Automatic door product recommendation method and system

Abstract

The invention belongs to the technical field of door product recommendation, and discloses a door product automatic recommendation method and system. The method comprises the following steps: an enterprise server downloads a model and obtains data of real-time door products; the enterprise server performs three-dimensional reconstruction on the plurality of real-time two-dimensional image data by using a three-dimensional reconstruction model, and performs decoration style analysis by using a decoration style analysis model; the enterprise server is used for generating a user portrait by using a user portrait generation model and generating a real-time door product recommendation list by using an automatic recommendation model; and the enterprise server performs three-dimensional simulation according to the target real-time three-dimensional simulation data to obtain and visualize a corresponding target real-time three-dimensional simulation model. The problems that in the prior art, the cost input is large, the communication efficiency is low, customization is lacked, the data processing capacity is insufficient, the recommendation efficiency is low, the accuracy is poor, the practicability is low, and the display effect is not visual enough are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of category product recommendation, and specifically relates to a method and system for automatically recommending category products. Background Art

[0002] With the rapid development of the internet and e-commerce, consumers are faced with an increasing amount of information when choosing products, making their choices more difficult. Recommending products to consumers and helping them make informed choices not only increases sales and profits for merchants but also enhances the consumer shopping experience. Therefore, automated product recommendation has become a key development direction.

[0003] The existing product recommendations have the following defects:

[0004] 1) Existing product recommendations mostly use paper brochures to introduce product categories. This recommendation method involves two costs: brochure design fees and paper product printing fees, which is costly and inconvenient to carry;

[0005] 2) Stores display products mostly based on descriptions and imaginations, and staff coordination and adjustment are frequent. This results in a large number of corresponding personnel in the factory, low communication efficiency, and requires customers to wait for renderings, which is easy to lose customers. The final product may deviate significantly from the customer's imagination, which is easy to cause disputes.

[0006] 3) Existing product recommendation technologies for various categories cannot fully understand users' personalized needs, resulting in recommended products that do not match users' actual preferences. This lack of customized recommendation solutions often leads to a high rate of misrecommendations and an inability to accurately match user needs.

[0007] 4) Existing product category recommendation technologies lack data processing capabilities and are not suitable for large-scale data or three-dimensional data processing, resulting in low recommendation efficiency. Furthermore, they are unable to analyze interior decoration styles, resulting in poor accuracy and low practicality.

[0008] 5) The existing product category recommendation technology mostly recommends product categories that are visualized in two-dimensional images. The display effect is not intuitive enough and cannot provide sufficient reference for users' choices, resulting in low value of the recommendation results. Summary of the Invention

[0009] In order to solve the problems of existing technologies such as high cost investment, low communication efficiency, lack of customization, insufficient data processing capabilities, low recommendation efficiency, poor accuracy, low practicality and less intuitive display effects, the purpose of the present invention is to provide a method and system for automatic recommendation of product categories.

[0010] The technical solution adopted in the present invention is:

[0011] A method for automatically recommending product categories includes the following steps:

[0012] The enterprise server downloads the pre-trained user portrait generation model, 3D reconstruction model, decoration style analysis model, and automatic recommendation model from the cloud data center, and obtains a number of real-time 2D image data and a number of first real-time 3D simulation data of different real-time categories of products provided by merchants;

[0013] The enterprise server uses the 3D reconstruction model to perform 3D reconstruction on the plurality of real-time 2D image data to obtain a plurality of second real-time 3D simulation data, and uses the decoration style analysis model to perform decoration style analysis on the plurality of first and / or second real-time 3D simulation data to obtain a plurality of real-time decoration style analysis results;

[0014] The enterprise server generates user portraits based on the user's real-time basic information and behavior data using a user portrait generation model. Furthermore, based on the real-time decoration style analysis results and the generated real-time user portraits, the enterprise server uses an automatic recommendation model to generate a real-time product recommendation list including a number of real-time product categories.

[0015] The enterprise server extracts the corresponding target real-time three-dimensional simulation data based on the target real-time category products in the real-time category product recommendation list selected by the user, performs three-dimensional simulation, and obtains and visualizes the corresponding target real-time three-dimensional simulation model.

[0016] Furthermore, the enterprise server downloads the pre-trained user portrait generation model, 3D reconstruction model, decoration style analysis model, and automatic recommendation model from the cloud data center, including the following steps:

[0017] The cloud data center collects historical basic information and behavior data of several users, as well as historical two-dimensional image data and historical three-dimensional simulation data of products in different historical categories;

[0018] Performing preprocessing to obtain a number of preprocessed historical basic information and behavior data, a number of preprocessed historical two-dimensional image data, and a number of preprocessed historical three-dimensional simulation data;

[0019] Based on some pre-processed historical basic information and behavioral data, a deep learning algorithm is used to build a user portrait generation model and generate several historical user portraits;

[0020] Based on a number of pre-processed historical two-dimensional image data, a three-dimensional reconstruction model is constructed using a deep learning-based three-dimensional reconstruction algorithm;

[0021] Based on some pre-processed historical 3D simulation data, a deep learning algorithm based on 3D data is used to build a decoration style analysis model and generate some historical decoration style analysis results;

[0022] Based on the analysis results of several historical user portraits and several historical decoration styles, an automatic recommendation model is constructed using a deep learning algorithm based on multimodal fusion;

[0023] The enterprise server connects to the cloud data center and downloads the user portrait generation model, 3D reconstruction model, decoration style analysis model, and automatic recommendation model.

[0024] Furthermore, the user portrait generation model is built based on the RF-MLP algorithm;

[0025] The 3D reconstruction model is built based on the DBN-GANs algorithm;

[0026] The decoration style analysis model is built based on the 3D CNN-DANN-Elman algorithm;

[0027] The automatic recommendation model is built based on the ICPO-GCN-LSTM-Attention-MLP algorithm.

