E-commerce personalized recommendation data system fused with deep learning

Abstract

The invention relates to the technical field of new user data recommendation, in particular to an e-commerce personalized recommendation data system fused with deep learning, which comprises a data acquisition and processing unit, a dynamic monitoring unit, a GAN data enhancement unit and a recommendation service unit. The multi-modal data of users and commodities is comprehensively collected, features are accurately extracted, the dynamic monitoring unit monitors and predicts data changes in real time and generates adjustment signals by applying quantum state space and causal inference network technologies, and the GAN data enhancement unit generates high-quality simulation data based on a deep convolutional generative adversarial network by means of meta-learning, knowledge distillation and other algorithms. And the recommendation service unit is used for training a model according to the enhanced data, so that accurate recommendation is realized, the exposure degree of new commodities is improved, and the accuracy and efficiency of recommendation are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of new user data recommendation, and more specifically, to an e-commerce personalized recommendation data system integrating deep learning. Background Art

[0002] New user data recommendation is an important technology. With the explosive growth of the variety of goods and the number of users on e-commerce platforms, users are faced with the problem of information overload, while merchants also expect new products to quickly find target customers. The solution to the cold start problem is imminent.

[0003] Traditional recommendation systems mainly rely on users' historical behavior data and product interaction records for recommendation. When a new user first visits the platform, due to the lack of their historical behavior data, it is difficult to accurately grasp their interest preferences, and only broad and non-targeted recommendations can be provided, resulting in poor user experience and an increased risk of user loss. For newly listed products, due to the lack of sufficient user browsing and purchase data, it is difficult to obtain exposure opportunities in the recommendation system, affecting the promotion and sales of products and causing waste of merchant resources. To solve this technical problem, we provide an e-commerce personalized recommendation data system integrating deep learning. Summary of the Invention

[0004] The purpose of the present invention is to provide an e-commerce personalized recommendation data system integrating deep learning to solve the problems raised in the above background art.

[0005] To achieve the above purpose, an e-commerce personalized recommendation data system integrating deep learning is provided, including a data acquisition and processing unit, a dynamic monitoring unit, a GAN data enhancement unit, and a recommendation service unit;

[0006] The data acquisition and processing unit collects the behavior data of new and old users, the multi-modal data of products, and the operation data from the e-commerce platform logs to form a multi-modal data set, and extracts data features from the multi-modal data set;

[0007] The dynamic monitoring unit sets user behavior characteristics, product characteristics, and platform operation indicators as data features, uses time series analysis to monitor and predict each indicator in real time, judges the change range of data distribution by comparing the actual value with the predicted value, and generates an adjustment signal according to the change range of data distribution, and applies the adjustment signal to the GAN data enhancement unit to determine the adjustment direction and amplitude of the training parameters of the generator and discriminator;

[0008] The GAN data augmentation unit adopts a deep convolutional generative adversarial network. The generator receives a random noise vector and outputs simulated user-item interaction data. The discriminator receives real and generated data and outputs the probability that the data is real data. It modifies the weights of the generator loss function according to the adjustment signal, generates simulated data that conforms to the data change trend, and finally adjusts the regularization parameters of the discriminator to adapt to the data distribution change. The generator and discriminator parameters are alternately updated using real data and the adjusted generated data for training until convergence;

[0009] When the recommendation service unit inputs new user data, the GAN model generates a personalized recommendation list based on their behavior data and collects feedback data from the new user on the recommendation results to retrain the GAN model.

[0010] As a further improvement of this technical solution, when the data acquisition and processing unit collects and processes data, the specific operations are as follows:

[0011] Embed a spatio-temporal perception module in the front-end interface of the e-commerce platform. The spatio-temporal perception module uses timestamp and geolocation technologies to record the specific time and geographical location information of each user-item interaction. At the same time, combined with the user's device information, a multi-dimensional user behavior data space is constructed, and a spatio-temporal association algorithm is used to perform association analysis on the behavior data of different users at different times and locations to obtain the user's behavior characteristics;

[0012] For the image, text, and audio data of the item, for the image data, a convolutional neural network is used to extract visual features, and for the text data, a pre-trained language model is used to extract semantic features. Then, these features of different modalities are input into a cross-modal fusion network. The cross-modal fusion network adopts a graph neural network with an attention mechanism. In the graph neural network, the feature nodes of different modalities are connected by edges, and the weights of the edges are dynamically adjusted by the attention mechanism. The attention mechanism assigns weights according to the contribution degree of different modality features to the final fusion feature to extract item features.

