Interface color matching scheme recommendation system based on artificial intelligence

Abstract

The invention generally relates to a system for generating recommendation, in particular to an interface color matching scheme recommendation system based on artificial intelligence, which comprises a data collection subsystem, a color matching scheme generation subsystem, an intelligent recommendation subsystem, a user interaction subsystem and an optimization subsystem. Edge feature extraction ensures that contour and boundary information of interface elements can be accurately obtained, and deep feature extraction further captures complex textures, shapes and color matching modes in interface samples; by combining meta learning, reinforcement learning and a Lagrangian multiplier method, the Lagrangian multiplier method is used for processing a multi-objective optimization problem when a color matching scheme is optimized; and a self-supervised heterogeneous graph neural network and collaborative filtering are combined in the intelligent recommendation subsystem, so that the accuracy and diversity of personalized recommendation are improved.

Description

Technical Field

[0001] The present invention generally relates to a system for generating recommendations, and more particularly to an interface color scheme recommendation system based on artificial intelligence. Background Art

[0002] With the rapid development of artificial intelligence technology, interface color scheme recommendation has gradually become a key part of improving user experience. To recommend interface color schemes, it is necessary to first collect existing color schemes, which involves feature extraction. The general feature extraction method has the problem of insufficient extraction, usually ignoring the fusion of multi-scale information, unable to capture local features and global features at the same time, and focusing more on the visual content itself, while ignoring the personalized preferences of users; constructing a color generation model in the color scheme generation subsystem, and performing hyperparameter tuning on the constructed color generation model. The traditional hyperparameter tuning method has the problem of one-sided optimization objectives, ignoring other constraint objectives or the balance between multiple objectives, resulting in compromises in certain dimensions of the color generation model; the ordinary recommendation system has the problem of inaccurate recommendations, and the relevance and personalization of the recommendation results are low, which cannot fully meet the needs of users. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides an interface color scheme recommendation system based on artificial intelligence. In view of the problem that general feature extraction methods are not detailed enough, the present invention uses two-dimensional feature extraction for interface collection samples, edge feature extraction ensures that the contour and boundary information of interface elements can be accurately obtained, and deep feature extraction further captures the complex textures, shapes and color patterns in interface samples; in view of the problem that traditional hyperparameter tuning methods have one-sided optimization targets, the present invention combines meta-learning, reinforcement learning and Lagrange multiplier method, and uses Lagrange multiplier method to deal with multi-objective optimization problems when optimizing color schemes, especially constraint targets, and provides corresponding feedback information for each color recommendation to ensure the stability and efficiency of the system under different conditions; in view of the problem that ordinary recommendation systems have inaccurate recommendations, the present invention uses a combination of self-supervised heterogeneous graph neural networks and collaborative filtering in the intelligent recommendation subsystem, and achieves accurate matching of user preferences and color schemes through the fusion of deep features and broad features, thereby improving the accuracy and diversity of personalized recommendations.

[0004] The technical solution adopted by the present invention is as follows: The present invention provides an interface color scheme recommendation system based on artificial intelligence, including a data collection subsystem, a color scheme generation subsystem, an intelligent recommendation subsystem, a user interaction subsystem and an optimization subsystem, specifically including the following contents:

[0005] The data collection subsystem collects interface samples, including existing color schemes, interface screenshots and layout structures, and also obtains user preference and behavior data;

[0006] The color scheme generation subsystem generates candidate color schemes by constructing a color scheme generation model;

[0007] The intelligent recommendation subsystem recommends personalized color schemes based on the generated candidate color schemes by combining self-supervised heterogeneous graph neural networks and collaborative filtering;

[0008] The user interaction subsystem provides a visual interface for users to view personalized color schemes in real time, and provides a feedback mechanism to obtain user feedback and usage data;

[0009] The optimization subsystem optimizes the color scheme generation model according to user feedback and usage data adjustment.

[0010] Furthermore, the color scheme generation subsystem constructs a color scheme generation model, which specifically includes the following steps:

[0011] Step S1: Data preparation, the receiving interface collects samples, user preferences and behavior data;

[0012] Step S2: feature extraction, using two-dimensional feature extraction on the interface collection samples to obtain interface features;

[0013] Step S3: Build a model, using the GAN model as the basic architecture to build a color matching generation model;

[0014] Step S4: Model training, using the interface to collect samples, user preferences and behavior data to train the color matching generation model, and repeatedly perform hyperparameter tuning;

[0015] Step S5: Color scheme generation: using the trained color scheme generation model to generate candidate color schemes.