[0028] Furthermore, based on some pre-processed historical basic information and behavioral data, a deep learning algorithm is used to build a user portrait generation model and generate several historical user portraits, including the following steps:

[0029] Dividing a plurality of pre-processed historical basic information and behavior data with real labels into a first model training set and a first model test set;

[0030] Use the RF-MLP algorithm to build an initial user portrait generation model, and input the first model training set to train and optimize the initial user portrait generation model to obtain an optimized user portrait generation model;

[0031] Input the first model test set to perform model testing on the optimized user portrait generation model to obtain the first model accuracy;

[0032] If the accuracy of the first model is greater than the first accuracy threshold, the final user portrait generation model is output and several historical user portraits are generated. Otherwise, the optimization training continues.

[0033] Furthermore, a 3D reconstruction model is constructed based on a plurality of pre-processed historical 2D image data using a 3D reconstruction algorithm based on deep learning, including the following steps:

[0034] Dividing a plurality of pre-processed historical two-dimensional image data with real labels into an adversarial training set, a second model training set, and a second model test set;

[0035] The DBN-GANs algorithm is used to construct an initial 3D reconstruction model. The initial 3D reconstruction model is pre-trained and fine-tuned by inputting several pre-processed historical 2D image data without real labels to obtain a pre-trained 3D reconstruction model.

[0036] Input the adversarial training set to perform adversarial training on the pre-trained 3D reconstruction model to obtain a trained 3D reconstruction model;

[0037] Inputting the second model training set to perform training optimization on the trained three-dimensional reconstruction model to obtain an optimized three-dimensional reconstruction model;

[0038] Input the second model test set to perform model testing on the optimized three-dimensional reconstructed model to obtain the accuracy of the second model;

[0039] If the accuracy of the second model is greater than the second accuracy threshold, the final three-dimensional reconstruction model is output; otherwise, the optimization training continues.

[0040] Furthermore, based on a number of pre-processed historical three-dimensional simulation data, a deep learning algorithm based on three-dimensional data is used to construct a decoration style analysis model and generate a number of historical decoration style analysis results, including the following steps:

[0041] Dividing a plurality of pre-processed historical three-dimensional simulation data with real labels into a third model training set and a third model test set;

[0042] Use the 3D CNN-DANN-Elman algorithm to build an initial decoration style analysis model and set a joint loss function;

[0043] Inputting a number of pre-processed historical three-dimensional simulation data without setting real labels to pre-train the initial decoration style analysis model, thereby obtaining a pre-trained decoration style analysis model;

[0044] Based on the joint loss function, the third model training set is input to train and optimize the pre-trained decoration style analysis model to obtain an optimized decoration style analysis model;

[0045] Input the third model test set to perform model testing on the optimized decoration style analysis model to obtain the accuracy of the third model;

[0046] If the accuracy of the third model is greater than the third accuracy threshold, the final decoration style analysis model is output and several historical decoration style analysis results are generated; otherwise, the optimization training continues.

[0047] Furthermore, based on the analysis results of several historical user portraits and several historical decoration styles, an automatic recommendation model is constructed using a deep learning algorithm based on multimodal fusion, which includes the following steps:

[0048] Associating each historical user portrait with several historical decoration style analysis results to obtain several automatic recommendation samples, and dividing the several automatic recommendation samples with real labels into a fourth model training set and a fourth model test set;

[0049] Use the ICPO-GCN-LSTM-Attention-MLP algorithm to build an initial automatic recommendation model, and adjust the initial network parameters of the initial automatic recommendation model to obtain an adjusted automatic recommendation model;

[0050] Inputting the fourth model training set to train and optimize the adjusted automatic recommendation model to obtain an optimized automatic recommendation model;

[0051] Input the fourth model test set to perform model testing on the optimized automatic recommendation model to obtain the accuracy of the fourth model;

[0052] If the accuracy of the fourth model is greater than the fourth accuracy threshold, the final automatic recommendation model is output; otherwise, the optimization training continues.

[0053] Furthermore, the enterprise server uses the 3D reconstruction model to perform 3D reconstruction on the plurality of real-time 2D image data to obtain a plurality of second real-time 3D simulation data, and uses the decoration style analysis model to perform decoration style analysis on the plurality of first and / or second real-time 3D simulation data to obtain a plurality of real-time decoration style analysis results, including the following steps:

[0054] The enterprise server inputs a plurality of real-time two-dimensional image data into the three-dimensional reconstruction model and extracts the real-time two-dimensional features of each real-time two-dimensional image data;

[0055] Performing three-dimensional reconstruction based on a number of real-time two-dimensional features to obtain a number of second real-time three-dimensional simulation data;

[0056] Inputting a plurality of first and / or second real-time three-dimensional simulation data into a decoration style analysis model to extract a plurality of real-time three-dimensional features;

[0057] Based on a number of real-time three-dimensional features, a decoration style analysis is performed to obtain a number of real-time decoration style analysis results.

[0058] Furthermore, the enterprise server generates a user profile using a user profile generation model based on the user's real-time basic information and behavior data, and generates a real-time category product recommendation list including a real-time category product list based on the real-time decoration style analysis results and the generated real-time user profile using an automatic recommendation model, including the following steps:

[0059] The enterprise server inputs the real-time basic information and behavior data into the user portrait generation model, screens the real-time basic information and behavior data for key features, and obtains several real-time key features;

[0060] Generate user portraits based on several real-time key features to obtain real-time user portraits of users;

[0061] Input the user's real-time user portrait and the real-time decoration style analysis results of several real-time categories of products into the automatic recommendation model;

[0062] Extract real-time graph features of real-time user portraits and real-time data features of several real-time decoration style analysis results;

[0063] According to the preset attention weight value, the real-time image features and real-time data features are weightedly fused to obtain real-time fusion features;

[0064] According to the real-time fusion features, recommendation scores are made to obtain the real-time recommendation scores of each real-time category product;

[0065] Several real-time category products whose real-time recommendation scores exceed a score threshold are sorted in descending order to generate a real-time category product recommendation list including the several real-time category products.