[0013] As a further improvement of this technical solution, the dynamic monitoring unit adopts the following method when monitoring and predicting indicators in real time:

[0014] Map the time series data of user behavior characteristics, item characteristics, and platform operation indicators to a quantum state space. The quantum state space uses quantum bits to represent the state of the data, and by simulating the quantum entanglement phenomenon, a correlation model between the data is established. Adopt the quantum entanglement state prediction algorithm, which is based on the principles of quantum mechanics and predicts the future values of the time series by calculating the evolution of the quantum state;

[0015] Construct a causal inference network with user behavior characteristics, product characteristics, and platform operation metrics as the nodes of the network. By analyzing the causal relationships between the nodes, judge the changes in the data distribution, and then use the causal inference algorithm based on the structural causal model to determine the causal effects between different metrics and generate corresponding adjustment signals.

[0016] As a further improvement of this technical solution, the operations of the dynamic monitoring unit when generating adjustment signals and determining the direction and amplitude of parameter adjustment are as follows:

[0017] When the dynamic monitoring unit generates adjustment signals, it can be regarded as an agent. Take the differences between the actual values and predicted values of user behavior characteristics, product characteristics, and platform operation metrics as the environmental state, and the adjustment signals as the actions of the agent. Then, through the deep deterministic policy gradient algorithm, realize the generation of adaptive adjustment signals;

[0018] Encode the loss function weights of the generator and the regularization parameters of the discriminator as chromosomes, initialize a population containing multiple chromosomes, each chromosome representing a set of parameter combinations, use the genetic algorithm, according to the generated adjustment signals, define the fitness function, and through continuous iterative evolution, find the chromosome with the highest fitness, that is, the optimal parameter adjustment amplitude.

[0019] As a further improvement of this technical solution, the operations of the generator of the GAN data augmentation unit when generating simulated data are as follows:

[0020] Before training the generator, use the meta-learning algorithm to optimize the initial parameters of the generator. Divide the historical data into multiple tasks, each task corresponding to a different data distribution, and then use the meta-learner to train on these tasks to generate simulated data that conforms to the data change trend;

[0021] Then, conduct adversarial training on the generator and the discriminator based on historical data to generate simulated user-product interaction data.

[0022] As a further improvement of this technical solution, the operations of the discriminator of the GAN data augmentation unit during training and adjustment are as follows:

[0023] Introduce a pre-trained discriminator as the teacher model and the discriminator as the student model. The teacher model is trained on historical data, and the student model is trained by imitating the output of the teacher model. Use the knowledge distillation algorithm to train the student model by minimizing the difference between the outputs of the student model and the teacher model;

[0024] Among them, during the training process of the discriminator, an adaptive regularization algorithm is adopted. The adaptive regularization algorithm adjusts the regularization parameters according to the change rate of the data distribution and the training error of the discriminator.

[0025] As a further improvement of this technical solution, the operation of the GAN data augmentation unit when alternately updating the parameters of the generator and the discriminator is as follows:

[0026] Regarding the generator and the discriminator as two agents, a coordinating agent is introduced to coordinate their parameter update processes. The coordinating agent formulates a parameter update strategy using a multi-agent deep reinforcement learning algorithm based on the adjustment signal generated by the dynamic monitoring unit and the training status of the generator and the discriminator;

[0027] Transform the parameter update problem of the generator and the discriminator into a combinatorial optimization problem, and use the quantum annealing algorithm to solve this optimization problem to find a parameter update scheme that optimizes the performance of the GAN model;

[0028] Among them, the quantum annealing algorithm searches for the optimal solution in the high-dimensional parameter space by simulating the annealing process of the quantum system. At each parameter update, the current parameter state is used as the initial state, and a better parameter state is found within a certain time through the quantum annealing algorithm, thereby optimizing the parameters of the generator and the discriminator.

[0029] As a further improvement of this technical solution, the operation of the GAN data augmentation unit during training convergence judgment and model evaluation is specifically as follows:

[0030] During the training process, calculate the information entropy of the simulated data generated by the generator and the real data, and judge whether the training converges by comparing the change trends of the information entropy of the simulated data and the real data;

[0031] Define an information entropy convergence index. When the difference between the information entropy of the simulated data and the real data is less than a preset threshold and remains stable within a certain number of training steps, it is considered that the training converges;

[0032] During model evaluation, by adding perturbations to the real data and the generated data, observe the discrimination results of the discriminator and the quality changes of the data generated by the generator, and determine whether the GAN model is trained completed according to the discrimination results of the discriminator and the quality changes of the data generated by the generator.

