditional vector;

[0030] Construct a deep generative adversarial network based on the conditional vector. The deep generative adversarial network includes a multi-branch generator and a discriminator. The multi-branch generator includes a shape generation branch, a color generation branch, and a composition generation branch. Each branch introduces a conditional normalization layer to inject the conditional vector into the generation process, and uses an adaptive weighting method to fuse the output features of each branch to generate an initial icon design scheme;

[0031] Input the initial icon design scheme into the discriminator for multi-scale authenticity evaluation. The discriminator includes a global discrimination branch and a local discrimination branch. The global discrimination branch evaluates the overall coordination of the icon, and the local discrimination branch evaluates the detailed features of the icon to generate an authenticity evaluation score.

[0032] Calculate the resonance coefficient of the initial icon design scheme based on the resonance score matrix. Use a sliding window to analyze the dynamic change of the resonance coefficient, smooth the resonance score result by an exponential weighting method, and comprehensively evaluate the authenticity evaluation score and the resonance score result. Screen out the design features with high resonance coefficients from the results of the comprehensive evaluation.

[0033] Map the design features with high resonance coefficients to the design space enhanced by the knowledge graph through a transfer learning module, and continuously optimize the design parameters in combination with deep reinforcement learning. According to the dynamic feedback of the user preference evolution curve, continuously adjust and output the final personalized icon design recommendation scheme, including:

[0034] Input the design features with high resonance coefficients into the transfer learning module. The transfer learning module includes a source domain encoder and a target domain mapper. The source domain encoder reduces the dimension and extracts features from the design features with high resonance coefficients to obtain a design feature vector, and the target domain mapper maps the design feature vector to the design space enhanced by the knowledge graph to obtain a mapped feature.

[0035] Construct a deep reinforcement learning model based on the mapped feature. The deep reinforcement learning model includes a parameter generation network and an evaluation network. The parameter generation network fuses the mapped feature with the current design parameters to generate a parameter adjustment scheme, and the evaluation network evaluates the feasibility of the parameter adjustment scheme to obtain an evaluation score.

[0036] Use the user preference evolution curve as a dynamic feedback signal to adjust the reward function weight of the deep reinforcement learning model. The parameter generation network continuously optimizes the design parameters based on the adjusted reward function to obtain optimized parameters, and the evaluation network evaluates the optimized parameters and feeds back the evaluation results to the parameter generation network.

[0037] Input the optimized parameters into the icon generation model to obtain a set of candidate icon design schemes. Calculate the preference matching degree of each scheme in the set of candidate icon design schemes based on the user preference evolution curve, and select the scheme with the highest preference matching degree as the final personalized icon design recommendation scheme.

[0038] Taking the user preference evolution curve as a dynamic feedback signal, adjust the weight of the reward function of the deep reinforcement learning model. The parameter generation network continuously optimizes the design parameters based on the adjusted reward function to obtain optimized parameters. The evaluation network evaluates the optimized parameters and feeds back the evaluation results to the parameter generation network, including:

[0039] Input the user preference evolution curve into the feature extraction module to extract the preference intensity feature and the preference direction feature. The preference intensity feature represents the degree of the user's preference for the design parameters, and the preference direction feature represents the changing trend of the user's preference. Construct the preference intensity feature and the preference direction feature into a dynamic feedback signal;

[0040] Based on the dynamic feedback signal, adjust the weight of the reward function of the deep reinforcement learning model. Adjust the preference matching degree weight in the reward function weight according to the preference intensity feature, and adjust the parameter constraint weight in the reward function weight according to the preference direction feature to generate an adjusted reward function;

[0041] The parameter generation network continuously optimizes the design parameters based on the adjusted reward function. Input the current design parameters and the dynamic feedback signal into the parameter generation network to obtain a parameter adjustment scheme, and evaluate the parameter adjustment scheme based on the adjusted reward function to obtain optimized parameters;

[0042] The evaluation network evaluates the optimized parameters, generates an evaluation result including a parameter feasibility score and a preference compliance score, and feeds back the evaluation result to the parameter generation network.

[0043] In the second aspect of the embodiments of the present invention,

[0044] Provide an intelligent recommendation and generation system for design styles based on user preferences, including:

[0045] The first unit is used to collect the design preference selections, rating feedbacks, annotation information, and interactive review texts in the user's explicit interaction data, and at the same time obtain the browsing trajectories, residence durations, and operation sequences in the user's implicit behavior data. Based on the explicit interaction data and the implicit behavior data, construct a user preference feature vector; input the user preference feature vector into a temporal deep learning network, analyze the user preference evolution law through a periodic pattern recognition algorithm, generate a user preference evolution curve, and calculate the importance weights of the features in each dimension based on the attention mechanism to output a feature importance score;

[0046] A second unit is configured to calculate the temporal correlation between icon design features and user preference features by using a multi-head attention network according to the user preference evolution curve and the feature importance score, and construct a resonance score matrix; based on the resonance score matrix, an adaptive resonance threshold screening model is designed, and the adaptive resonance threshold screening model tracks the dynamic change of the resonance coefficient through a bidirectional long short-term memory network, and combines a swarm intelligence algorithm to optimize the resonance score at the group level, generating group resonance enhanced features, and the group resonance enhanced features are used to guide the dynamic adjustment of subsequent icon design schemes;

[0047] A third unit is configured to input the resonance score matrix and the group resonance enhanced features into a deep generative adversarial network, and create an initial icon design scheme through a conditional multi-branch generator; perform multi-scale authenticity evaluation and resonance degree calculation on the initial icon design scheme, and screen out design features with high resonance coefficients; map the design features with high resonance coefficients to a knowledge graph-enhanced design space through a transfer learning module, and continuously optimize the design parameters by combining deep reinforcement learning, and continuously adjust and output a final personalized icon design recommendation scheme according to the dynamic feedback of the user preference evolution curve.

[0048] In a third aspect of the embodiments of the present invention,

[0049] there is provided an electronic device, including:

[0050] a processor;

[0051] a memory for storing processor-executable instructions;

[0052] wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0053] In a fourth aspect of the embodiments of the present invention,

[0054] there is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0055] The beneficial effects of the present application are as follows:

[0056] 1. By collecting explicit interaction data and implicit behavior data of users, constructing user preference feature vectors, and using a temporal deep learning network to analyze the user preference evolution law and generate a user preference evolution curve, the present method realizes the accurate capture and dynamic tracking of user design preferences, laying a solid foundation for subsequent icon design recommendations.

[0057] 2. This method introduces an adaptive resonance threshold screening model and a swarm intelligence algorithm. By calculating the temporal correlation degree between the icon design features and the user preference features, a resonance scoring matrix is constructed and optimized at the group level to generate group resonance enhanced features, effectively improving the matching degree between the icon design scheme and the user preferences and enhancing the accuracy and personalization of the recommendation results.