[0016] Furthermore, in step S2, two-dimensional feature extraction is performed on the interface collection samples, specifically including the following steps:

[0017] Step S21: edge feature extraction, using a filter to convolve the interface collection samples, extracting the edge extraction in the interface collection samples, performing multi-layer convolution processing on the interface collection samples through the ResNet structure, extracting deep features, and generating a multi-level feature map, in which each pixel position represents the feature of a part of the interface collection sample;

[0018] Step S22: dense connection feature fusion, using a dense connection mechanism between convolutional layers, the output of each layer is passed to the next layer and returned to each subsequent layer, and a multi-scale feature map is generated by reusing deep features;

[0019] Step S23: spatial channel attention mechanism, for each multi-scale feature map, calculates the attention weight of each pixel position, and weights the channel dimension of the multi-scale feature map to generate a weighted feature map;

[0020] Step S24: Feature extraction, combining edge features and deep features, performing convolution operations to extract interface features.

[0021] Further, step S4, repeatedly performing hyperparameter tuning, specifically includes the following steps:

[0022] Step S41: Initialization, initializing the environment, experience replay pool, internal parameters and external hyperparameters. The internal parameters include the parameters of the policy network, the objective function of the policy network and the Lagrange multiplier. The external hyperparameters include the safety threshold and entropy temperature.

[0023] Step S42: interact with the environment, use the policy network to sample in the environment, obtain batch data of state-action-reward data and store it in the experience replay pool;

[0024] Step S43: Update internal parameters, sample batch data from the experience replay pool, and optimize the objective function of the policy network by maximizing rewards and entropy regularization;

[0025] Step S44: Optimize the Lagrange multiplier, combine the main objective and the constraint objective, and use the gradient downward method to update the policy network parameters;

[0026] Step S45: dynamically adjusting the Lagrange multiplier according to the constraint target;

[0027] Step S46: Optimize the hyperparameter target, define the hyperparameter regularization target, and optimize the hyperparameter;

[0028] Step S47: Set the convergence standard, terminate the tuning after the main objective function converges, and output the optimized hyperparameters.

[0029] Furthermore, in the intelligent recommendation subsystem, the method of combining the self-supervised heterogeneous graph neural network and collaborative filtering specifically includes the following steps:

[0030] Step Q1: User preference data processing, obtaining the user's historical color selection, click behavior, rating data and user portrait;

[0031] Step Q2: Construct a user-color matching graph structure, taking each candidate color matching scheme as a node and the user as a node of another category, initializing the feature vector of the node, and the edge weight between the user and the historical color matching selection represents the degree of preference. The edges between the candidate color matching schemes are constructed by similarity;

[0032] Step Q3: Collect deep features, define the target node as all candidate color schemes, set the domain range to the 1-hop neighbor to the k-hop neighbor of the target node, use the graph neural network to process the features of the neighbors in the neighborhood layer by layer, assign weights to neighbors with different hop numbers, and fuse the features of all layers to form a deep feature representation of the target node;

[0033] Step Q4: Collect broad features, define the direct domain as the 1-hop neighbors of the target node, use a single-layer graph convolutional network to aggregate the direct neighborhood of the target node, and generate a broad feature representation of the target node;

[0034] Step Q5: Collaborative contrastive learning, constructing deep view and wide view according to the relationship between the target node and the deep neighbor and the direct neighbor, constructing positive and negative samples of the two features, setting the deep feature representation and the wide feature representation from the same target node as positive samples, and the deep feature representation and the wide feature representation from different target nodes as negative samples, for each pair of positive and negative samples, calculating the contrast loss function and the total loss function, and optimizing the deep feature representation and the wide feature representation;

[0035] Step Q6: Generate node embedding, fuse the optimized deep feature representation and broad feature representation into the final embedding vector of the node, and normalize it;

[0036] Step Q7: Collaborative filtering score prediction, using the final embedding vector of the node for score prediction, and using mean square error loss to optimize the score prediction to obtain the predicted score;

[0037] Step Q8: Recommendation generation, generating a personalized color scheme recommendation list based on the predicted score.