[0033] Compared with the prior art, the beneficial effects of the present invention:

[0034] In the e-commerce personalized recommendation data system that integrates deep learning, for new users, the data collection and processing unit uses the spatiotemporal perception module to combine time, geographic location and device information to build a multi-dimensional behavioral data space, and uses the spatiotemporal correlation algorithm to accurately analyze behavioral characteristics, so that the system can quickly understand the preferences of new users and improve the platform experience and retention rate of new users. For new products, the GAN data enhancement unit optimizes the initial parameters through the generator through the meta-learning algorithm, combines historical data for adversarial training, generates high-quality simulated user-product interaction data, and enriches the "virtual" interaction records of new products. The discriminator uses the knowledge distillation algorithm and adaptive regularization training to accurately distinguish between real and generated data, provide effective feedback to the generator, optimize the quality of simulated data, help merchants promote new products, and improve sales conversion efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is an overall block diagram of the present invention.

[0036] The meaning of each number in the figure is:

[0037] 1. Data acquisition and processing unit; 2. Dynamic monitoring unit; 3. GAN data enhancement unit; 4. Recommendation service unit. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] The present invention provides an e-commerce personalized recommendation data system integrating deep learning, please refer to Figure 1 As shown, it includes a data acquisition and processing unit 1, a dynamic monitoring unit 2, a GAN data enhancement unit 3 and a recommendation service unit 4;

[0040] The data collection and processing unit 1 collects the behavior data of new and old users, the multimodal data of commodities and the operation data from the e-commerce platform log to form a multimodal data set, and extracts data features from the multimodal data set.

[0041] When the data acquisition and processing unit 1 collects and processes data, the specific operations are as follows:

[0042] In order to capture user behavior information more comprehensively and accurately, it is far from enough to just record regular interactive operations. Time and geographic location factors often have an important impact on users' purchasing decisions and behavior patterns. At the same time, the device information used by users can also reflect their usage habits and preferences. Therefore, embedding a time-space perception module in the front-end interface of the e-commerce platform can help build a richer and multi-dimensional user behavior data space.

[0043] A space-time perception module is embedded in the front-end interface of the e-commerce platform. The space-time perception module uses timestamp and geo-positioning technology to record the specific time and geographic location information of each user's interaction with a product. At the same time, it combines the user's device information to build a multi-dimensional user behavior data space, which can more realistically reflect the overall picture of the user's behavior and help to gain a deeper understanding of user needs.

[0044] The collected multi-dimensional user behavior data is scattered, and it is necessary to use effective methods to explore the potential patterns and characteristics therein. Using spatiotemporal correlation algorithms, the behavior data of different users at different times and places can be correlated and analyzed to obtain user behavior characteristics, providing a more targeted basis for personalized recommendations.

[0045] For product image data, convolutional neural networks are used to extract visual features. Convolutional neural networks have good feature extraction capabilities and generalization performance, and can accurately extract key visual information from images. For product text data, pre-trained language models are used to extract semantic features, which can clearly express the attributes and characteristics of the products, help to better understand the products, and improve the relevance of recommendations.

[0046] These features of different modalities are input into the cross-modal fusion network. The cross-modal fusion network adopts a graph neural network with an attention mechanism. In the graph neural network, the feature nodes of different modalities are connected by edges, and the weights of the edges are dynamically adjusted through the attention mechanism. The attention mechanism assigns weights according to the contribution of different modal features to the final fusion features. The extracted fusion features can comprehensively and accurately describe the products, combining the advantages of each modal data, greatly improving the accuracy and comprehensiveness of product recommendations, and providing users with product recommendations that better meet their needs.

[0047] The dynamic monitoring unit 2 sets user behavior characteristics, product characteristics and platform operation indicators as data characteristics, and uses time series analysis method to monitor and predict various indicators in real time. It judges the amplitude of data distribution change by comparing the difference between actual value and predicted value, and generates adjustment signal according to the amplitude of data distribution change. The adjustment signal is applied to the GAN data enhancement unit 3 to determine the adjustment direction and amplitude of the generator and discriminator training parameters.

[0048] The dynamic monitoring unit 2 uses the following methods to monitor and predict indicators in real time:

[0049] Traditional time series analysis methods have certain limitations when dealing with complex, high-dimensional, and non-linearly correlated data. The quantum state space has unique characteristics. Mapping the time series data of user behavior characteristics, commodity characteristics, and platform operation indicators to the quantum state space can utilize these characteristics to more accurately capture the complex correlations between data. The quantum state space uses qubits to represent the state of data and establishes a correlation model between data by simulating the phenomenon of quantum entanglement, providing a more powerful data representation basis for subsequent accurate prediction and helping to improve the ability to capture the changing trends of indicators.

[0050] Adopt the quantum entanglement state prediction algorithm, which is based on the principles of quantum mechanics and predicts the future values of time series by calculating the evolution of quantum states, so as to more accurately predict in advance the changes in indicators such as user behavior, commodity sales, and platform operation, providing a more reliable basis for the decision-making of e-commerce platforms.