[0058] 3. This method uses a deep generative adversarial network combined with a multi-branch generator with conditional constraints to create an initial icon design scheme, and continuously optimizes the design parameters through transfer learning and deep reinforcement learning, realizing intelligent icon design generation based on user preferences. It not only ensures the innovation and diversity of the design scheme but also can continuously adjust the output results according to the dynamic changes of user preferences, greatly improving the efficiency and quality of icon design. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is a schematic flowchart of the intelligent recommendation and generation method for design styles based on user preferences according to an embodiment of the present invention;

[0060] Figure 2 is a schematic structural diagram of the intelligent recommendation and generation system for design styles based on user preferences according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0062] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0063] Figure 1 is a schematic flowchart of the intelligent recommendation and generation method for design styles based on user preferences according to an embodiment of the present invention, as Figure 1 shown, the method includes:

[0064] Collect design preference selections, rating feedback, annotation information, and interactive review texts in the explicit interaction data of users. At the same time, obtain browsing trajectories, stay durations, and operation sequences in the implicit behavior data of users. Based on the explicit interaction data and implicit behavior data, construct a user preference feature vector; input the user preference feature vector into a temporal deep learning network, analyze the evolution law of user preferences through a periodic pattern recognition algorithm, generate a user preference evolution curve, and calculate the importance weights of features in each dimension based on the attention mechanism, and output feature importance scores;

[0065] According to the user preference evolution curve and feature importance scores, use a multi-head attention network to calculate the temporal correlation between icon design features and user preference features, and construct a resonance score matrix; based on the resonance score matrix, design an adaptive resonance threshold screening model. The adaptive resonance threshold screening model tracks the dynamic changes of resonance coefficients through a bidirectional long short-term memory network, and combines a swarm intelligence algorithm to optimize the resonance scores at the group level, generating group resonance enhanced features, which are used to guide the dynamic adjustment of subsequent icon design schemes;

[0066] Input the resonance score matrix and group resonance enhanced features into a deep generative adversarial network, and create an initial icon design scheme through a conditional multi-branch generator; perform multi-scale authenticity evaluation and resonance degree calculation on the initial icon design scheme, and screen out design features with high resonance coefficients; map the design features with high resonance coefficients to a knowledge graph-enhanced design space through a transfer learning module, and continuously optimize the design parameters in combination with deep reinforcement learning. According to the dynamic feedback of the user preference evolution curve, continuously adjust and output the final personalized icon design recommendation scheme.

[0067] In an alternative embodiment, inputting the user preference feature vector into a temporal deep learning network, analyzing the evolution law of user preferences through a periodic pattern recognition algorithm, generating a user preference evolution curve, and calculating the importance weights of features in each dimension based on the attention mechanism, and outputting feature importance scores includes:

[0068] Input the user preference feature vector into a temporal deep learning network. The temporal deep learning network includes a feature extraction layer and a temporal analysis layer. The feature extraction layer uses a bidirectional gated recurrent unit to extract the temporal features of the user preference feature vector, and the temporal analysis layer uses a long short-term memory network to perform long-term dependence analysis on the temporal features to obtain a user preference temporal feature sequence;

[0069] Perform periodic pattern analysis on the user preference time series feature sequence, extract periodic features through Fourier transform, identify local periodic patterns by combining the moving average method with an adaptive window, and use the dynamic programming algorithm to optimize the combination of the local periodic patterns to obtain the evolution law of user preferences;

[0070] Based on the evolution law of user preferences, use a piecewise polynomial regression model to fit the changing trend of user preferences, determine key change nodes through an adaptive breakpoint selection algorithm, and generate a user preference evolution curve by combining spline smoothing technology;

[0071] Construct a multi-layer attention network to calculate the importance of each dimension feature in the user preference evolution curve. The multi-layer attention network captures the mutual relationship between feature dimensions through the self-attention mechanism and analyzes the evolution characteristics of features over time using the cross-attention mechanism;

[0072] Based on the output results of the multi-layer attention network, use the analytic hierarchy process to calculate the initial weights of each dimension feature in the user preference curve, dynamically adjust the initial weights by combining time series correlation analysis, and obtain the feature importance score through normalization processing.

[0073] The present invention proposes a method for analyzing user preference evolution based on time series deep learning and attention mechanism. This method first inputs the user preference feature vector into a time series deep learning network, and obtains the user preference time series feature sequence through feature extraction and time series analysis. Then, perform periodic pattern analysis on this sequence to identify the evolution law of user preferences. Next, generate a user preference evolution curve based on the evolution law, and finally use a multi-layer attention network to calculate the importance weights of each dimension feature and output the feature importance score.

[0074] First, input the user preference feature vector into a time series deep learning network. This network includes a feature extraction layer and a time series analysis layer. The feature extraction layer uses a bidirectional gated recurrent unit (BiGRU) to extract the time series features of the user preference feature vector. BiGRU can consider both past and future context information and effectively capture the time series dependence relationship of features. The time series analysis layer then uses a long short-term memory network (LSTM) to perform long-term dependence analysis on the extracted time series features. The gating mechanism of LSTM can selectively remember and forget information and is suitable for processing long sequence data. Through the processing of these two layers, the user preference time series feature sequence can be obtained.

[0075] For example, assume that the user preference feature vector contains 10 dimensions, where each dimension represents the user's preference degree for a certain type of commodity, and the value range is 0 - 1. When inputting the user preference feature vectors for 30 consecutive days, the BiGRU layer will extract the temporal features of each dimension, and the LSTM layer will further analyze the long-term change trends of these features, and finally output a 30 * 10 user preference temporal feature sequence matrix.

[0076] Next, perform a periodic pattern analysis on the user preference temporal feature sequence. First, extract the periodic features of the sequence through Fourier transform, convert the time-domain signal into a frequency-domain signal, and identify the main periodic components. Then, combine the moving average method with an adaptive window to identify local periodic patterns. Specifically, set an initial window size, slide this window on the sequence and calculate the average value of the data within the window, and identify local periodic patterns by comparing the changes in the average values of adjacent windows. The window size can be adaptively adjusted according to the degree of data change. Finally, use the dynamic programming algorithm to combine and optimize the identified local periodic patterns to obtain the globally optimal user preference evolution law.

[0077] For example, for the above-mentioned 30 * 10 feature sequence matrix, it may be identified that the user's preference for a certain type of commodity shows a periodic change every 7 days, while the preference for another type of commodity has a cycle every 15 days. Through the dynamic programming algorithm, these local periodic patterns can be combined into an overall user preference evolution law.

[0078] Based on the identified user preference evolution law, the next step is to use a piecewise polynomial regression model to fit the user preference change trend. First, determine the key change nodes through an adaptive piecewise point selection algorithm. This algorithm calculates the slope change between data points, and when the slope change exceeds a preset threshold, marks this point as a potential piecewise point. Then, evaluate and screen these potential piecewise points, and finally determine the optimal set of piecewise points. Next, apply the polynomial regression model to fit the data within each segment. The order of the polynomial can be adaptively selected according to the complexity of the data. Finally, combine the spline smoothing technique to smooth the connection at the piecewise points and generate a continuous and smooth user preference evolution curve.

[0079] For example, for a certain feature dimension, there may be obvious turning points in the change trend on the 10th day and the 20th day. Then the data can be divided into three segments and fitted with polynomials of different orders. Then, apply cubic spline interpolation at the connection between the 10th day and the 20th day to ensure the smoothness and continuity of the overall curve.

[0080] To calculate the importance of each dimension feature in the user preference evolution curve, a multi-layer attention network is constructed. This network first captures the mutual relationship between feature dimensions through the self-attention mechanism. The self-attention layer calculates the correlation between each feature and all other features, generating an attention weight matrix. Then, the cross-attention mechanism is used to analyze the evolution characteristics of features over time. The cross-attention layer takes time information as the query, and feature information as the key and value, calculating the importance of features at different time points. By stacking multiple layers, more complex feature relationships and time dependencies can be captured.

[0081] For example, for 10 feature dimensions, the self-attention layer will generate a 10*10 weight matrix representing the mutual influence between each dimension. The cross-attention layer will generate a 10-dimensional weight vector for each time point, representing the importance of each feature at that time point.

[0082] Finally, based on the output results of the multi-layer attention network, the analytic hierarchy process is used to calculate the initial weights of each dimension feature in the user preference curve. The analytic hierarchy process obtains the relative importance weights of each feature by constructing a judgment matrix, performing consistency tests and normalization processing. Then, the initial weights are dynamically adjusted in combination with time series correlation analysis. Specifically, the correlation coefficient between each feature and the overall preference change trend is calculated, and this correlation coefficient is used as an adjustment factor to correct the initial weights. Finally, through normalization processing, the final feature importance scores are obtained.