[0066] A system for automatically recommending product categories is used to implement a method for automatically recommending product categories. The system includes a cloud data center and an enterprise server. The cloud data center is communicatively connected to the enterprise server. The enterprise server includes a model download and data confirmation unit, a three-dimensional reconstruction and decoration style analysis unit, a user portrait generation and automatic recommendation unit, and a three-dimensional simulation and visualization unit, which are connected in sequence.

[0067] The beneficial effects of the present invention are:

[0068] The present invention discloses a method and system for automatically recommending category products, which adopts digital and information-based category product display and recommendation, avoids the use of traditional paper product albums, and reduces cost investment; by intuitively displaying category products for customers to choose, and sorting out orders according to customers' real-time opinions, communication efficiency is improved, there is no need to wait, the customer experience is improved, and finished product deviation is avoided; by constructing a user portrait generation model, the user's basic information, behavior data and preferences are comprehensively analyzed to achieve highly personalized category product recommendations, meet the user's personalized needs, and improve the recommendation accuracy; the constructed three-dimensional reconstruction model, decoration style analysis model and automatic recommendation model can realize automated and intelligent three-dimensional reconstruction, decoration style analysis and automatic recommendation, provide powerful data processing capabilities, reduce manual intervention, improve recommendation efficiency and stability, and are suitable for large-scale data; the decoration style analysis model takes into account the adaptability of category products to decoration styles, improves the accuracy, practicality and user experience of recommendations; the three-dimensional reconstruction model converts the two-dimensional image of the category product into three-dimensional simulation data, realizes the conversion from two-dimensional to three-dimensional, enhances the display effect, provides sufficient reference effects for the user's choice, and improves the value of the recommendation results.

[0069] Other beneficial effects of the present invention will be further described in the specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a flowchart of the automatic recommendation method for category products in the present invention.

[0071] Figure 2 It is a structural block diagram of the automatic recommendation system for category products in the present invention. DETAILED DESCRIPTION

[0072] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0073] Example 1:

[0074] like Figure 1 As shown, this embodiment provides a method for automatically recommending product categories, including the following steps:

[0075] S1: The enterprise server downloads the pre-trained user portrait generation model, 3D reconstruction model, decoration style analysis model, and automatic recommendation model from the cloud data center, and obtains a number of real-time 2D image data and a number of first real-time 3D simulation data of different real-time categories of products provided by the merchant, including the following steps:

[0076] S1-1: Cloud data center, which collects historical basic information and behavior data of several users, as well as historical two-dimensional image data and historical three-dimensional simulation data of products in different historical categories;

[0077] S1-2: performing preprocessing to obtain a number of preprocessed historical basic information and behavior data, a number of preprocessed historical two-dimensional image data, and a number of preprocessed historical three-dimensional simulation data;

[0078] Preprocessing of basic information and behavioral data includes data cleaning, format conversion, and normalization. This eliminates duplicate and erroneous data while performing standardized conversions to eliminate magnitude differences and noise, and converting the data into a format recognizable by subsequent models. Preprocessing of 2D image data also includes image enhancement and resizing to improve image quality. Preprocessing of 3D simulation data also includes meshing and boundary condition definition to improve 3D data quality.

[0079] S1-3: Based on some pre-processed historical basic information and behavioral data, use deep learning algorithms to build a user profile generation model and generate several historical user profiles;

[0080] The user portrait generation model is built based on the Random Forest (RF)-Multilayer Perceptron (MLP) algorithm, and includes a key feature screening module built based on the RF algorithm and a user portrait generation module built based on the MLP algorithm, which are connected in sequence;

[0081] The key feature screening module uses an internal Classification and Regression Tree (CART) to screen the feature components of the input basic information and behavioral data. It can process a large number of feature components, generate a key feature importance score for each feature component, and select the most stable and discriminative feature components based on the key feature importance score. After the key feature screening module has been trained, it can directly screen the newly input basic information and behavioral data based on the selected key features. The user profile generation module, as a fully connected network, can predict user profile labels based on the screened key features.

[0082] Based on some pre-processed historical basic information and behavioral data, a deep learning algorithm is used to build a user profile generation model and generate several historical user profiles, including the following steps:

[0083] S1-3-1: Divide the pre-processed historical basic information and behavior data with real labels into a first model training set and a first model test set;

[0084] Basic information and behavioral data include basic information of users (age, occupation, gender, etc.), behavioral data (purchase history, browsing history, click history) and preferences;

[0085] S1-3-2: Use the RF-MLP algorithm to build an initial user portrait generation model, and input the first model training set to train and optimize the initial user portrait generation model to obtain an optimized user portrait generation model;

[0086] S1-3-3: Input the first model test set to perform model testing on the optimized user portrait generation model to obtain the first model accuracy;

[0087] S1-3-4: If the accuracy of the first model is greater than the first accuracy threshold, the final user profile generation model is output and several historical user profiles are generated. Otherwise, optimization training continues;

[0088] S1-4: Based on some pre-processed historical 2D image data, a 3D reconstruction model is constructed using a deep learning-based 3D reconstruction algorithm;

[0089] The 3D reconstruction model is built based on the Deep Belief Nets (DBN)-Generative Adversarial Networks (GANs) algorithm, and includes a two-dimensional feature extraction module based on the DBN algorithm and a three-dimensional data generation module based on the GANs algorithm.