[0051] Merely knowing the changing trend of data is not enough. It is also necessary to understand the causal relationships between different indicators so as to accurately find the reasons and take corresponding measures when the data distribution changes. Construct a causal inference network with user behavior characteristics, commodity characteristics, and platform operation indicators as the nodes of the network, and judge the changes in data distribution by analyzing the causal relationships between the nodes, so as to more accurately locate the reasons for the changes in data distribution and provide a clear direction for subsequent adjustments.

[0052] The structural causal model can formally describe the causal relationships between variables. The causal inference algorithm based on this model can accurately determine the causal effects between different indicators through intervening in the data and counterfactual reasoning, thereby generating reasonable adjustment signals. Let the causal graph , where is the set of nodes, including user behavior characteristics, commodity characteristics, and platform operation indicators, is the set of edges representing causal relationships, then the structural causal model can be expressed as , where is the variable of node , is the set of variables of its direct parent nodes, is the exogenous variable, and the causal effect is calculated through the backdoor criterion, that is , where represents intervening in variable . According to the calculated causal effect and combined with the preset rules, generate an adjustment signal to help the e-commerce platform quickly adjust its strategies to adapt to market changes and improve the operation efficiency and user experience of the platform.

[0053] The operations of the dynamic monitoring unit 2 when generating the adjustment signal and determining the direction and amplitude of parameter adjustment are as follows:

[0054] The dynamic monitoring unit 2 needs to generate an appropriate adjustment signal according to the complex and changeable environmental state. By abstracting it as an intelligent agent and using the idea of reinforcement learning, the dynamic monitoring unit 2 can continuously learn and optimize the generation strategy of the adjustment signal in the interaction with the environment. When the dynamic monitoring unit 2 generates the adjustment signal, it can be regarded as an intelligent agent. The differences between the actual values and the predicted values of user behavior characteristics, commodity characteristics, and platform operation indicators are used as the environmental state, and the adjustment signal is used as the action of the intelligent agent. This abstraction method transforms the problem into a reinforcement learning problem, enabling the dynamic monitoring unit 2 to use the rich algorithms and theories of reinforcement learning to solve complex decision-making problems and enhancing the adaptive ability of the system.

[0055] Then, the adaptive adjustment signal generation is realized through the deep deterministic policy gradient algorithm, which is as follows:

[0056] Policy network Used to generate actions, that is, adjustment signals, where is the difference between the actual value and the predicted value, are the parameters of the policy network, and the value network is used to evaluate the value of the state-action pair, where is the action.

[0057] Then the update formula of the policy network is ;

[0058] The update target of the value network is where is the reward, is the discount factor, and are the parameters of the target network;

[0059] The loss function of the value network is ; The deep deterministic policy gradient algorithm can effectively learn the optimal policy in the continuous action space. By continuously interacting with the environment and updating the network parameters, the generated adjustment signal can be dynamically optimized according to the real-time environmental changes, improving the flexibility and stability of the system.

[0060] ​​Encode the loss function weights of the generator and the regularization parameters of the discriminator as chromosomes, and initialize a population containing multiple chromosomes. Each chromosome represents a set of parameter combinations. The encoding method of the chromosomes transforms the parameter optimization problem into a form that can be processed by the genetic algorithm, facilitating the use of the selection, crossover, and mutation operations of the genetic algorithm for global search. Use the genetic algorithm, define the fitness function according to the generated adjustment signal, and through continuous iterative evolution, find the chromosome with the highest fitness, that is, the optimal parameter adjustment amplitude, as follows:

[0061] Define the fitness function according to the adjustment signal and the performance indicators of the system , adopt the roulette wheel selection method, and the probability of each chromosome being selected ; where is the th chromosome, is the population size, select two chromosomes for crossover with a certain crossover probability to generate new chromosomes, and mutate the genes of the chromosomes with a certain mutation probability in order to determine the optimal parameter adjustment amplitude. The determined optimal parameter adjustment amplitude can enable the generator and the discriminator to better adapt to the changes in the data distribution, improve the performance of the GAN data augmentation unit 3, and further enhance the effect of the entire e-commerce personalized recommendation data system.

[0062] The GAN data augmentation unit 3 adopts a deep convolutional generative adversarial network. The generator receives a random noise vector and outputs simulated user-item interaction data. The discriminator receives real and generated data and outputs the probability that the data is real data, and modifies the loss function weights of the generator according to the adjustment signal to generate simulated data that conforms to the data change trend. Finally, adjust the regularization parameters of the discriminator to adapt to the data distribution change, and use real data and the adjusted generated data to alternately update the parameters of the generator and the discriminator for training until convergence.