[0083] The solution of this application can:

[0084] By combining a time series deep learning network and a periodic pattern recognition algorithm, this method can effectively capture the long-term evolution law and short-term fluctuation characteristics of user preferences, improving the accuracy and comprehensiveness of user preference analysis. By adopting a multi-layer attention network and a dynamic weight adjustment mechanism, it realizes the precise quantification of the importance of user preference features, can adaptively identify key influencing factors, and provides strong support for personalized recommendation and precision marketing. This method has strong generality and scalability, can adapt to the user preference analysis needs of different fields and different scales, and provides an important decision-making basis for enterprises to formulate refined operation strategies and improve user experience.

[0085] In an alternative implementation, based on the resonance scoring matrix, an adaptive resonance threshold screening model is designed. The adaptive resonance threshold screening model tracks the dynamic changes of the resonance coefficient through a bidirectional long short-term memory network, and combines a swarm intelligence algorithm to optimize the resonance scores at the group level, generating group resonance enhanced features. The group resonance enhanced features are used to guide the dynamic adjustment of subsequent icon design schemes, including:

[0086] Construct an adaptive resonance threshold screening model based on the resonance scoring matrix. The adaptive resonance threshold screening model uses a sliding window to segment the resonance scoring matrix, extracts the scoring distribution features through kernel density estimation, dynamically calculates the resonance threshold by combining the distribution kurtosis and skewness, and obtains the threshold screening result;

[0087] Input the threshold screening result into a bidirectional long short-term memory network. The bidirectional long short-term memory network includes a gating adjustment layer and a feature fusion layer. Among them, the gating adjustment layer uses a gating mechanism to adjust the information retention ratio, and the feature fusion layer introduces residual connection and layer normalization processing, fuses bidirectional features through an attention mechanism, and tracks the dynamic changes of the resonance coefficient;

[0088] Based on the dynamic changes of the resonance coefficient, construct a swarm intelligence optimization algorithm. The swarm intelligence optimization algorithm takes the scoring weight and threshold parameter as the optimization objectives, uses the local optimal solution item to enhance the search ability, combines the non-linear decreasing strategy to dynamically adjust the inertia weight, and optimizes the resonance score at the swarm level to obtain the swarm-level optimization result;

[0089] Input the swarm-level optimization result into a feature fusion network. The feature fusion network adjusts the historical features through a time decay factor, analyzes the feature correlation using a spatial attention mechanism, integrates the swarm opinions based on a confidence-weighted method, and generates a swarm resonance enhanced feature. The swarm resonance enhanced feature is used to guide the dynamic adjustment of the icon design scheme.

[0090] The present invention proposes an adaptive resonance threshold screening model based on a resonance scoring matrix for guiding the dynamic adjustment of an icon design scheme. This model is implemented through multiple steps, specifically as follows:

[0091] First, construct an adaptive resonance threshold screening model based on the resonance scoring matrix. This model uses a sliding window to segment the resonance scoring matrix, and the window size can be set to 50 scoring units. For the scoring data within each window, extract the scoring distribution features through kernel density estimation. Specifically, use a Gaussian kernel function to smooth the scoring data to obtain a continuous probability density curve. Then calculate the kurtosis and skewness of this curve. The kurtosis reflects the sharpness of the distribution, and the skewness reflects the symmetry of the distribution. Dynamically calculate the resonance threshold based on the weighted combination of kurtosis and skewness, and the weights can be set to 0.6 and 0.4 respectively. Apply the calculated threshold to the original scoring data, and screen out the scores higher than the threshold as the threshold screening result.

[0092] Next, the threshold screening results are input into a bidirectional long short-term memory network to track the dynamic changes of the resonance coefficient. This network includes a gated adjustment layer and a feature fusion layer. The gated adjustment layer adopts three gating mechanisms: an input gate, a forget gate, and an output gate, which respectively control the reception of new input information, the forgetting of historical information, and the output information at the current moment. The gating parameters are mapped to the range of 0-1 through the sigmoid function to adjust the information retention ratio. The feature fusion layer introduces a residual connection, adding the input directly to the output to alleviate the problem of gradient disappearance. At the same time, layer normalization is used to normalize the input of each neuron to accelerate the network convergence. Finally, the forward and backward hidden states are fused through an attention mechanism to obtain a comprehensive feature representation. The attention weights can be calculated through the softmax function. Through this network, the changing trend of the resonance coefficient over time can be effectively captured.

[0093] Then, based on the dynamic changes of the resonance coefficient, a swarm intelligence optimization algorithm is constructed. This algorithm takes the scoring weight and threshold parameter as the optimization objectives and uses the particle swarm optimization method to solve them. Based on the standard particle swarm algorithm, a local optimal solution term is introduced to enhance the search ability. Specifically, at each iteration, with a certain probability (such as 0.3), the local optimal solution rather than the global optimal solution is selected to update the particle velocity. At the same time, a non-linear decreasing strategy is adopted to dynamically adjust the inertia weight, with the initial value set to 0.9 and gradually decreasing to 0.4 as the number of iterations increases, to avoid falling into local optima. Through multiple rounds of iterative optimization, the optimal scoring weight and threshold parameter at the group level are obtained.

[0094] Finally, the optimization results at the group level are input into the feature fusion network to generate group resonance enhanced features. This network first adjusts the influence of historical features through a time decay factor, and the decay coefficient can be set to 0.9. Then, a spatial attention mechanism is used to analyze the correlation between features, calculate the similarity matrix between features, and obtain the attention weights through the softmax function. Finally, weighted averaging is performed based on the confidence of each feature to integrate group opinions. The confidence can be comprehensively calculated from the stability and consistency of the features. The finally output group resonance enhanced features contain temporal information, spatial correlation, and group wisdom, which can effectively guide the dynamic adjustment of the icon design scheme.

[0095] In practical applications, 100 icon design schemes can be selected as the initial samples, and each scheme contains 10 scoring dimensions. After being processed by the above model, 20 high-quality schemes are screened out, and specific improvement suggestions are given. For example, for a certain icon scheme, the model finds that its scores are relatively low in the two dimensions of "recognizability" and "implication", but it performs excellently in the "aesthetic" dimension. Based on this, it can be suggested that the designer enhance the uniqueness and deep meaning of the icon while maintaining the aesthetic, so as to improve the overall design quality.

[0096] The solution of this application can:

[0097] Through the adaptive resonance threshold screening model, the dynamic screening of resonance scores is realized. This model can adaptively adjust the threshold according to the score distribution characteristics, effectively improving the accuracy and flexibility of screening. This dynamic screening mechanism can better adapt to the scoring characteristics in different scenarios, enhancing the generality and robustness of the model. The bidirectional long short-term memory network is used to track the dynamic changes of the resonance coefficient, effectively capturing the time series characteristics. Through the gating mechanism and attention mechanism, this network can store important information for a long time and focus on key time steps, thus accurately depicting the change trend of the resonance coefficient. This provides a reliable time series basis for subsequent optimization and decision-making. Combining the swarm intelligence algorithm and the feature fusion network, the optimization at the swarm level and feature enhancement are realized. By introducing local search and dynamic weight adjustment, the convergence performance of the optimization algorithm is improved. At the same time, the feature fusion network comprehensively considers time decay, spatial correlation, and group opinions, and the generated group resonance enhanced features have stronger representativeness and guiding significance. The integration of this group wisdom greatly improves the scientificity and effectiveness of the icon design scheme adjustment.