[0090] DBN is a generative model that can learn deep representations of two-dimensional data for extracting two-dimensional features. GAN consists of a generator and a discriminator. The generator attempts to generate realistic three-dimensional simulation data, while the discriminator attempts to distinguish the generated three-dimensional simulation data from the real three-dimensional simulation data. The generator and the discriminator are trained alternately. The generator tries to deceive the discriminator, while the discriminator tries not to be deceived by the generator.

[0091] Based on some pre-processed historical 2D image data, a 3D reconstruction model is constructed using a deep learning-based 3D reconstruction algorithm, which includes the following steps:

[0092] S1-4-1: Divide a plurality of pre-processed historical two-dimensional image data with real labels into an adversarial training set, a second model training set, and a second model test set;

[0093] S1-4-2: Use the DBN-GANs algorithm to build an initial 3D reconstruction model. Input a number of pre-processed historical 2D image data without real labels to pre-train and fine-tune the 2D feature extraction module of the initial 3D reconstruction model, thereby obtaining a pre-trained 2D feature extraction module.

[0094] S1-4-3: Input the adversarial training set to perform adversarial training on the 3D data generation module of the initial 3D reconstruction model to obtain an adversarially trained 3D data generation module, and combine it with the pre-trained 2D feature extraction module to obtain a trained 3D reconstruction model;

[0095] S1-4-4: inputting the second model training set to train and optimize the trained 3D reconstruction model to obtain an optimized 3D reconstruction model;

[0096] S1-4-5: input the second model test set to perform model testing on the optimized 3D reconstructed model to obtain the second model accuracy;

[0097] S1-4-6: If the accuracy of the second model is greater than the second accuracy threshold, the final 3D reconstruction model is output; otherwise, optimization training is continued;

[0098] S1-5: Based on a number of pre-processed historical 3D simulation data, a deep learning algorithm based on 3D data is used to construct a decoration style analysis model and generate a number of historical decoration style analysis results;

[0099] The decoration style analysis model is built based on the 3D Convolutional Neural Network (3D CNN)-DANN-Elman algorithm, and includes a 3D feature extraction module based on the 3D CNN algorithm, a domain adversarial module based on the Domain-Adversarial Neural Network (DANN) algorithm, and a decoration style analysis module based on the Elman algorithm, which are connected in sequence.

[0100] The domain adversarial module includes a gradient reversal layer and a domain classification layer connected in sequence. The three-dimensional feature extraction module, the gradient reversal layer and the decoration style analysis module are connected in sequence. The three-dimensional feature extraction module is used to extract the three-dimensional features of the three-dimensional simulation data. The gradient reversal layer is located between the feature extractor and the domain classifier. Its function is to reverse the gradient during the back propagation process, so that the domain classifier can learn to distinguish the features of the source domain and the target domain during the training process, while the feature extractor is trained to generate domain-invariant features, thereby realizing adversarial training. The domain classification layer is used to predict whether the feature belongs to the source domain or the target domain based on the inverted features output by the gradient reversal layer. During the training process, the three-dimensional feature extraction module attempts to generate domain-invariant features, while the domain classifier attempts to distinguish which domain these features come from. Through this adversarial training, the three-dimensional feature extraction module gradually learns to ignore domain-related information and only retains task-related information. Through multiple iterative training, the model gradually learns to accurately classify in the source domain while extracting features that also have good generalization ability in the target domain. The decoration style analysis module predicts the decoration style based on the three-dimensional features extracted by the three-dimensional feature extraction module.

[0101] Based on some pre-processed historical 3D simulation data, a deep learning algorithm based on 3D data is used to build a decoration style analysis model and generate some historical decoration style analysis results, including the following steps:

[0102] S1-5-1: Divide a plurality of pre-processed historical three-dimensional simulation data with real labels into a third model training set and a third model test set;

[0103] S1-5-2: Use the 3D CNN-DANN-Elman algorithm to build an initial decoration style analysis model. The first loss function of the 3D feature extraction module of the initial decoration style analysis model and the second loss function of the domain adversarial module are weighted together to obtain a joint loss function.

[0104] S1-5-3: Inputting a number of pre-processed historical three-dimensional simulation data without real labels to pre-train the initial decoration style analysis model, thereby obtaining a pre-trained decoration style analysis model;

[0105] S1-5-4: Based on the joint loss function, input the third model training set to train and optimize the pre-trained decoration style analysis model to obtain an optimized decoration style analysis model;

[0106] S1-5-5: Input the third model test set to perform model testing on the optimized decoration style analysis model to obtain the accuracy of the third model;

[0107] S1-5-6: If the accuracy of the third model is greater than the third accuracy threshold, the final decoration style analysis model is output and several historical decoration style analysis results are generated; otherwise, the optimization training is continued;

[0108] The decoration style analysis results include overall style prediction results, color matching prediction results, material and texture prediction results, decorative element prediction results, etc.

[0109] S1-6: Based on the analysis results of several historical user portraits and several historical decoration styles, an automatic recommendation model is constructed using a deep learning algorithm based on multimodal fusion;

[0110] The automatic recommendation model is based on the Improved Crested Porcupine Optimizer (ICPO)-Graph Convolutional Network (GCN)-Long Short-Term Memory (LSTM)-Attention-MLP algorithm. The automatic recommendation model includes a network parameter optimization module based on the ICPO algorithm, a graph feature extraction module based on the GCN algorithm, a data feature extraction module based on the LSTM algorithm, an attention weight module based on the Attention mechanism, and a recommendation scoring module based on the MLP algorithm. The network parameter optimization module is connected to the graph feature extraction module and the data feature extraction module, respectively. The graph feature extraction module and the data feature extraction module are both connected to the attention weight module, and the attention weight module is connected to the recommendation scoring module.