[0063] The operations of the generator of the GAN data augmentation unit 3 when generating simulated data are as follows:

[0064] The meta-learning algorithm enables the generator to quickly learn general initialization parameters from multiple different data distributions, enabling the generator to adapt faster and generate high-quality simulated data when facing new data distributions. Before training the generator, use the meta-learning algorithm to optimize the initial parameters of the generator. Divide the historical data into multiple tasks, each task corresponding to a different data distribution, and then use the meta-learner to train on these tasks.

[0065] Let the goal of the meta-learner be to minimize the sum of losses on multiple tasks. Let the task set , each task has its own dataset , for each task , when the generator parameter is , the loss function is . The objective function of meta-learning can be expressed as . In actual implementation, the model-agnostic meta-learning algorithm is adopted, and its steps are as follows:

[0066] For each task , first use the current parameter to calculate the gradient on the task , and then obtain the temporary parameter through one-step gradient update , where is the learning rate of the inner loop. Finally, calculate the meta-loss, and the formula for the meta-loss is , and update the parameter according to the meta-loss. The update formula is , where is the learning rate of the outer loop. By optimizing the initial parameters through meta-learning, the generator can adapt to the new data distribution faster, reducing the training time and improving the training efficiency.

[0067] The core idea of the generative adversarial network is to improve the ability of the generator to generate simulated data through the adversarial training of the generator and the discriminator. The specific operation of conducting adversarial training on the generator and the discriminator based on historical data is as follows:

[0068] Suppose the input of the generator is the random noise vector , and the output is the simulated user-item interaction data . The input of the discriminator is the data , and the output is the probability that the data is real data . Among them, the loss function of the discriminator is , where is the distribution of real data, is the distribution of random noise. The goal of the discriminator is to maximize this loss function, that is, to distinguish real data and generated data as accurately as possible.

[0069] The loss function of the generator is . The goal of the generator is to minimize this loss function, that is, the generated simulated data can make the discriminator mistake it for real data. During the training process, the parameters of the discriminator and the generator are updated alternately. For the discriminator, the gradient ascent method is used to update the parameters ; for the generator, the gradient descent method is used to update the parameters , where and They are the learning rates of the discriminator and the generator respectively. Adversarial training can give full play to the mutual promotion between the generator and the discriminator, enabling the generator to continuously learn how to generate more realistic simulated data and the discriminator to continuously improve its discrimination ability.

[0070] The operations of the discriminator of the GAN data augmentation unit 3 during training and adjustment are as follows:

[0071] When training the discriminator, starting from scratch may face problems such as slow convergence speed and easy overfitting. Introduce a pre-trained discriminator as the teacher model and the discriminator as the student model. The teacher model is trained on historical data, and the student model is trained by mimicking the output of the teacher model and using the knowledge distillation algorithm to train the student model by minimizing the difference between the outputs of the student model and the teacher model.

[0072] Let the output of the teacher model be , and the output of the student model be , and the loss function of knowledge distillation usually consists of two parts:

[0073] The distillation loss is used to measure the difference between the outputs of the student model and the teacher model and is calculated using KL divergence, that is . In actual calculation, for classification problems, if and are probability distribution vectors, ;

[0074] The original discriminator loss : that is, the original loss function of the discriminator, which is used to distinguish real data and generated data. The final knowledge distillation loss function is , where is a parameter used to balance the weights of the distillation loss and the original discriminator loss. When training the student model, the gradient descent method is used to update the model parameters , and the update formula is , where is the learning rate. The discriminator can converge to a better state faster, the discrimination accuracy is improved, it can distinguish real data and generated data more accurately, and provide more effective feedback for the generator, promoting the generator to generate higher-quality simulated data.

[0075] During the training process of the discriminator, the data distribution may change over time. The adaptive regularization algorithm is adopted. The adaptive regularization algorithm adjusts the regularization parameter according to the change rate of the data distribution and the training error of the discriminator. Let the original loss function of the discriminator be , and the regularization term be . Usually, regularization is adopted, that is , where are the parameters of the discriminator, is the regularization parameter.

[0076] The discriminator loss function with regularization is .

[0077] The change rate of the data distribution can be measured by calculating the change in the statistical features of the data in adjacent training batches. The training error of the discriminator can be obtained by calculating the error of the discriminator on the validation set. The adaptive regularization parameter update formula is , where is the regularization parameter of the th training batch, is the update step size, is an adjustment function calculated according to the change rate of the data distribution and the training error . In each training batch, the loss function with adaptive regularization is used to update the parameters of the discriminator , and the update formula is , where is the learning rate. The adaptive regularization algorithm can dynamically adjust the regularization parameter according to the actual change of the data distribution and the training state of the discriminator, so that the discriminator can maintain good generalization ability and discrimination performance under different data distributions, enable the generator to generate simulated data more in line with the real data distribution, and improve the performance of the entire GAN data augmentation unit 3.