[0098] In an alternative embodiment, based on the dynamic changes of the resonance coefficient, a swarm intelligence optimization algorithm is constructed. The swarm intelligence optimization algorithm takes the scoring weight and threshold parameter as the optimization objectives, uses the local optimal solution item to enhance the search ability, combines the nonlinear decreasing strategy to dynamically adjust the inertia weight, and optimizes the resonance score at the swarm level. The swarm-level optimization results include:

[0099] Based on the dynamic changes of the resonance coefficient, a swarm intelligence optimization algorithm is constructed. The swarm intelligence optimization algorithm takes the scoring weight and threshold parameter as the optimization objectives. The scoring weight is used to quantify the importance of each dimension feature in the resonance score, and the threshold parameter is used to control the screening criteria of the resonance score;

[0100] Initialize the population of the swarm intelligence optimization algorithm, calculate the fitness values of each individual in the initial population, and screen high-quality solutions according to the fitness values to construct a local optimal solution set, which is used to enhance the local search ability of the algorithm;

[0101] In the iterative optimization process, according to the dynamic change characteristics of the resonance coefficient, select high-quality solutions from the local optimal solution set for local search, dynamically adjust the range and accuracy of the local search, and enhance the search ability of the algorithm through the local optimal solution;

[0102] Use the nonlinear decreasing strategy to dynamically adjust the inertia weight, determine the weight decreasing coefficient according to the dynamic change degree of the resonance coefficient, and adaptively update the inertia weight using the weight decreasing coefficient. The inertia weight is used to regulate the group search behavior;

[0103] Iteratively optimize the population based on the updated inertia weight, calculate the fitness value of the optimized population, update the local optimal solution set according to the population fitness value, and continue to iterate until convergence, and output the resonance score result optimized at the population level.

[0104] This embodiment proposes a swarm intelligence optimization algorithm based on the dynamic change of the resonance coefficient for optimizing the resonance score at the population level. This algorithm takes the scoring weight and threshold parameter as the optimization objectives, enhances the search ability by using the local optimal solution items, combines the non-linear decreasing strategy to dynamically adjust the inertia weight, and finally obtains the optimization result at the population level.

[0105] First, construct a swarm intelligence optimization algorithm. This algorithm takes the scoring weight and threshold parameter as the optimization objectives. The scoring weight is used to quantify the importance of each dimension feature in the resonance score. For example, the emotional dimension weight can be set to 0.3, the cognitive dimension weight to 0.4, and the behavioral dimension weight to 0.3. The threshold parameter is used to control the screening criteria of the resonance score. For example, the threshold can be set to 0.7, that is, samples with a resonance score greater than 0.7 are considered to have strong resonance.

[0106] Next, perform population initialization. Assume that the population size is 100, and each individual contains two parts: the scoring weight and the threshold parameter. 100 groups of initial solutions can be randomly generated. For example, the initial solution of the first individual may be [0.35, 0.35, 0.3, 0.65], indicating that the weights of the three dimensions are 0.35, 0.35, 0.3 respectively, and the threshold is 0.65. Then calculate the fitness value of each individual, and the fitness value can be measured based on the accuracy of the resonance score. Select the top 20% of the individuals with higher fitness values to construct the local optimal solution set to enhance the local search ability of the algorithm.

[0107] In the iterative optimization process, according to the dynamic change characteristics of the resonance coefficient, select high-quality solutions from the local optimal solution set for local search. For example, when it is detected that the resonance coefficient rises rapidly, the 3 solutions with the highest fitness can be selected for more refined local search; when the resonance coefficient changes gently, 5 - 10 solutions can be selected for relatively rough local search. The range and accuracy of the local search will also be dynamically adjusted according to the change of the resonance coefficient. For example, when the resonance coefficient fluctuates violently, the search range can be expanded to ±20% of the current solution, and the search accuracy is set to 0.01; when the resonance coefficient is relatively stable, the search range can be reduced to ±10%, and the search accuracy is increased to 0.001.

[0108] Meanwhile, a non-linear decreasing strategy is adopted to dynamically adjust the inertia weight. The weight decreasing coefficient is determined according to the dynamic change degree of the resonance coefficient. For example, the change rate of the resonance coefficient can be mapped to the range between 0.9 and 0.99 as the weight decreasing coefficient. When the resonance coefficient changes drastically, the weight decreasing coefficient approaches 0.9, causing the inertia weight to decrease rapidly; when the resonance coefficient changes gently, the weight decreasing coefficient approaches 0.99, causing the inertia weight to decrease slowly. Assume the initial inertia weight is 0.9, and the inertia weight after each iteration is equal to the current inertia weight multiplied by the weight decreasing coefficient.

[0109] Iteratively optimize the population based on the updated inertia weight. Use the updated inertia weight to adjust the search behavior of individuals. When the inertia weight is large, individuals tend to maintain the original search direction; when the inertia weight is small, individuals are more likely to change the search direction. Calculate the fitness value of the optimized population and update the set of local optimal solutions according to the new fitness value. For example, the 20 solutions with the highest fitness can be retained as the new set of local optimal solutions.

[0110] Repeat the above process until the algorithm converges or reaches the preset maximum number of iterations. Finally, output the resonance score results optimized at the population level, including the optimal scoring weights and threshold parameters. For example, the final optimization result may be [0.32, 0.38, 0.3, 0.72], indicating that the optimal weights for the three dimensions are 0.32, 0.38, 0.3, and the optimal threshold is 0.72.

[0111] The solution of this application can:

[0112] This method realizes the population-level optimization of resonance scoring by constructing a swarm intelligence optimization algorithm based on the dynamic change of the resonance coefficient, improving the accuracy and reliability of resonance scoring. Adopting the strategy of enhancing the search ability with local optimal solution items, combined with the dynamically adjusted local search range and precision, significantly improves the balance between the local search ability and the global search ability of the algorithm, effectively avoiding the problem of falling into local optimal solutions. Introducing a non-linear decreasing strategy to dynamically adjust the inertia weight and adaptively update the weight according to the dynamic change characteristics of the resonance coefficient enhances the adaptability of the algorithm to resonance characteristics in different stages, improving the optimization efficiency and convergence speed.

[0113] In an alternative embodiment, input the resonance scoring matrix and the population resonance enhancement features into a deep generative adversarial network, and create an initial icon design scheme through a conditional-constrained multi-branch generator; conduct multi-scale authenticity evaluation and resonance degree calculation on the initial icon design scheme, and the screened high-resonance coefficient design features include:

[0114] Input the resonance scoring matrix and the group resonance enhancement features into a feature fusion network. The feature fusion network uses a multi-head attention mechanism to calculate feature weights, performs principal component analysis dimensionality reduction on the resonance scoring matrix, and adaptively fuses the dimensionality-reduced features with the group resonance enhancement features to generate a conditional vector.

[0115] Construct a deep generative adversarial network based on the conditional vector. The deep generative adversarial network includes a multi-branch generator and a discriminator. The multi-branch generator contains a shape generation branch, a color generation branch, and a composition generation branch. Each branch introduces a conditional normalization layer to inject the conditional vector into the generation process, and adaptively weights the output features of each branch to generate an initial icon design scheme.

[0116] Input the initial icon design scheme into the discriminator for multi-scale authenticity evaluation. The discriminator includes a global discriminant branch and a local discriminant branch. The global discriminant branch evaluates the overall coordination of the icon, and the local discriminant branch evaluates the detailed features of the icon to generate an authenticity evaluation score.

[0117] Calculate the resonance coefficient of the initial icon design scheme based on the resonance scoring matrix, analyze the dynamic changes of the resonance coefficient using a sliding window, smooth the resonance scoring results through an exponential weighting method, comprehensively evaluate the authenticity evaluation score and the resonance scoring results, and screen out high-resonance coefficient design features from the comprehensive evaluation results.

[0118] The present invention provides an icon design method based on resonance scoring and a deep generative adversarial network. This method first inputs a resonance scoring matrix and group resonance enhancement features into a deep generative adversarial network, creates an initial icon design scheme through a multi-branch generator with conditional constraints, then performs multi-scale authenticity evaluation and resonance degree calculation on the initial design scheme, and finally screens out design features with high resonance coefficients.

[0119] First, input the resonance scoring matrix and group resonance enhancement features into a feature fusion network. The feature fusion network uses a multi-head attention mechanism to calculate feature weights, performs principal component analysis dimensionality reduction on the resonance scoring matrix, adaptively fuses the dimensionality-reduced features with the group resonance enhancement features, and generates a conditional vector. Specifically, an 8-head attention mechanism can be used, with each head having a dimension of 64, and the attention weights are obtained by calculating the dot product similarity between the query, key, and value. For the resonance scoring matrix, the principal components that explain 95% of the variance can be retained, usually reduced to 50 - 100 dimensions. Then, the dimensionality-reduced features are concatenated with the group resonance enhancement features and adaptively fused through a fully connected layer to obtain a 256-dimensional conditional vector.