[0111] The ICPO algorithm optimizes the initial network parameters of the model, preventing the GCN network and LSTM network from falling into local optimal values ​​and being unable to jump out, reducing the sensitivity to the initial values ​​of the network, and improving the robustness of the model, the efficiency and accuracy of feature extraction; the GCN network propagates features on the user portrait through operations similar to convolution, extracts the node features of the user portrait, the edge features between nodes, and constructs the graph features of the user portrait; the LSTM network of the data feature extraction module can learn to retain important information in long time series and ignore unimportant information through the gate mechanism, which makes the LSTM network have significant advantages in processing sequence data. The LSTM network can capture long-term dependencies in the data and extract data features in the decoration style analysis results; the attention weight module automatically assigns different levels of attention to different parts of the input feature sequence by setting the attention weight value, which significantly improves the model's ability to process multimodal long sequence features because it allows the model to focus on the most relevant part of the input feature sequence when predicting each output; the recommendation scoring module, as a fully connected network, can recommend score predictions based on weighted fusion features;

[0112] Based on the analysis results of several historical user portraits and several historical decoration styles, we use a deep learning algorithm based on multimodal fusion to build an automatic recommendation model, which includes the following steps:

[0113] S1-6-1: Associating each historical user profile with several historical decoration style analysis results to obtain several automatic recommendation samples, and dividing the several automatic recommendation samples with real labels into a fourth model training set and a fourth model test set;

[0114] S1-6-2: Use the ICPO-GCN-LSTM-Attention-MLP algorithm to build an initial automatic recommendation model, and use the network parameter optimization module to adjust the initial network parameters of the initial automatic recommendation model with the goal of minimizing the mean square error. This results in an adjusted automatic recommendation model, including the following steps:

[0115] S1-6-2-1: Use the ICPO-GCN-LSTM-Attention-MLP algorithm to build the initial automatic recommendation model; the automatic recommendation model includes a network parameter optimization module, a graph feature extraction module, a data feature extraction module, an attention weight module, and a recommendation scoring module;

[0116] S1-6-2-2: Set the initial network parameters of the graph feature extraction module and the data feature extraction module to the individual vector encoding format of the parameter optimization model;

[0117] S1-6-2-3: Take minimizing the mean square error as the optimization goal, and according to the optimization goal, set the fitness function of the network parameter optimization module, and set the ICPO population parameters and the maximum number of iterations of the network parameter optimization module;

[0118] The formula is:

[0119]

[0120] Where, fit is the fitness function; MSE is the mean square error function; Y n' is the true value of the n'th sample; is the predicted value of the n'th sample; n' is the sample indicator; N is the total number of samples;

[0121] S1-6-2-4: Based on the individual vector encoding format and ICPO population parameters, the Circle chaotic mapping sequence is used to initialize the ICPO population to obtain an initial ICPO population containing several initial ICPO individuals;

[0122] The formula is:

[0123]

[0124] Where, is the initial ICPO individual of the Circle chaos map, corresponding to an initial solution; is the randomly generated initial ICPO individual; i' is the ICPO individual indicator;

[0125] S1-6-2-5: Introduce a cyclic population reduction mechanism to limit the number of individuals in the ICPO population parameters and obtain the updated ICPO population parameters for the next iteration;

[0126] The formula is:

[0127]

[0128] Where S t+1 is the number of individuals in the ICPO population parameter of the t+1th iteration; S t is the number of individuals in the ICPO population parameter of the tth iteration; S min is the minimum number of individuals in the ICPO population parameter; a'% is the function evaluation parameter; V is the function evaluation cycle parameter; V max is the maximum function evaluation loop parameter; t is the number of iterations indicator;

[0129] S1-6-2-6: Calculate the initial fitness value of the initial ICPO individuals in the initial ICPO population according to the fitness function;

[0130] S1-6-2-7: Based on the initial fitness value and the updated ICPO population parameters, the first defense strategy, the second defense strategy, the third defense strategy, and the fourth defense strategy are used to update the initial ICPO population to obtain an updated ICPO population;

[0131] The formula for the first defense strategy is:

[0132]

[0133] Where, For the updated ICPO individuals within the first defense range; is the initial ICPO individual within the first defense range; τ1 is a random number based on normal distribution; τ2 is a random value in the interval [0,1]; It is the optimal solution within the first defense range; is the vector generated between the true optimal solution within the first defense range and the optimal solution randomly selected from the ICPO population; i' is the ICPO individual indicator; t is the iteration indicator;

[0134] The formula for the second defense strategy is:

[0135]

[0136] Where, For the updated ICPO individuals within the second defense range; The initial ICPO individual within the second defense range; is the search upper limit vector of the second defense range; τ3 is a random value in the interval [0,1]; are the r1th and r2th initial ICPO individuals respectively; r1 and r2 are two random integers between [1, S]; is the vector generated between the true optimal solution within the second defense range and the optimal solution randomly selected from the ICPO population;

[0137] The formula for the third defense strategy is:

[0138]

[0139] Where, For the updated ICPO individuals within the third defense range; The initial ICPO individual within the third defense range; is the search upper limit vector of the third defense range; are the r2th and r3th initial ICPO individuals respectively; r3 is a random integer between [1, S]; The odor diffusion factor defined for the fitness function; λ t It is a defense factor; Control parameters for search direction;

[0140] The formula for the fourth defense strategy is:

[0141]

[0142] Where, For the updated ICPO individuals within the fourth defense range; For the initial ICPO individuals within the fourth defense range; is the optimal solution within the fourth defense range; τ4 and τ5 are both random values ​​in the interval [0,1]; λ t It is a defense factor; Control parameters for search direction; is the average force affecting the search direction; a' is the convergence speed factor;