[0038] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0039] (1) In view of the problem that general feature extraction methods are not detailed enough, the present invention uses two-dimensional feature extraction for interface collection samples. Edge feature extraction ensures that the contour and boundary information of interface elements can be accurately obtained, and deep feature extraction further captures the complex textures, shapes and color patterns in interface samples;

[0040] (2) In view of the problem that traditional hyperparameter tuning methods have one-sided optimization objectives, the present invention combines meta-learning, reinforcement learning and Lagrange multiplier method, and uses Lagrange multiplier method to deal with multi-objective optimization problems when optimizing color schemes, especially constraint objectives, and provides corresponding feedback information for each color recommendation, thereby ensuring the stability and efficiency of the system under different conditions;

[0041] (3) To address the problem of inaccurate recommendations in common recommendation systems, the present invention combines self-supervised heterogeneous graph neural networks and collaborative filtering in the intelligent recommendation subsystem. By fusing deep features with broad features, it achieves accurate matching of user preferences and color schemes, thereby improving the accuracy and diversity of personalized recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a schematic diagram of an interface color scheme recommendation system based on artificial intelligence proposed by the present invention.

[0043] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0045] Example 1, see Figure 1 The present invention provides an interface color scheme recommendation system based on artificial intelligence, including a data collection subsystem, a color scheme generation subsystem, an intelligent recommendation subsystem, a user interaction subsystem and an optimization subsystem, which specifically includes the following contents:

[0046] The data collection subsystem collects interface samples, including existing color schemes, interface screenshots and layout structures, and also obtains user preference and behavior data;

[0047] The color scheme generation subsystem generates candidate color schemes by constructing a color scheme generation model;

[0048] The intelligent recommendation subsystem recommends personalized color schemes based on the generated candidate color schemes by combining self-supervised heterogeneous graph neural networks and collaborative filtering;

[0049] The user interaction subsystem provides a visual interface for users to view personalized color schemes in real time, and provides a feedback mechanism to obtain user feedback and usage data;

[0050] The optimization subsystem optimizes the color scheme generation model according to user feedback and usage data adjustment.

[0051] In this embodiment, when a designer needs to select a new color scheme for his mobile application interface, the system will generate several candidate color schemes by uploading the interface layout and theme requirements and push them to the designer through the intelligent recommendation subsystem. The designer will view the color scheme effect through the interactive interface, score and make minor adjustments to the scheme. The system will store the designer's feedback data and use it to optimize the color scheme generation model.

[0052] Embodiment 2: Based on the above embodiment, the color scheme generation subsystem constructs a color scheme generation model, specifically including the following steps:

[0053] Step S1: Data preparation, the receiving interface collects samples, user preferences and behavior data;

[0054] Step S2: feature extraction, using two-dimensional feature extraction on the interface collection samples to obtain interface features;

[0055] Step S3: Build a model, using the GAN model as the basic architecture to build a color matching generation model;

[0056] Step S4: Model training, using the interface to collect samples, user preferences and behavior data to train the color matching generation model, and repeatedly perform hyperparameter tuning;

[0057] Step S5: Color scheme generation: using the trained color scheme generation model to generate candidate color schemes.

[0058] In this embodiment, the GAN model is used as the basic architecture in step S3, which specifically includes the following contents:

[0059] Generator: receives the input interface feature vector, user preference data and random noise, and generates a color scheme vector;

[0060] Discriminator: compares the generated color scheme with the real user historical preference data to distinguish between true and false color schemes;

[0061] The discriminator's loss function combines perceptual loss and color scheme consistency loss to ensure that the generated color scheme is both in line with user aesthetics and highly adaptable to the interface layout.

[0062] In this embodiment, after the model training is completed in step S5, the designer inputs the feature vector of the target interface layout and the desired style constraints, including simple cold tones or lively warm tones, and the color generation model generates a series of candidate color schemes, including:

[0063] Color scheme A: The primary color is light blue, the secondary color is off-white, and the accent color is orange-yellow;

[0064] Color scheme B: The primary color is dark green, the secondary color is light beige, and the accent color is bright yellow.