[0078] The operations of the GAN data augmentation unit 3 when alternately updating the parameters of the generator and the discriminator are as follows:

[0079] Regard the generator and the discriminator as two agents, introduce a coordination agent to coordinate their parameter update processes, and the coordination agent formulates a parameter update strategy using the multi-agent deep reinforcement learning algorithm according to the adjustment signal generated by the dynamic monitoring unit 2 and the training states of the generator and the discriminator.

[0080] Let the parameters of the generator agent be , and the parameters of the discriminator agent be . The state information received by the coordination agent includes the adjustment signal generated by the dynamic monitoring unit 2 and the training states of the generator and the discriminator, that is, . The coordination agent Generate actions , actions represent the parameter update strategies of the generator and discriminator, and define the reward function to measure the quality of the coordination strategy. The goal is to make the losses of the generator and discriminator as close as possible to maintain the balance of training.

[0081] The coordination agent is trained using the Deep Deterministic Policy Gradient algorithm. The update formula for its policy network is ; where is the value network, which is used to evaluate the value of state-action pairs. The generator and discriminator update their parameters according to the policy given by the coordination agent. The update formula for the generator's parameters is , and the update formula for the discriminator's parameters is , where and are the learning rates determined according to the policy . The training convergence speed of the GAN model is accelerated, and the generated simulated data has higher quality and better diversity, which can more accurately reflect the distribution of real data and provide better training data for the subsequent recommendation model.

[0082] Transform the parameter update problem of the generator and discriminator into a combinatorial optimization problem, and use the quantum annealing algorithm to solve this optimization problem to find the parameter update scheme that optimizes the performance of the GAN model. Among them, the quantum annealing algorithm searches for the optimal solution in the high-dimensional parameter space by simulating the annealing process of the quantum system. At each parameter update, the current parameter state is used as the initial state, and a better parameter state is found within a certain time through the quantum annealing algorithm.

[0083] Let the parameter vectors of the generator and discriminator be , and define the objective function to measure the performance of the GAN model. The quantum annealing algorithm describes the energy of the quantum system through the Hamiltonian , where is the annealing parameter. Initially , and finally . The Hamiltonian consists of two parts , where is the initial Hamiltonian, which is used to introduce quantum fluctuations is the final Hamiltonian, which is related to the objective function .

[0084] During the annealing process, as gradually increases from 0 to 1, the quantum system gradually evolves from the initial state to the final state. The evolution of the quantum system is simulated by solving the Schrödinger equation , where is a quantum state, is the reduced Planck constant.

[0085] At each parameter update, the current parameter status As the initial state, the quantum annealing algorithm is used for a certain period of time. Find it in Minimum parameter status , and then update the parameters of the generator and discriminator as , the performance of the GAN model has been significantly improved, the generated simulated data has higher quality and is more similar to the real data, which can better meet the needs of e-commerce personalized recommendation data system for data enhancement and improve the accuracy and generalization ability of the recommendation model.

[0086] The operations of GAN data enhancement unit 3 during training convergence judgment and model evaluation are as follows:

[0087] In the GAN training process, judging whether the training has converged is a key issue. The information entropy of the simulated data and real data generated by the generator can be calculated, and the change trend of the information entropy of the simulated data and the real data can be compared to judge whether the training has converged.

[0088] For discrete data, the calculation formula for information entropy is: ,in is a data random variable, Is the data value The probability of calculating the information entropy of real data And the information entropy of the generated data ,Information entropy starts from the essential characteristics of data distribution, and can comprehensively consider the diversity and distribution of data. Compared with the simple loss function judgment, it can more accurately reflect the similarity between generated data and real data, and thus more reliably judge whether the training has converged.

[0089] Simply observing the trend of information entropy changes is not accurate enough. In order to more clearly judge the convergence of training, it is necessary to define a specific convergence indicator. The information entropy convergence indicator is defined as the difference between the information entropy of the simulated data and the real data is less than a preset threshold. , and at a certain number of training steps When ∈R remains stable, the training is considered to have converged.

[0090] Assume that the difference in information entropy ,when , and in succession In the training steps, The range of change is less than another small threshold , it is determined that the training converges. The clear convergence index makes the convergence judgment more objective and quantifiable, reduces the uncertainty of subjective judgment, and helps with the comparison and reproduction between different experiments.