[0143] S1-6-2-8: Use the dynamic reverse learning algorithm to perform dynamic reverse learning on the updated ICPO population to generate a dynamic reverse ICPO population;

[0144] The formula is:

[0145]

[0146] Where, is the ICPO individual in dynamic reverse direction; γ' is the decreasing inertia coefficient; L max 、L min are the maximum and minimum values ​​of the vector space respectively; For the updated ICPO individual;

[0147] S1-6-2-9: According to the fitness function, calculate the fitness values ​​of all ICPO individuals in the updated ICPO population and the dynamically reversed ICPO population, take the ICPO individual with the minimum fitness value as the optimal individual, and retain the optimal individual;

[0148] S1-6-2-10: If the number of iterations of iterative optimization reaches the maximum number of iterations or the fitness value of the optimal individual meets the requirements, the optimal solution corresponding to the optimal individual is output;

[0149] S1-6-2-11: Decode the individual vector of the optimal solution to generate optimal initial network parameters, and adjust the initial automatic recommendation model based on the optimal initial network parameters to obtain an adjusted automatic recommendation model;

[0150] S1-6-3: Input the fourth model training set to train and optimize the adjusted automatic recommendation model to obtain an optimized automatic recommendation model;

[0151] S1-6-4: Input the fourth model test set to perform model testing on the optimized automatic recommendation model to obtain the fourth model accuracy;

[0152] S1-6-5: If the accuracy of the fourth model is greater than the fourth accuracy threshold, output the final automatic recommendation model; otherwise, continue optimization training;

[0153] S1-7: Enterprise server, connected to the cloud data center, and downloads the user portrait generation model, 3D reconstruction model, decoration style analysis model, and automatic recommendation model;

[0154] S1-8: Acquire a plurality of real-time two-dimensional image data and a plurality of first real-time three-dimensional simulation data of different real-time categories of products provided by merchants;

[0155] S2: The enterprise server uses the 3D reconstruction model to perform 3D reconstruction on the plurality of real-time 2D image data to obtain a plurality of second real-time 3D simulation data, and uses the decoration style analysis model to perform decoration style analysis on the plurality of first and / or second real-time 3D simulation data to obtain a plurality of real-time decoration style analysis results, including the following steps:

[0156] S2-1: The enterprise server inputs a number of real-time 2D image data into the 3D reconstruction model and uses the 2D feature extraction module of the 3D reconstruction model to extract the real-time 2D features of each real-time 2D image data;

[0157] S2-2: performing 3D reconstruction using a 3D data generation module of a 3D reconstruction model based on the plurality of real-time 2D features to obtain a plurality of second real-time 3D simulation data;

[0158] S2-3: Inputting the plurality of first and / or second real-time 3D simulation data into the decoration style analysis model, and extracting a plurality of real-time 3D features using the 3D feature extraction module of the decoration style analysis model;

[0159] S2-4: performing decoration style analysis using a decoration style analysis module of a decoration style analysis model based on the multiple real-time three-dimensional features to obtain multiple real-time decoration style analysis results;

[0160] S3: The enterprise server generates user profiles based on the user's real-time basic information and behavior data using a user profile generation model. Based on the real-time decoration style analysis results and the generated real-time user profiles, it uses an automatic recommendation model to generate a real-time product recommendation list that includes several real-time product categories. This includes the following steps:

[0161] S3-1: The enterprise server inputs the real-time basic information and behavior data into the user profile generation model. The key feature screening module of the user profile generation model is used to screen the real-time basic information and behavior data for key features, thereby obtaining several real-time key features.

[0162] S3-2: Based on several real-time key features, use the user portrait generation module of the user portrait generation model to generate a user portrait to obtain a real-time user portrait of the user;

[0163] S3-3: Input the user's real-time user profile and the real-time decoration style analysis results of several real-time categories of products into the automatic recommendation model;

[0164] S3-4: Use the graph feature extraction module of the automatic recommendation model to extract real-time graph features of real-time user portraits. Use the data feature extraction module of the automatic recommendation model to extract real-time data features of several real-time decoration style analysis results.

[0165] S3-5: Based on the preset attention weight value, the attention weight module of the automatic recommendation model is used to perform weighted fusion on the real-time graph features and real-time data features to obtain real-time fusion features;

[0166] S3-6: Based on the real-time fusion features, the recommendation scoring module of the automatic recommendation model is used to perform recommendation scoring to obtain the real-time recommendation score for each real-time category product;

[0167] S3-7: sorting the real-time category products whose real-time recommendation scores exceed the score threshold in descending order to generate a real-time category product recommendation list including the real-time category products;

[0168] S4: The enterprise server extracts the corresponding target real-time three-dimensional simulation data based on the target real-time category products in the real-time category product recommendation list selected by the user, performs three-dimensional simulation, and obtains and visualizes the corresponding target real-time three-dimensional simulation model.

[0169] Example 2:

[0170] like Figure 2 As shown, this embodiment provides an automatic recommendation system for product categories, which is used to implement an automatic recommendation method for product categories. The system includes a cloud data center and an enterprise server. The cloud data center is communicatively connected to the enterprise server. The enterprise server includes a model download and data confirmation unit, a 3D reconstruction and decoration style analysis unit, a user portrait generation and automatic recommendation unit, and a 3D simulation and visualization unit connected in sequence.