[0065] Embodiment 3: This embodiment is based on the above embodiment. In step S2, two-dimensional feature extraction is used for the interface collection samples, which specifically includes the following steps:

[0066] Step S21: edge feature extraction, using a filter to convolve the interface collection samples, extracting the edge extraction in the interface collection samples, performing multi-layer convolution processing on the interface collection samples through the ResNet structure, extracting deep features, and generating a multi-level feature map, in which each pixel position represents the feature of a part of the interface collection sample;

[0067] Step S22: dense connection feature fusion, using a dense connection mechanism between convolutional layers, the output of each layer is passed to the next layer and returned to each subsequent layer, and a multi-scale feature map is generated by reusing deep features;

[0068] Step S23: spatial channel attention mechanism, for each multi-scale feature map, calculates the attention weight of each pixel position, and weights the channel dimension of the multi-scale feature map to generate a weighted feature map;

[0069] Step S24: Feature extraction, combining edge features and deep features, performing convolution operations to extract interface features.

[0070] In this embodiment, in step S21, for example, the interface of the online education platform contains rich graphic elements and text layouts. The system first applies a Sobel filter to the interface sample for convolution to highlight the boundary features and detect the outlines of elements such as buttons, icons, and picture frames; a deep convolution network based on the ResNet structure is used to perform multi-layer convolution processing on elements such as buttons, icons, and picture frames to extract deep features, such as the texture details of icons and the font features of title text. Finally, a multi-level feature map is generated, which clearly describes the structural information in the interface sample.

[0071] In step S22, during the convolution process, in order to avoid information loss in the deep feature map, the system uses a dense connection mechanism; for example, the output of the first layer is directly input into the second layer, and is simultaneously passed to the third and fourth layers; this feature reuse mechanism ensures that each layer of convolution can fuse shallow and deep information to generate a multi-scale feature map. The shallow features retain the interface structure information, such as the shape of the button, and the deep features describe the texture and color distribution, such as the change of the gradient background.

[0072] In step S23, for the generated multi-scale feature map, the system calculates the attention weight of each pixel. For example, the system finds that certain button areas have a greater impact on user behavior and assigns higher weights, while background areas have lower weights. At the same time, for the channel dimensions of the feature map, such as the color channel and the edge feature channel, the system enhances the dimensional signals related to color matching through weighted operations, such as enhancing the saliency of the color gradient area and weakening irrelevant gray background information.

[0073] Embodiment 4, based on the above embodiment, step S4, repeatedly performs hyperparameter tuning, specifically comprising the following steps:

[0074] Step S41: Initialization, initializing the environment, experience replay pool, internal parameters and external hyperparameters. The internal parameters include the parameters of the policy network, the objective function of the policy network and the Lagrange multiplier. The external hyperparameters include the safety threshold and entropy temperature.

[0075] Step S42: interact with the environment, use the policy network to sample in the environment, obtain batch data of state-action-reward data and store it in the experience replay pool;

[0076] Step S43: Update internal parameters, sample batch data from the experience replay pool, and optimize the objective function of the policy network by maximizing rewards and entropy regularization;

[0077] Step S44: Optimize the Lagrange multiplier, combine the main objective and the constraint objective, and use the gradient downward method to update the policy network parameters;

[0078] Step S45: dynamically adjusting the Lagrange multiplier according to the constraint target;

[0079] Step S46: Optimize the hyperparameter target, define the hyperparameter regularization target, and optimize the hyperparameter;

[0080] Step S47: Set the convergence standard, terminate the tuning after the main objective function converges, and output the optimized hyperparameters.

[0081] In this embodiment, for example, a company is developing an interface color scheme recommendation system to meet the personalized needs of users for interface aesthetics and readability. In the color scheme generation subsystem of the system, the generation model is hyperparameter tuned as follows:

[0082] Step S41: Initialize the environment: define the task scenario for color scheme generation, including sample interface layout and user behavior simulation;

[0083] Initialize the experience replay pool: create a buffer to store data for recording the state-action-reward data generated in each interaction;

[0084] Initialize internal parameters: including the initial parameters of the policy network (such as the weights of the color scheme generation network) and the Lagrange multipliers;

[0085] Initialize external hyperparameters: set safety thresholds, such as the minimum color contrast ratio of 4.5:1 and the initial entropy temperature of 0.2;

[0086] Step S42: displaying the generated interface in a simulation environment to simulate the interaction between the user and the interface;

[0087] Collect data generated by the interaction, including interface state (such as layout features), actions (selected color schemes), and rewards (user preference scores or system-calculated aesthetic scores), and store them in the experience replay pool;