[0120] Next, a deep generative adversarial network is constructed based on the conditional vector. This network includes a multi-branch generator and a discriminator. The multi-branch generator contains a shape generation branch, a color generation branch, and a composition generation branch. Each branch introduces a conditional normalization layer to inject the conditional vector into the generation process. Specifically, the shape generation branch can adopt a 5-layer convolutional network, followed by conditional batch normalization and ReLU activation function for each layer. The color generation branch uses a 3-layer fully connected network to output RGB color values. The composition generation branch adopts a transformer structure with an attention mechanism to generate relative position information of elements. Finally, the output features of each branch are fused by an adaptive weighting method to generate an initial icon design scheme with a resolution of 128x128.

[0121] Then, the initial icon design scheme is input into the discriminator for multi-scale authenticity evaluation. The discriminator includes a global discriminant branch and a local discriminant branch. The global discriminant branch uses a 5-layer convolutional network to evaluate the overall coordination of the icon and outputs a authenticity score between 0 and 1. The local discriminant branch uses a 3x3 sliding window to evaluate the authenticity of local detail features at different scales, and finally the global and local scores are weighted and averaged to obtain the final authenticity evaluation score.

[0122] Finally, the resonance coefficient of the initial icon design scheme is calculated based on the resonance scoring matrix. A 5x5 sliding window is used to analyze the dynamic change of the resonance coefficient, and the resonance scoring result is smoothed by the exponential weighted average method, with the decay factor set to 0.9. The authenticity evaluation score and the resonance scoring result are comprehensively evaluated, with the authenticity scoring weight of 0.4 and the resonance scoring weight of 0.6. The top 10% of the design schemes are selected from the comprehensive evaluation results, and their design features are extracted as high-resonance coefficient design features.

[0123] Through the above steps, this method can effectively fuse resonance scoring information and group resonance characteristics, generate icon design schemes with high resonance, and screen out high-quality design features through multi-scale evaluation. In practical applications, this method can adjust the structure and parameters of the feature fusion network according to the characteristics of different industries and target groups to meet the needs of different scenarios.

[0124] The solution of this application can:

[0125] This method can effectively improve the quality and resonance of icon design by combining a deep generative adversarial network and multi-scale evaluation. By integrating a resonance scoring matrix and group resonance enhancement features, the generated icon design better conforms to the aesthetic preferences and psychological needs of the target group, enhancing brand recognition and attractiveness. The use of a multi-branch generator and conditional constraint mechanism enables the icon design to reflect resonance characteristics in terms of shape, color, and composition, improving the overall coordination and aesthetics of the design. The combination of multi-scale authenticity evaluation and dynamic resonance calculation not only ensures the authenticity and feasibility of the design but also guarantees a high resonance effect, effectively improving the quality of icon design. This method has strong flexibility and scalability, can be adjusted and optimized according to the characteristics of different industries and target groups, and has a wide range of applications. By means of automated generation and screening, the efficiency of icon design is greatly improved, providing strong technical support for brand building.

[0126] In an alternative embodiment, the high-resonance coefficient design features are mapped to a knowledge graph-enhanced design space through a transfer learning module, and the design parameters are continuously optimized in combination with deep reinforcement learning. Based on the dynamic feedback of the user preference evolution curve, the final personalized icon design recommendation scheme is continuously adjusted and output, including:

[0127] Input the high-resonance coefficient design features into the transfer learning module. The transfer learning module includes a source domain encoder and a target domain mapper. The source domain encoder reduces the dimension and extracts features from the high-resonance coefficient design features to obtain a design feature vector, and the target domain mapper maps the design feature vector to the knowledge graph-enhanced design space to obtain a mapped feature;

[0128] Based on the mapped features, a deep reinforcement learning model is constructed. The deep reinforcement learning model includes a parameter generation network and an evaluation network. The parameter generation network fuses the mapped features with the current design parameters to generate a parameter adjustment scheme, and the evaluation network evaluates the feasibility of the parameter adjustment scheme to obtain an evaluation score;

[0129] Use the user preference evolution curve as a dynamic feedback signal to adjust the reward function weight of the deep reinforcement learning model. The parameter generation network continuously optimizes the design parameters based on the adjusted reward function to obtain optimized parameters, and the evaluation network evaluates the optimized parameters and feeds the evaluation results back to the parameter generation network;

[0130] Input the optimized parameters into the icon generation model to obtain a set of candidate icon design schemes. Calculate the preference matching degree of each scheme in the set of candidate icon design schemes based on the user preference evolution curve, and select the scheme with the highest preference matching degree as the final personalized icon design recommendation scheme.

[0131] The present invention provides a personalized icon design method based on transfer learning and deep reinforcement learning. First, the design features with high resonance coefficients are mapped to the knowledge graph-enhanced design space through a transfer learning module, then the design parameters are continuously optimized in combination with deep reinforcement learning, and finally, according to the dynamic feedback of the user preference evolution curve, the final personalized icon design recommendation scheme is continuously adjusted and output.

[0132] First, the design features with high resonance coefficients are input into the transfer learning module. This module includes two key components: a source domain encoder and a target domain mapper. The source domain encoder adopts a multi-layer convolutional neural network structure to reduce the dimension and extract features from the input design features with high resonance coefficients, obtaining a low-dimensional design feature vector. For example, the input design features with high resonance coefficients may contain information in multiple dimensions such as color, shape, layout, etc. After being processed by the source domain encoder, a 128-dimensional feature vector can be obtained. The target domain mapper then adopts a fully connected neural network structure to map the above design feature vector to the knowledge graph-enhanced design space, obtaining a mapped feature. The knowledge graph-enhanced design space here refers to a high-dimensional space that integrates a large amount of icon design knowledge, including various design elements, styles, industry characteristics, etc. Through this mapping process, the original design features are transformed into representations in the knowledge graph-enhanced design space, providing richer semantic information for subsequent optimization.

[0133] Next, a deep reinforcement learning model is constructed based on the above mapped features. This model includes two core components: a parameter generation network and an evaluation network. The parameter generation network adopts a recurrent neural network structure to fuse the mapped features with the current design parameters to generate a parameter adjustment scheme. For example, the current design parameters may include font size, color ratio, position of graphic elements, etc. The parameter generation network will propose adjustment suggestions for these parameters according to the information of the mapped features. The evaluation network then adopts a multi-layer perceptron structure to evaluate the feasibility of the parameter adjustment scheme, obtaining an evaluation score between 0 and 1. The evaluation criteria include multiple aspects such as the aesthetics, uniqueness, and industry relevance of the design.

[0134] Then, the user preference evolution curve is used as a dynamic feedback signal to adjust the reward function weight of the deep reinforcement learning model. The user preference evolution curve is a function that changes over time, reflecting the change in the user's preference for different design styles and elements. For example, the user may prefer a simple style at the beginning, but gradually tend to more complex designs over time. The system will dynamically adjust the weights of each evaluation index in the reward function according to this curve. The parameter generation network continuously optimizes the design parameters based on the adjusted reward function to obtain optimized parameters. The evaluation network then evaluates the optimized parameters and feeds back the evaluation results to the parameter generation network, forming a closed-loop optimization process.

[0135] Finally, the optimized parameters are input into the icon generation model to obtain a set of candidate icon design solutions. The icon generation model adopts the structure of a generative adversarial network (GAN) and can generate diverse icon designs based on the input parameters. The system calculates the preference matching degree of each solution in the set of candidate icon design solutions based on the user preference evolution curve. The calculation method is to perform a correlation analysis between the features of each design solution and the user preference evolution curve to obtain a matching degree score between 0 and 1. Finally, the system selects the solution with the highest preference matching degree as the recommended solution for personalized icon design.