[0091] In practical applications, the GAN model needs to have a certain degree of robustness, that is, it can still maintain good performance when the data is slightly perturbed. By adding perturbations to the real data and the generated data, observing the changes in the discrimination results of the discriminator and the quality of the data generated by the generator, the robustness and generalization ability of the model can be evaluated, and it can be judged whether the model is truly trained.

[0092] For real data and generated data add perturbations. Let the perturbation function be . The data after adding perturbations is and . Calculate the change in the discrimination accuracy of the discriminator for the original data and the perturbed data. Let the discrimination accuracy of the discriminator for the original real data be , the discrimination accuracy for the perturbed real data be , the discrimination accuracy for the original generated data be , and the discrimination accuracy for the perturbed generated data be . Define the accuracy change rate . For the change in the quality of the data generated by the generator, let the value of the original generated data be , and the FID value of the perturbed generated data be . Define the change rate . When and are both less than the preset threshold, it is considered that the GAN model training is completed. This evaluation method takes into account the noise and perturbation situations that the model may encounter in practical applications, can more comprehensively evaluate the robustness and generalization ability of the model, avoids the problem that the model performs well on clean data but is easily interfered in actual scenarios, and enables the quality of the generated data to meet the actual needs of e-commerce personalized recommendations.

[0093] When the recommendation service unit 4 inputs new user data, the GAN model generates a personalized recommendation list based on its behavior data and collects the feedback data of the new user on the recommendation results to retrain the GAN model.

[0094] In the present invention, the data acquisition and processing unit 1 comprehensively collects multi-modal data of users and commodities by embedding a spatio-temporal perception module, accurately extracts features, the dynamic monitoring unit 2 uses quantum state space and causal inference network technologies to monitor and predict data changes in real time, generates adjustment signals, the GAN data augmentation unit 3 generates high-quality simulated data based on a deep convolutional generative adversarial network, with the help of algorithms such as meta-learning and knowledge distillation to improve the data augmentation effect, and the recommendation service unit 4 trains a model based on the augmented data to achieve accurate recommendation, improve the exposure of new commodities, and enhance the accuracy and efficiency of recommendation.

[0095] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification are only preferred examples of the present invention and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. The e-commerce personalized recommendation data system integrating deep learning is characterized by: It includes a data collection and processing unit (1), a dynamic monitoring unit (2), a GAN data enhancement unit (3) and a recommendation service unit (4); The data collection and processing unit (1) collects behavioral data of new and old users, multimodal data of commodities, and operation data from the e-commerce platform logs to form a multimodal data set, and extracts data features from the multimodal data set; The dynamic monitoring unit (2) sets user behavior characteristics, product characteristics and platform operation indicators as data characteristics, uses time series analysis method to monitor and predict various indicators in real time, judges the data distribution change range by comparing the difference between actual value and predicted value, and generates adjustment signals according to the data distribution change range, applies the adjustment signals to the GAN data enhancement unit (3), and determines the adjustment direction and range of the generator and discriminator training parameters; The GAN data enhancement unit (3) adopts a deep convolutional generative adversarial network, wherein the generator receives a random noise vector and outputs simulated user-product interaction data, and the discriminator receives real and generated data and outputs the probability that the data is real data, and modifies the weight of the generator loss function according to the adjustment signal to generate simulated data that conforms to the data change trend, and finally adjusts the regularization parameter of the discriminator to adapt to the change of data distribution, and uses the real data and the adjusted generated data to alternately update the parameters of the generator and the discriminator for training until convergence; When new user data is input into the recommendation service unit (4), the GAN model generates a personalized recommendation list based on the new user's behavior data, and collects feedback data on the recommendation results from the new user to retrain the GAN model.

2. The e-commerce personalized recommendation data system integrating deep learning according to claim 1 is characterized by: When collecting and processing data, the data acquisition and processing unit (1) specifically operates as follows: The spatiotemporal perception module is embedded in the front-end interface of the e-commerce platform. The spatiotemporal perception module uses timestamp and geolocation technology to record the specific time and geographic location information of each user's interaction with the product. At the same time, it combines the user's device information to build a multi-dimensional user behavior data space, and uses the spatiotemporal association algorithm to associate and analyze the behavior data of different users at different times and locations to obtain the user's behavior characteristics; For the images, texts, and audio data of products, a convolutional neural network is used to extract visual features for image data, and a pre-trained language model is used to extract semantic features for text data. Then, the features of these different modalities are input into a cross-modal fusion network, which uses a graph neural network with an attention mechanism. In the graph neural network, feature nodes of different modalities are connected by edges, and the weights of the edges are dynamically adjusted through the attention mechanism. The attention mechanism assigns weights according to the contribution of different modal features to the final fusion features to extract product features.