[0171] Cloud data center, used to build user portrait generation models, 3D reconstruction models, decoration style analysis models, and automatic recommendation models using deep learning algorithms;

[0172] A model downloading and data confirmation unit is used to download the pre-trained user portrait generation model, 3D reconstruction model, decoration style analysis model, and automatic recommendation model in the cloud data center, and obtain a number of real-time 2D image data and a number of first real-time 3D simulation data of different real-time categories of products provided by merchants;

[0173] a 3D reconstruction and decoration style analysis unit, configured to use the 3D reconstruction model to perform 3D reconstruction on a plurality of real-time 2D image data to obtain a plurality of second real-time 3D simulation data, and use the decoration style analysis model to perform decoration style analysis on a plurality of first and / or second real-time 3D simulation data to obtain a plurality of real-time decoration style analysis results;

[0174] A user portrait generation and automatic recommendation unit is used to generate a user portrait based on the user's real-time basic information and behavior data using a user portrait generation model, and to generate a real-time product category recommendation list including a number of real-time product categories based on a number of real-time decoration style analysis results and the generated real-time user portrait using an automatic recommendation model;

[0175] The three-dimensional simulation and visualization unit is used to extract the corresponding target real-time three-dimensional simulation data according to the target real-time category products in the real-time category product recommendation list selected by the user, perform three-dimensional simulation, and obtain and visualize the corresponding target real-time three-dimensional simulation model.

[0176] The present invention discloses a method and system for automatically recommending category products, which adopts digital and information-based category product display and recommendation, avoids the use of traditional paper product albums, and reduces cost investment; by intuitively displaying category products for customers to choose, and sorting out orders according to customers' real-time opinions, communication efficiency is improved, there is no need to wait, the customer experience is improved, and finished product deviation is avoided; by constructing a user portrait generation model, the user's basic information, behavior data and preferences are comprehensively analyzed to achieve highly personalized category product recommendations, meet the user's personalized needs, and improve the recommendation accuracy; the constructed three-dimensional reconstruction model, decoration style analysis model and automatic recommendation model can realize automated and intelligent three-dimensional reconstruction, decoration style analysis and automatic recommendation, provide powerful data processing capabilities, reduce manual intervention, improve recommendation efficiency and stability, and are suitable for large-scale data; the decoration style analysis model takes into account the adaptability of category products to decoration styles, improves the accuracy, practicality and user experience of recommendations; the three-dimensional reconstruction model converts the two-dimensional image of the category product into three-dimensional simulation data, realizes the conversion from two-dimensional to three-dimensional, enhances the display effect, provides sufficient reference effects for the user's choice, and improves the value of the recommendation results.

[0177] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for automatically recommending product categories, characterized by: The steps include: The enterprise server downloads the pre-trained user portrait generation model, 3D reconstruction model, decoration style analysis model, and automatic recommendation model from the cloud data center, and obtains a number of real-time 2D image data and a number of first real-time 3D simulation data of different real-time categories of products provided by merchants; The enterprise server uses the 3D reconstruction model to perform 3D reconstruction on the plurality of real-time 2D image data to obtain a plurality of second real-time 3D simulation data, and uses the decoration style analysis model to perform decoration style analysis on the plurality of first and / or second real-time 3D simulation data to obtain a plurality of real-time decoration style analysis results; The enterprise server generates user portraits based on the user's real-time basic information and behavior data using a user portrait generation model. Furthermore, based on the real-time decoration style analysis results and the generated real-time user portraits, the enterprise server uses an automatic recommendation model to generate a real-time product recommendation list including a number of real-time product categories. The enterprise server extracts the corresponding target real-time three-dimensional simulation data based on the target real-time category products in the real-time category product recommendation list selected by the user, performs three-dimensional simulation, and obtains and visualizes the corresponding target real-time three-dimensional simulation model.

2. The method for automatically recommending door products according to claim 1, characterized in that: The enterprise server downloads the pre-trained user portrait generation model, 3D reconstruction model, decoration style analysis model, and automatic recommendation model from the cloud data center, including the following steps: The cloud data center collects historical basic information and behavior data of several users, as well as historical two-dimensional image data and historical three-dimensional simulation data of products in different historical categories; Performing preprocessing to obtain a number of preprocessed historical basic information and behavior data, a number of preprocessed historical two-dimensional image data, and a number of preprocessed historical three-dimensional simulation data; Based on some pre-processed historical basic information and behavioral data, a deep learning algorithm is used to build a user portrait generation model and generate several historical user portraits; Based on a number of pre-processed historical two-dimensional image data, a three-dimensional reconstruction model is constructed using a deep learning-based three-dimensional reconstruction algorithm; Based on some pre-processed historical 3D simulation data, a deep learning algorithm based on 3D data is used to build a decoration style analysis model and generate some historical decoration style analysis results; Based on the analysis results of several historical user portraits and several historical decoration styles, an automatic recommendation model is constructed using a deep learning algorithm based on multimodal fusion; The enterprise server connects to the cloud data center and downloads the user portrait generation model, 3D reconstruction model, decoration style analysis model, and automatic recommendation model.

3. The method for automatically recommending door products according to claim 2, characterized in that: The user portrait generation model is constructed based on the RF-MLP algorithm; The three-dimensional reconstruction model is constructed based on the DBN-GANs algorithm; The decoration style analysis model is constructed based on the 3D CNN-DANN-Elman algorithm; The automatic recommendation model is built based on the ICPO-GCN-LSTM-Attention-MLP algorithm.

4. The method for automatically recommending door products according to claim 3, characterized in that: Based on some pre-processed historical basic information and behavioral data, a deep learning algorithm is used to build a user profile generation model and generate several historical user profiles, including the following steps: Dividing a plurality of pre-processed historical basic information and behavior data with real labels into a first model training set and a first model test set; Use the RF-MLP algorithm to build an initial user portrait generation model, and input the first model training set to train and optimize the initial user portrait generation model to obtain an optimized user portrait generation model; Input the first model test set to perform model testing on the optimized user portrait generation model to obtain the first model accuracy; If the accuracy of the first model is greater than the first accuracy threshold, the final user portrait generation model is output and several historical user portraits are generated. Otherwise, the optimization training continues.