[0136] Throughout the process, the system will continuously iterate and optimize. Each time the user's feedback on the recommended solution is used to update the user preference evolution curve, thereby affecting the next round of the optimization process. This dynamic adjustment mechanism ensures that the system can continuously adapt to the changing needs and preferences of users.

[0137] The solution of this application can:

[0138] This method effectively utilizes the existing design knowledge through transfer learning, greatly improving the starting point and efficiency of icon design. The transfer learning module can map the design features with high resonance coefficients to the design space enhanced by the knowledge graph, enabling the system to make full use of the existing design experience and knowledge, and avoiding the low efficiency problem of designing from scratch. The introduction of the deep reinforcement learning model enables the system to continuously optimize the design parameters and continuously improve the design quality. Through the cooperation of the parameter generation network and the evaluation network, the system can quickly find the optimal solution in the huge design space, greatly improving the efficiency and quality of the design. The introduction of the user preference evolution curve enables the system to have the ability of personalized recommendation and can dynamically adjust the design solution according to the changing needs of users. This dynamic feedback mechanism ensures that the system can continuously meet the personalized needs of users and improve user satisfaction.

[0139] In an alternative embodiment, taking the user preference evolution curve as a dynamic feedback signal, adjusting the reward function weight of the deep reinforcement learning model, the parameter generation network continuously optimizes the design parameters based on the adjusted reward function to obtain optimized parameters, and the evaluation network evaluates the optimized parameters and feeds back the evaluation result to the parameter generation network, including:

[0140] Inputting the user preference evolution curve into the feature extraction module to extract the preference intensity feature and the preference direction feature. The preference intensity feature represents the degree of inclination of the user to the design parameters, and the preference direction feature represents the changing trend of the user's preference. Constructing the preference intensity feature and the preference direction feature into a dynamic feedback signal;

[0141] Adjust the reward function weights of the deep reinforcement learning model based on the dynamic feedback signal, adjust the preference matching degree weights in the reward function weights according to the preference intensity feature, and adjust the parameter constraint weights in the reward function weights according to the preference direction feature to generate an adjusted reward function;

[0142] The parameter generation network continuously optimizes the design parameters based on the adjusted reward function, inputs the current design parameters and the dynamic feedback signal into the parameter generation network to obtain a parameter adjustment scheme, and evaluates the parameter adjustment scheme based on the adjusted reward function to obtain optimized parameters;

[0143] The evaluation network evaluates the optimized parameters, generates an evaluation result including a parameter feasibility score and a preference compliance score, and feeds back the evaluation result to the parameter generation network.

[0144] The present invention proposes a method for optimizing the parameters of a deep reinforcement learning model based on a user preference evolution curve. The method first uses the user preference evolution curve as a dynamic feedback signal to adjust the reward function weights of the deep reinforcement learning model. Then, the parameter generation network continuously optimizes the design parameters based on the adjusted reward function to obtain optimized parameters. Finally, the evaluation network evaluates the optimized parameters and feeds back the evaluation result to the parameter generation network.

[0145] First, input the user preference evolution curve into the feature extraction module. The feature extraction module adopts a convolutional neural network structure, including multiple convolutional layers and pooling layers. The convolutional layers use convolutional kernels of different sizes to extract features from the input curve, and the pooling layers perform dimensionality reduction operations on the feature maps. Through these operations, the preference intensity feature and the preference direction feature are extracted from the user preference evolution curve.

[0146] The preference intensity feature characterizes the degree of inclination of the user towards the design parameters. For example, for a certain design parameter, if the user preference evolution curve shows a high peak near this parameter, it indicates that the user has a strong preference for this parameter. The preference direction feature characterizes the change trend of the user's preference. For example, if the curve shows an upward trend, it indicates that the user's preference degree for this parameter is increasing; conversely, if it shows a downward trend, it indicates that the user's preference degree for this parameter is decreasing.

[0147] The extracted preference intensity feature and preference direction feature are constructed into a dynamic feedback signal. This signal is a multi-dimensional vector, where each dimension corresponds to a design parameter, and the value of the vector reflects the preference intensity and preference direction of the user for this parameter.

[0148] Next, adjust the weights of the reward function of the deep reinforcement learning model based on the dynamic feedback signal. The reward function consists of two main parts: preference matching degree and parameter constraint. The preference matching degree measures the matching degree between the generated parameters and the user's preferences, and the parameter constraint ensures that the generated parameters meet the design requirements.

[0149] Adjust the weight of the preference matching degree in the reward function according to the preference intensity feature. If the preference intensity of a certain parameter is high, then increase the weight of this parameter in the calculation of the preference matching degree. For example, if the preference intensity of a certain parameter for the user is 0.8 (assuming the preference intensity ranges from 0 to 1), then the weight of this parameter in the calculation of the preference matching degree can be set to 0.8.

[0150] Adjust the weight of the parameter constraint in the reward function according to the preference direction feature. If the preference direction of a certain parameter shows an upward trend, then appropriately relax the constraint condition of this parameter and increase the flexibility of the parameter value. For example, if the value range of a certain parameter is originally [0,100] and the preference direction shows an upward trend, then the value range can be expanded to [-10,110], and at the same time, reduce the weight of the parameter constraint in the reward function.

[0151] Through the above adjustments, a new reward function is generated. This adjusted reward function can better reflect the user's real-time preferences and is conducive to generating design parameters that better meet the user's needs.

[0152] Then, the parameter generation network continuously optimizes the design parameters based on the adjusted reward function. The parameter generation network adopts a recurrent neural network structure, which can process sequential data and maintain long-term dependencies. The input of the network includes the current design parameters and the dynamic feedback signal.

[0153] The parameter generation network first concatenates the current design parameters and the dynamic feedback signal to form a comprehensive input vector. This vector is processed by a multi-layer recurrent neural network to generate a parameter adjustment scheme. For example, if the current value of a certain design parameter is 50 and the dynamic feedback signal shows that the user's preference intensity for this parameter is 0.7 and the preference direction is upward, then the parameter generation network may generate an adjustment scheme to increase the parameter value to 60.

[0154] The generated parameter adjustment scheme is then evaluated through the adjusted reward function. During the evaluation process, calculate the preference matching degree and the degree of satisfaction of the parameter constraint of the adjusted parameters, and perform a weighted sum according to the weights of the reward function to obtain a comprehensive score. If the score is higher than the score of the current parameters, then accept this adjustment scheme and update it as the new optimized parameters.

[0155] Finally, the evaluation network evaluates the optimized parameters. The evaluation network adopts a feedforward neural network structure and consists of multiple fully connected layers. The input of the network is the optimized design parameters, and the output is the evaluation result including the parameter feasibility score and the preference compliance score.

[0156] The parameter feasibility score measures whether the optimized parameters meet the design requirements and constraints. For example, if the value range of a certain parameter is [0, 100] and the optimized parameter value is 80, then the feasibility score of this parameter is relatively high. The preference compliance score measures the matching degree between the optimized parameters and the user preferences. For example, if the user's preferred value for a certain parameter is 70 and the optimized parameter value is 75, then the preference compliance score of this parameter is relatively high.

[0157] The evaluation result generated by the evaluation network is then fed back to the parameter generation network. The parameter generation network adjusts its internal state and parameters according to this feedback information to generate a better parameter adjustment scheme in the next round of optimization.

[0158] Through continuous iteration of the above steps, the deep reinforcement learning model can continuously optimize the design parameters to make them meet both the design requirements and better conform to the dynamic preferences of users.

[0159] The solution of this application can:

[0160] This method can capture the changes in user preferences in real time and convert them into quantifiable dynamic feedback signals to guide the parameter optimization process of the deep reinforcement learning model. This dynamic adjustment mechanism enables the model to respond more quickly and accurately to the changes in user needs, improving the efficiency and accuracy of parameter optimization. By decomposing the user preference evolution curve into preference intensity features and preference direction features, this method can more comprehensively understand the internal structure of user preferences. This fine-grained feature extraction helps to generate design parameters that better meet the actual needs of users, improving the quality of the optimization results and user satisfaction. This method adopts a dual-network structure combining a parameter generation network and an evaluation network, realizing the decoupling of parameter generation and evaluation. This structural design enhances the generalization ability and robustness of the model, enabling it to adapt to a wider range of design scenarios and user groups, and also improving the interpretability and controllability of the model.