[0088] Step S43: Sample batch data from the experience replay pool, and optimize the objective function of the policy network by maximizing rewards (user ratings) and entropy regularization (improving solution diversity);

[0089] Step S44: combining the main goal (maximizing user rating) and the constraint goal (achieving the minimum color contrast standard), the policy network parameters are updated using the gradient descent method. When the generated color scheme does not meet the contrast standard, the system adjusts the parameters to be more inclined to generate a high-contrast scheme;

[0090] Step S45: dynamically adjust the Lagrangian multiplier according to the achievement of the constraint target (such as color matching contrast). If the color matching schemes generated multiple times do not meet the contrast standard, the system increases the Lagrangian multiplier to strengthen the optimization of the constraint conditions;

[0091] Step S46: define hyperparameter regularization targets, such as entropy temperature adjustment range (0.1 to 0.5) and safety threshold dynamic range (4.0 to 6.0);

[0092] Step S47: When the objective function (user rating) and constraint target (contrast compliance rate) of the policy network both reach the set convergence criteria, stop tuning and output the optimized hyperparameters, with the entropy temperature being 0.35 and the contrast safety threshold being 5.2.

[0093] Embodiment 5, based on the above embodiment, in the intelligent recommendation subsystem, the method of combining self-supervised heterogeneous graph neural network and collaborative filtering specifically includes the following steps:

[0094] Step Q1: User preference data processing, obtaining the user's historical color selection, click behavior, rating data and user portrait;

[0095] Step Q2: Construct a user-color matching graph structure, taking each candidate color matching scheme as a node and the user as a node of another category, initializing the feature vector of the node, and the edge weight between the user and the historical color matching selection represents the degree of preference. The edges between the candidate color matching schemes are constructed by similarity;

[0096] Step Q3: Collect deep features, define the target node as all candidate color schemes, set the domain range to the 1-hop neighbor to the k-hop neighbor of the target node, use the graph neural network to process the features of the neighbors in the neighborhood layer by layer, assign weights to neighbors with different hop numbers, and fuse the features of all layers to form a deep feature representation of the target node;

[0097] Step Q4: Collect broad features, define the direct domain as the 1-hop neighbors of the target node, use a single-layer graph convolutional network to aggregate the direct neighborhood of the target node, and generate a broad feature representation of the target node;

[0098] Step Q5: Collaborative contrastive learning, constructing deep view and wide view according to the relationship between the target node and the deep neighbor and the direct neighbor, constructing positive and negative samples of the two features, setting the deep feature representation and the wide feature representation from the same target node as positive samples, and the deep feature representation and the wide feature representation from different target nodes as negative samples, for each pair of positive and negative samples, calculating the contrast loss function and the total loss function, and optimizing the deep feature representation and the wide feature representation;

[0099] Step Q6: Generate node embedding, fuse the optimized deep feature representation and broad feature representation into the final embedding vector of the node, and normalize it;

[0100] Step Q7: Collaborative filtering score prediction, using the final embedding vector of the node for score prediction, and using mean square error loss to optimize the score prediction to obtain the predicted score;

[0101] Step Q8: Recommendation generation, generating a personalized color scheme recommendation list based on the predicted score.

[0102] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0103] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0104] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. The present invention provides an artificial intelligence-based interface color scheme recommendation system, which includes a data collection subsystem, a color scheme generation subsystem, an intelligent recommendation subsystem, a user interaction subsystem and an optimization subsystem, and specifically includes the following contents: The data collection subsystem collects interface samples, including existing color schemes, interface screenshots and layout structures, and also obtains user preference and behavior data; The color scheme generation subsystem generates candidate color schemes by constructing a color scheme generation model; The intelligent recommendation subsystem recommends personalized color schemes based on the generated candidate color schemes by combining self-supervised heterogeneous graph neural networks and collaborative filtering; The user interaction subsystem provides a visual interface for users to view personalized color schemes in real time, and provides a feedback mechanism to obtain user feedback and usage data; The optimization subsystem optimizes the color scheme generation model according to user feedback and usage data adjustment.

2. The artificial intelligence-based interface color scheme recommendation system according to claim 1, characterized in that: The color scheme generation subsystem builds a color scheme generation model, which specifically includes the following steps: Step S1: Data preparation, the receiving interface collects samples, user preferences and behavior data; Step S2: feature extraction, using two-dimensional feature extraction on the interface collection samples to obtain interface features; Step S3: Build a model, using the GAN model as the basic architecture to build a color matching generation model; Step S4: Model training, using the interface to collect samples, user preferences and behavior data to train the color matching generation model, and repeatedly perform hyperparameter tuning; Step S5: Color scheme generation: using the trained color scheme generation model to generate candidate color schemes.