5. The method for automatically recommending door products according to claim 3, characterized in that: Based on some pre-processed historical 2D image data, a 3D reconstruction model is constructed using a deep learning-based 3D reconstruction algorithm, which includes the following steps: Dividing a plurality of pre-processed historical two-dimensional image data with real labels into an adversarial training set, a second model training set, and a second model test set; The DBN-GANs algorithm is used to construct an initial 3D reconstruction model. The initial 3D reconstruction model is pre-trained and fine-tuned by inputting several pre-processed historical 2D image data without real labels to obtain a pre-trained 3D reconstruction model. Input the adversarial training set to perform adversarial training on the pre-trained 3D reconstruction model to obtain a trained 3D reconstruction model; Inputting the second model training set to perform training optimization on the trained three-dimensional reconstruction model to obtain an optimized three-dimensional reconstruction model; Input the second model test set to perform model testing on the optimized three-dimensional reconstructed model to obtain the accuracy of the second model; If the accuracy of the second model is greater than the second accuracy threshold, the final three-dimensional reconstruction model is output; otherwise, the optimization training continues.

6. The method for automatically recommending door products according to claim 5, characterized in that: Based on some pre-processed historical 3D simulation data, a deep learning algorithm based on 3D data is used to build a decoration style analysis model and generate some historical decoration style analysis results, including the following steps: Dividing a plurality of pre-processed historical three-dimensional simulation data with real labels into a third model training set and a third model test set; Use the 3D CNN-DANN-Elman algorithm to build an initial decoration style analysis model and set a joint loss function; Inputting a number of pre-processed historical three-dimensional simulation data without setting real labels to pre-train the initial decoration style analysis model, thereby obtaining a pre-trained decoration style analysis model; Based on the joint loss function, the third model training set is input to train and optimize the pre-trained decoration style analysis model to obtain an optimized decoration style analysis model; Input the third model test set to perform model testing on the optimized decoration style analysis model to obtain the accuracy of the third model; If the accuracy of the third model is greater than the third accuracy threshold, the final decoration style analysis model is output and several historical decoration style analysis results are generated; otherwise, the optimization training continues.

7. The method for automatically recommending door products according to claim 4, characterized in that: Based on the analysis results of several historical user portraits and several historical decoration styles, we use a deep learning algorithm based on multimodal fusion to build an automatic recommendation model, which includes the following steps: Associating each historical user portrait with several historical decoration style analysis results to obtain several automatic recommendation samples, and dividing the several automatic recommendation samples with real labels into a fourth model training set and a fourth model test set; Use the ICPO-GCN-LSTM-Attention-MLP algorithm to build an initial automatic recommendation model, and adjust the initial network parameters of the initial automatic recommendation model to obtain an adjusted automatic recommendation model; Inputting the fourth model training set to train and optimize the adjusted automatic recommendation model to obtain an optimized automatic recommendation model; Input the fourth model test set to perform model testing on the optimized automatic recommendation model to obtain the accuracy of the fourth model; If the accuracy of the fourth model is greater than the fourth accuracy threshold, the final automatic recommendation model is output; otherwise, the optimization training continues.

8. The method for automatically recommending door products according to claim 6, characterized in that: The enterprise server uses the 3D reconstruction model to perform 3D reconstruction on the plurality of real-time 2D image data to obtain a plurality of second real-time 3D simulation data, and uses the decoration style analysis model to perform decoration style analysis on the plurality of first and / or second real-time 3D simulation data to obtain a plurality of real-time decoration style analysis results, including the following steps: The enterprise server inputs a plurality of real-time two-dimensional image data into the three-dimensional reconstruction model and extracts the real-time two-dimensional features of each real-time two-dimensional image data; Performing three-dimensional reconstruction based on a number of real-time two-dimensional features to obtain a number of second real-time three-dimensional simulation data; Inputting a plurality of first and / or second real-time three-dimensional simulation data into a decoration style analysis model to extract a plurality of real-time three-dimensional features; Based on a number of real-time three-dimensional features, a decoration style analysis is performed to obtain a number of real-time decoration style analysis results.

9. The method for automatically recommending door products according to claim 7, characterized in that: The enterprise server generates user portraits based on the user's real-time basic information and behavior data using a user portrait generation model. Based on the real-time decoration style analysis results and the generated real-time user portraits, the enterprise server uses an automatic recommendation model to generate a real-time product recommendation list that includes a number of real-time product categories. This includes the following steps: The enterprise server inputs the real-time basic information and behavior data into the user portrait generation model, screens the real-time basic information and behavior data for key features, and obtains several real-time key features; Generate user portraits based on several real-time key features to obtain real-time user portraits of users; Input the user's real-time user portrait and the real-time decoration style analysis results of several real-time categories of products into the automatic recommendation model; Extract real-time graph features of real-time user portraits and real-time data features of several real-time decoration style analysis results; According to the preset attention weight value, the real-time image features and real-time data features are weightedly fused to obtain real-time fusion features; According to the real-time fusion features, recommendation scores are made to obtain the real-time recommendation scores of each real-time category product; Several real-time category products whose real-time recommendation scores exceed a score threshold are sorted in descending order to generate a real-time category product recommendation list including the several real-time category products.

10. An automatic door product recommendation system, used to implement the automatic door product recommendation method according to any one of claims 1 to 9, characterized in that: The system includes a cloud data center and an enterprise server. The cloud data center is communicatively connected to the enterprise server. The enterprise server includes a model download and data confirmation unit, a three-dimensional reconstruction and decoration style analysis unit, a user portrait generation and automatic recommendation unit, and a three-dimensional simulation and visualization unit, which are connected in sequence.

Images