[0161] Figure 2 This is a schematic structural diagram of the intelligent recommendation and generation system for design styles based on user preferences according to an embodiment of the present invention, as Figure 2 shown, the system includes:

[0162] The first unit is used to collect design preference selections, rating feedback, annotation information, and interaction review texts in the explicit interaction data of users, and at the same time obtain browsing trajectories, stay durations, and operation sequences in the implicit behavior data of users. Based on the explicit interaction data and implicit behavior data, a user preference feature vector is constructed; the user preference feature vector is input into a time-series deep learning network, and the user preference evolution law is analyzed through a periodic pattern recognition algorithm to generate a user preference evolution curve, and the importance weights of each dimension feature are calculated based on the attention mechanism, and the feature importance scores are output;

[0163] The second unit is used to calculate the time-series correlation degree between the icon design features and the user preference features by using a multi-head attention network according to the user preference evolution curve and the feature importance scores, and construct a resonance scoring matrix; based on the resonance scoring matrix, an adaptive resonance threshold screening model is designed. The adaptive resonance threshold screening model tracks the dynamic changes of the resonance coefficient through a bidirectional long short-term memory network, and combines a swarm intelligence algorithm to optimize the resonance scores at the group level to generate group resonance enhanced features, and the group resonance enhanced features are used to guide the dynamic adjustment of subsequent icon design schemes;

[0164] The third unit is used to input the resonance scoring matrix and the group resonance enhanced features into a deep generative adversarial network, and create an initial icon design scheme through a conditional constrained multi-branch generator; perform multi-scale authenticity evaluation and resonance degree calculation on the initial icon design scheme, and screen out design features with high resonance coefficients; map the high resonance coefficient design features to a knowledge graph-enhanced design space through a transfer learning module, and continuously optimize the design parameters in combination with deep reinforcement learning, and continuously adjust and output the final personalized icon design recommendation scheme according to the dynamic feedback of the user preference evolution curve.

[0165] In the third aspect of the embodiments of the present invention,

[0166] An electronic device is provided, including:

[0167] A processor;

[0168] A memory for storing instructions executable by the processor;

[0169] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0170] In the fourth aspect of the embodiments of the present invention,

[0171] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0172] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for performing various aspects of the present invention.

[0173] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An intelligent recommendation and generation method for design styles based on user preferences, characterized in that Including: Collect the design preference selections, rating feedback, annotation information, and interaction review texts in the explicit interaction data of users. At the same time, obtain the browsing trajectories, stay durations, and operation sequences in the implicit behavior data of users. Based on the explicit interaction data and implicit behavior data, construct a user preference feature vector; input the user preference feature vector into a temporal deep learning network, analyze the evolution law of user preferences through a periodic pattern recognition algorithm, generate a user preference evolution curve, and calculate the importance weights of features in each dimension based on the attention mechanism, and output a feature importance score; According to the user preference evolution curve and the feature importance score, use a multi-head attention network to calculate the temporal correlation degree between the icon design features and the user preference features, and construct a resonance score matrix; based on the resonance score matrix, design an adaptive resonance threshold screening model. The adaptive resonance threshold screening model tracks the dynamic changes of the resonance coefficient through a bidirectional long short-term memory network, and combines a swarm intelligence algorithm to optimize the resonance score at the group level, generating a group resonance enhanced feature, and the group resonance enhanced feature is used to guide the dynamic adjustment of the subsequent icon design scheme; Input the resonance score matrix and the group resonance enhanced feature into a deep generative adversarial network, and create an initial icon design scheme through a conditional constrained multi-branch generator; perform multi-scale authenticity evaluation and resonance degree calculation on the initial icon design scheme, and screen to obtain high resonance coefficient design features; map the high resonance coefficient design features to the knowledge graph enhanced design space through a transfer learning module, continuously optimize the design parameters in combination with deep reinforcement learning, and continuously adjust and output the final personalized icon design recommendation scheme according to the dynamic feedback of the user preference evolution curve.

2. The method according to claim 1, characterized in that, Input the user preference feature vector into a temporal deep learning network, analyze the evolution law of user preferences through a periodic pattern recognition algorithm, generate a user preference evolution curve, and calculate the importance weights of features in each dimension based on the attention mechanism, and the output feature importance score includes: Input the user preference feature vector into a temporal deep learning network. The temporal deep learning network includes a feature extraction layer and a temporal analysis layer. The feature extraction layer uses a bidirectional gated recurrent unit to extract the temporal features of the user preference feature vector. The temporal analysis layer uses a long short-term memory network to perform long-term dependence analysis on the temporal features to obtain a user preference temporal feature sequence; Perform periodic pattern analysis on the user preference temporal feature sequence, extract periodic features through Fourier transform, combine the moving average method with an adaptive window to identify local periodic patterns, and use a dynamic programming algorithm to optimize the combination of the local periodic patterns to obtain the evolution law of user preferences; Based on the evolution law of user preferences, use a piecewise polynomial regression model to fit the change trend of user preferences, determine key change nodes through an adaptive breakpoint selection algorithm, and generate a user preference evolution curve in combination with spline smoothing technology; Construct a multi-layer attention network to calculate the importance of each dimension feature in the user preference evolution curve. The multi-layer attention network captures the mutual relationship between feature dimensions through the self-attention mechanism and analyzes the evolution characteristics of features over time using the cross-attention mechanism; Based on the output result of the multi-layer attention network, use the analytic hierarchy process to calculate the initial weights of each dimension feature in the user preference curve, dynamically adjust the initial weights by combining time series correlation analysis, and obtain the feature importance score through normalization processing.

3. The method according to claim 1, wherein Based on the resonance score matrix, design an adaptive resonance threshold screening model. The adaptive resonance threshold screening model tracks the dynamic changes of the resonance coefficient through a bidirectional long short-term memory network, and combines a swarm intelligence algorithm to optimize the resonance score at the group level, generating group resonance enhanced features. The group resonance enhanced features are used to guide the dynamic adjustment of the subsequent icon design scheme, including: Construct an adaptive resonance threshold screening model based on the resonance score matrix. The adaptive resonance threshold screening model uses a sliding window to segment the resonance score matrix, extracts the score distribution characteristics through kernel density estimation, dynamically calculates the resonance threshold by combining the distribution kurtosis and skewness, and obtains the threshold screening result; Input the threshold screening result into a bidirectional long short-term memory network. The bidirectional long short-term memory network includes a gating adjustment layer and a feature fusion layer. The gating adjustment layer uses a gating mechanism to adjust the information retention ratio, and the feature fusion layer introduces residual connection and layer normalization processing, fuses bidirectional features through the attention mechanism, and tracks the dynamic changes of the resonance coefficient; Based on the dynamic changes of the resonance coefficient, construct a swarm intelligence optimization algorithm. The swarm intelligence optimization algorithm takes the scoring weight and threshold parameter as optimization objectives, enhances the search ability by using the local optimal solution item, dynamically adjusts the inertia weight by combining the non-linear decreasing strategy, and optimizes the resonance score at the group level to obtain the group-level optimization result; Input the group-level optimization result into a feature fusion network. The feature fusion network adjusts the historical features through a time decay factor, analyzes the feature correlation using a spatial attention mechanism, integrates the group opinions based on the confidence weighted method, and generates group resonance enhanced features. The group resonance enhanced features are used to guide the dynamic adjustment of the icon design scheme.