3. The e-commerce personalized recommendation data system integrating deep learning according to claim 2 is characterized by: The dynamic monitoring unit (2) adopts the following method when monitoring and predicting indicators in real time: Mapping the time series data of user behavior characteristics, product characteristics and platform operation indicators into the quantum state space, which uses quantum bits to represent the state of data, and establishes a correlation model between data by simulating quantum entanglement phenomena. The quantum entanglement state prediction algorithm is used. The algorithm is based on the principles of quantum mechanics and predicts the future value of the time series by calculating the evolution of quantum states. Construct a causal inference network, use user behavior characteristics, product characteristics and platform operation indicators as network nodes, judge the changes in data distribution by analyzing the causal relationship between nodes, and then use the causal inference algorithm based on the structural causal model to determine the causal effects between different indicators and generate corresponding adjustment signals.

4. The e-commerce personalized recommendation data system integrating deep learning according to claim 3 is characterized by: The operation of the dynamic monitoring unit (2) when generating an adjustment signal and determining a parameter adjustment direction and amplitude is specifically as follows: When the dynamic monitoring unit generates an adjustment signal, it can be regarded as an intelligent agent. The difference between the actual value and the predicted value of the user behavior characteristics, product characteristics and platform operation indicators is taken as the environmental state, and the adjustment signal is taken as the action of the intelligent agent. Then, the adaptive adjustment signal generation is realized through the deep deterministic policy gradient algorithm. The loss function weights of the generator and the regularization parameters of the discriminator are encoded as chromosomes, and a population of multiple chromosomes is initialized. Each chromosome represents a set of parameter combinations. A genetic algorithm is used to define the fitness function based on the generated adjustment signal. Through continuous iterative evolution, the chromosome with the highest fitness, that is, the optimal parameter adjustment range, is found.

5. The e-commerce personalized recommendation data system integrating deep learning according to claim 4 is characterized by: The operation of the generator of the GAN data enhancement unit (3) when generating simulated data is as follows: Before training the generator, a meta-learning algorithm is used to optimize the initial parameters of the generator, and the historical data is divided into multiple tasks. Each task corresponds to a different data distribution. The meta-learner is then used to train on these tasks to generate simulated data that conforms to the data change trend. The generator and discriminator are then adversarially trained based on historical data to generate simulated user-product interaction data.

6. The e-commerce personalized recommendation data system integrating deep learning according to claim 5 is characterized by: The operation of the discriminator of the GAN data enhancement unit (3) during training and adjustment is as follows: A pre-trained discriminator is introduced as the teacher model, and the discriminator is used as the student model. The teacher model is trained on historical data, and the student model is trained by imitating the output of the teacher model. The knowledge distillation algorithm is used to train the student model by minimizing the difference between the output of the student model and the teacher model. The discriminator uses an adaptive regularization algorithm during training, and the adaptive regularization algorithm adjusts the regularization parameter according to the rate of change of data distribution and the training error of the discriminator.

7. The e-commerce personalized recommendation data system integrating deep learning according to claim 6 is characterized by: The operation of the GAN data augmentation unit (3) when alternatingly updating the parameters of the generator and the discriminator is as follows: The generator and the discriminator are regarded as two intelligent agents, and a coordination agent is introduced to coordinate their parameter update process. The coordination agent uses a multi-agent deep reinforcement learning algorithm to formulate a parameter update strategy according to the adjustment signal generated by the dynamic monitoring unit (2) and the training status of the generator and the discriminator; The parameter update problem of the generator and discriminator is transformed into a combinatorial optimization problem, and the quantum annealing algorithm is used to solve this optimization problem to find the parameter update scheme that optimizes the performance of the GAN model. The quantum annealing algorithm searches for the optimal solution in a high-dimensional parameter space by simulating the annealing process of a quantum system. Each time the parameters are updated, the current parameter state is used as the initial state, and a better parameter state is found within a certain period of time through the quantum annealing algorithm, thereby optimizing the parameters of the generator and the discriminator.

8. The e-commerce personalized recommendation data system integrating deep learning according to claim 7 is characterized by: The operation of the GAN data enhancement unit (3) during training convergence judgment and model evaluation is as follows: During the training process, the information entropy of the simulated data and the real data generated by the generator is calculated, and by comparing the change trends of the information entropy of the simulated data and the real data, it is determined whether the training has converged; Define the information entropy convergence index. When the difference between the information entropy of the simulated data and the real data is less than a preset threshold and remains stable within a certain number of training steps, the training is considered to have converged. When evaluating the model, by adding perturbations to the real data and generated data, the discriminator's discrimination results and the quality changes of the generator's generated data are observed, and whether the GAN model is trained is determined based on the discriminator's discrimination results and the quality changes of the generator's generated data.

Images