3. The interface color scheme recommendation system based on artificial intelligence according to claim 2, characterized in that: In step S2, two-dimensional feature extraction is performed on the interface collection samples, specifically including the following steps: Step S21: edge feature extraction, using a filter to convolve the interface collection samples, extracting the edge extraction in the interface collection samples, performing multi-layer convolution processing on the interface collection samples through the ResNet structure, extracting deep features, and generating a multi-level feature map, in which each pixel position represents the feature of a part of the interface collection sample; Step S22: dense connection feature fusion, using a dense connection mechanism between convolutional layers, the output of each layer is passed to the next layer and returned to each subsequent layer, and a multi-scale feature map is generated by reusing deep features; Step S23: spatial channel attention mechanism, for each multi-scale feature map, calculates the attention weight of each pixel position, and weights the channel dimension of the multi-scale feature map to generate a weighted feature map; Step S24: Feature extraction, combining edge features and deep features, performing convolution operations to extract interface features.

4. The interface color scheme recommendation system based on artificial intelligence according to claim 2, characterized in that: Step S4, repeatedly performing hyperparameter tuning, specifically includes the following steps: Step S41: Initialization, initializing the environment, experience replay pool, internal parameters and external hyperparameters. The internal parameters include the parameters of the policy network, the objective function of the policy network and the Lagrange multiplier. The external hyperparameters include the safety threshold and entropy temperature. Step S42: interact with the environment, use the policy network to sample in the environment, obtain batch data of state-action-reward data and store it in the experience replay pool; Step S43: Update internal parameters, sample batch data from the experience replay pool, and optimize the objective function of the policy network by maximizing rewards and entropy regularization; Step S44: Optimize the Lagrange multiplier, combine the main objective and the constraint objective, and use the gradient downward method to update the policy network parameters; Step S45: dynamically adjusting the Lagrange multiplier according to the constraint target; Step S46: Optimize the hyperparameter target, define the hyperparameter regularization target, and optimize the hyperparameter; Step S47: Set the convergence standard, terminate the tuning after the main objective function converges, and output the optimized hyperparameters.

5. The interface color scheme recommendation system based on artificial intelligence according to claim 1, characterized in that: In the intelligent recommendation subsystem, the method of combining self-supervised heterogeneous graph neural network and collaborative filtering specifically includes the following steps: Step Q1: User preference data processing, obtaining the user's historical color selection, click behavior, rating data and user portrait; Step Q2: Construct a user-color matching graph structure, taking each candidate color matching scheme as a node and the user as a node of another category, initializing the feature vector of the node, and the edge weight between the user and the historical color matching selection represents the degree of preference. The edges between the candidate color matching schemes are constructed by similarity; Step Q3: Collect deep features, define the target node as all candidate color schemes, set the domain range to the 1-hop neighbor to the k-hop neighbor of the target node, use the graph neural network to process the features of the neighbors in the neighborhood layer by layer, assign weights to neighbors with different hop numbers, and fuse the features of all layers to form a deep feature representation of the target node; Step Q4: Collect broad features, define the direct domain as the 1-hop neighbors of the target node, use a single-layer graph convolutional network to aggregate the direct neighborhood of the target node, and generate a broad feature representation of the target node; Step Q5: Collaborative contrastive learning, constructing deep view and wide view according to the relationship between the target node and the deep neighbor and the direct neighbor, constructing positive and negative samples of the two features, setting the deep feature representation and the wide feature representation from the same target node as positive samples, and the deep feature representation and the wide feature representation from different target nodes as negative samples, for each pair of positive and negative samples, calculating the contrast loss function and the total loss function, and optimizing the deep feature representation and the wide feature representation; Step Q6: Generate node embedding, fuse the optimized deep feature representation and broad feature representation into the final embedding vector of the node, and normalize it; Step Q7: Collaborative filtering score prediction, using the final embedding vector of the node for score prediction, and using mean square error loss to optimize the score prediction to obtain the predicted score; Step Q8: Recommendation generation, generating a personalized color scheme recommendation list based on the predicted score.

Images