4. The method according to claim 3, characterized in that, Based on the dynamic changes of the resonance coefficient, construct a swarm intelligence optimization algorithm. The swarm intelligence optimization algorithm takes the scoring weight and threshold parameter as optimization objectives, enhances the search ability by using the local optimal solution item, dynamically adjusts the inertia weight by combining the non-linear decreasing strategy, and optimizes the resonance score at the group level to obtain the group-level optimization result, including: Construct a swarm intelligence optimization algorithm based on the dynamic changes of the resonance coefficient. The swarm intelligence optimization algorithm takes the scoring weight and threshold parameter as optimization objectives. The scoring weight is used to quantify the importance of each dimension feature in the resonance score, and the threshold parameter is used to control the screening criteria of the resonance score; Initialize the population of the swarm intelligence optimization algorithm, calculate the fitness values of each individual in the initial population, and screen high-quality solutions according to the fitness values to construct a local optimal solution set, which is used to enhance the local search ability of the algorithm; During the iterative optimization process, according to the dynamic change characteristics of the resonance coefficient, select high-quality solutions from the local optimal solution set for local search, dynamically adjust the scope and accuracy of local search, and enhance the search ability of the algorithm through the local optimal solutions; Adopt a non-linear decreasing strategy to dynamically adjust the inertia weight, determine the weight decreasing coefficient according to the dynamic change degree of the resonance coefficient, and adaptively update the inertia weight using the weight decreasing coefficient, where the inertia weight is used to adjust the swarm search behavior; Iteratively optimize the population based on the updated inertia weight, calculate the optimized swarm fitness value, update the local optimal solution set according to the swarm fitness value, and continuously iterate until convergence, and output the resonance scoring result optimized at the swarm level.

5. The method according to claim 1, wherein Input the resonance scoring matrix and the swarm resonance enhancement feature into a deep generative adversarial network, and create an initial icon design scheme through a conditional multi-branch generator; Conduct multi-scale authenticity evaluation and resonance degree calculation on the initial icon design scheme, and the screened high-resonance coefficient design features include: Input the resonance scoring matrix and the swarm resonance enhancement feature into a feature fusion network. The feature fusion network uses a multi-head attention mechanism to calculate feature weights, performs principal component analysis and dimensionality reduction processing on the resonance scoring matrix, and adaptively fuses the dimensionality-reduced features with the swarm resonance enhancement feature to generate a conditional vector; Construct a deep generative adversarial network based on the conditional vector. The deep generative adversarial network includes a multi-branch generator and a discriminator. The multi-branch generator includes a shape generation branch, a color generation branch, and a composition generation branch. Each branch introduces a conditional normalization layer to inject the conditional vector into the generation process, and uses an adaptive weighting method to fuse the output features of each branch to generate an initial icon design scheme; Input the initial icon design scheme into the discriminator for multi-scale authenticity evaluation. The discriminator includes a global discriminant branch and a local discriminant branch. The global discriminant branch evaluates the overall coordination of the icon, and the local discriminant branch evaluates the detailed features of the icon to generate an authenticity evaluation score; Calculate the resonance coefficient of the initial icon design scheme based on the resonance scoring matrix, analyze the dynamic change of the resonance coefficient using a sliding window, smooth the resonance scoring result through an exponentially weighted method, comprehensively evaluate the authenticity evaluation score and the resonance scoring result, and screen high-resonance coefficient design features from the results of the comprehensive evaluation.

6. The method according to claim 1, wherein Map the high-resonance coefficient design features to the knowledge graph-enhanced design space through a transfer learning module, continuously optimize the design parameters in combination with deep reinforcement learning, and continuously adjust and output the final personalized icon design recommendation scheme according to the dynamic feedback of the user preference evolution curve, including: Input the high resonance coefficient design feature into the transfer learning module, which includes a source domain encoder and a target domain mapper. The source domain encoder reduces the dimension and extracts features from the high resonance coefficient design feature to obtain a design feature vector, and the target domain mapper maps the design feature vector to the knowledge graph enhanced design space to obtain a mapped feature; Construct a deep reinforcement learning model based on the mapped feature. The deep reinforcement learning model includes a parameter generation network and an evaluation network. The parameter generation network fuses the mapped feature with the current design parameter to generate a parameter adjustment scheme, and the evaluation network evaluates the feasibility of the parameter adjustment scheme to obtain an evaluation score; Use the user preference evolution curve as a dynamic feedback signal to adjust the reward function weight of the deep reinforcement learning model. The parameter generation network continuously optimizes the design parameter based on the adjusted reward function to obtain an optimized parameter, and the evaluation network evaluates the optimized parameter and feeds back the evaluation result to the parameter generation network; Input the optimized parameter into the icon generation model to obtain a set of candidate icon design schemes. Calculate the preference matching degree of each scheme in the set of candidate icon design schemes based on the user preference evolution curve, and select the scheme with the highest preference matching degree as the final personalized icon design recommendation scheme.

7. The method according to claim 6, wherein Using the user preference evolution curve as a dynamic feedback signal to adjust the reward function weight of the deep reinforcement learning model, the parameter generation network continuously optimizes the design parameter based on the adjusted reward function to obtain an optimized parameter, and the evaluation network evaluates the optimized parameter and feeds back the evaluation result to the parameter generation network includes: Input the user preference evolution curve into the feature extraction module to extract the preference intensity feature and the preference direction feature. The preference intensity feature characterizes the degree of the user's preference for the design parameter, and the preference direction feature characterizes the change trend of the user's preference. Construct the preference intensity feature and the preference direction feature into a dynamic feedback signal; Adjust the reward function weight of the deep reinforcement learning model based on the dynamic feedback signal. Adjust the preference matching degree weight in the reward function weight according to the preference intensity feature, and adjust the parameter constraint weight in the reward function weight according to the preference direction feature to generate an adjusted reward function; The parameter generation network continuously optimizes the design parameter based on the adjusted reward function. Input the current design parameter and the dynamic feedback signal into the parameter generation network to obtain a parameter adjustment scheme, and evaluate the parameter adjustment scheme based on the adjusted reward function to obtain an optimized parameter; The evaluation network evaluates the optimized parameter, generates an evaluation result including a parameter feasibility score and a preference compliance score, and feeds back the evaluation result to the parameter generation network.

8. An intelligent design style recommendation and generation system based on user preferences, which is used to implement the method described in any one of the foregoing claims 1-7, and is characterized in that, Includes: The first unit is used to collect design preference selections, rating feedback, annotation information, and interactive review texts in the explicit interaction data of users. Meanwhile, it obtains browsing trajectories, stay durations, and operation sequences in the implicit behavior data of users. Based on the explicit interaction data and implicit behavior data, it constructs a user preference feature vector; inputs the user preference feature vector into a temporal deep learning network, analyzes the evolution law of user preferences through a periodic pattern recognition algorithm, generates a user preference evolution curve, and calculates the importance weights of features in each dimension based on the attention mechanism, and outputs a feature importance score; The second unit is used to calculate the temporal correlation degree between icon design features and user preference features by using a multi-head attention network according to the user preference evolution curve and the feature importance score, and construct a resonance score matrix; based on the resonance score matrix, it designs an adaptive resonance threshold screening model. The adaptive resonance threshold screening model tracks the dynamic changes of the resonance coefficient through a bidirectional long short-term memory network, and combines a swarm intelligence algorithm to optimize the resonance score at the group level, generating a group resonance enhancement feature, which is used to guide the dynamic adjustment of subsequent icon design schemes; The third unit is used to input the resonance score matrix and the group resonance enhancement feature into a deep generative adversarial network, and create an initial icon design scheme through a conditional constrained multi-branch generator; conduct multi-scale authenticity evaluation and resonance degree calculation on the initial icon design scheme, and screen out design features with high resonance coefficients; map the design features with high resonance coefficients to a knowledge graph-enhanced design space through a transfer learning module, and continuously optimize the design parameters in combination with deep reinforcement learning, and continuously adjust and output a final personalized icon design recommendation scheme according to the dynamic feedback of the user preference evolution curve.

9. An electronic device, characterized in that, Comprising: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images