Personalized recommendation method and system based on optimization meta-learning, and storage medium

Abstract

The application discloses a kind of personalized recommendation method, system and storage medium based on optimization meta-learning, adopt the time series prediction model based on decomposition to the user behavior time series data decomposition, extract the trend feature and seasonal feature therein, to realize future behavior prediction;Meta-learning method is used to train the time series prediction model based on decomposition, summarize different users, and obtain a better generalization initialization model;Optimal path finding in model training process is realized and the optimal solution of the model is realized using the multi-objective optimization algorithm based on Pareto optimality;The personalized user behavior habit model is obtained by fine-tuning the initialization model, which is used to realize personalized recommendation for different users.The application can deeply mine the behavior data of household users, construct the user behavior habit model and interest model, generate personalized recommendation strategies for users, and greatly improve the accuracy of user recommendation strategies.

Description

Technical Field

[0001] This invention belongs to the technical field of personalized recommendation based on deep learning, specifically relating to a personalized recommendation method, system, and storage medium based on optimization meta-learning. Background Technology

[0002] With societal development and technological advancements, people's living standards have improved significantly. The most obvious improvement is the increasing number and variety of smart devices in homes, greatly enhancing comfort, convenience, and security. The technological sophistication of homes continues to rise, and smart homes now offer a variety of applications: smart lighting, smart locks, smart curtains, smart TVs, and smart speakers. Smart lighting automatically adjusts brightness based on user habits and needs; smart locks use facial recognition and fingerprint scanning for unlocking and burglar alarms; smart curtains automatically adjust their opening based on lighting conditions; and smart TVs recommend movies and TV shows based on user preferences. With the development of artificial intelligence, big data, and the Internet of Things, people's demands for smart homes will continue to rise. The intelligentization, humanization, and widespread adoption of smart homes are inevitable trends, becoming an integral part of future home life and bringing people a smarter, more comfortable, and healthier lifestyle.

[0003] One of the key aspects of smart home technology is personalized and precise recommendations based on user behavior and preferences. As the number of devices and services in smart home systems increases, users face a greater abundance of information and choices. A crucial challenge in smart home technology is identifying user interests from this information and providing personalized recommendations. Currently, personalized recommendations in smart home scenarios are mostly focused on individual products. For example, smart speakers use voice recognition and natural language processing to identify user commands and recommend personalized music; smart air conditioners provide personalized temperature and humidity control based on historical usage records, temperature, and humidity; and smart lighting systems offer personalized lighting scenes based on user behavior and lighting needs. The future of smart homes will inevitably extend to the entire house, predicting and recommending user behavior based on user behavior data, allowing users to enjoy a comfortable living environment without pressing buttons or using voice commands.

[0004] However, while the aforementioned personalized recommendation methods perform well for individual devices, they fall short when it comes to personalized recommendations for the entire home. This is because traditional methods have simple rules and cannot delve deeper into user behavior and preferences. Furthermore, as user data accumulates, the models cannot be effectively updated, thus failing to provide accurate and effective recommendation services. These issues significantly impact the mining of behavioral habits, interests, and social relationships in smart home scenarios, thereby affecting the accuracy and effectiveness of recommendation strategies. Summary of the Invention

[0005] The purpose of this invention is to provide a personalized recommendation method, system, and storage medium based on optimization meta-learning. By using a deep learning model to mine and model user behavior data, it can quickly fine-tune the model with a small amount of user data, thereby achieving rapid personalized prediction and recommendation for different users.

[0006] Effective mining of user behavior data is a crucial component of personalized recommendations. Data sets include home behavior trajectories, actions, and the usage of various smart appliances. This invention uses big data technology to model the behavioral habits of a large number of users, obtaining the living and device habit characteristics of smart home users. The model is then adjusted for specific users, enabling the rapid and accurate establishment of personalized user behavior models. Furthermore, based on the algorithm model, the user's lifestyle and habits are evaluated, and future behavior is predicted, allowing for the generation of precise service strategies and the provision of personalized services. This can comprehensively improve the intelligence level of smart home living, effectively enhance the quality of life, and have a significant impact on the smart home industry.

[0007] This invention is mainly achieved through the following technical solutions:

[0008] A personalized recommendation method based on optimization meta-learning is proposed. This method employs a decomposition-based time series prediction model to decompose user behavior time-series data, extracting trend and seasonal features to predict future behavior. A meta-learning approach is used to train the decomposition-based time series prediction model, summarizing different users to obtain an initial model with good generalization. A Pareto optimality-based multi-objective optimization algorithm is used to find the optimal path during model training and achieve the optimal solution for the model. Finally, the initial model is fine-tuned to obtain a personalized user behavior habit model, which is used to implement personalized recommendations for different users.

[0009] To better realize the present invention, the following steps are further included:

[0010] Step S1: Perform data preprocessing on the user behavior time series data, and then divide the dataset into training set, validation set, and test set according to a 6:2:2 ratio; the training set is used for model training; the validation set is used to test the network structure and adjust the model hyperparameters to determine whether overfitting occurs; the test set is used to test the generalization ability of the model.

[0011] Step S2: Use the decomposition-based time series prediction model as the base model and train the base model using the training set. Decompose the time series data layer by layer through the residual structure to finally obtain the trend features and periodic features between different levels of user behavior time series data.

[0012] Step S3: Train the basic model using the optimal meta-learning method. By summarizing the commonalities among different users, mining the model's prior knowledge, and obtaining robust initialization parameters, an initialization model applicable to all users is obtained.

[0013] Step S4: When a new user joins, the initial model is fine-tuned using a small amount of data to obtain a user behavior habit model for the new user, which is used to achieve personalized recommendations for different users.

[0014] To better implement the present invention, further, in step S1, the data preprocessing includes: performing linear normalization on the user behavior time series data to transform the original user behavior data to the range of [0,1], so as to balance the impact of different behaviors on the model; for user behavior time series data with default values ​​in some time periods, uniformly filling in the upper and lower data to ensure the consistency of the time granularity of the input data.

[0015] To better implement the present invention, in step S2, the time series prediction model based on decomposition includes several stacks arranged from top to bottom. Each stack consists of several blocks arranged from top to bottom. Each block includes an encoder, a self-attention module, and two fully connected layers arranged in parallel, arranged from front to back.

[0016] The encoder consists of several LSTM units, used to extract short- and long-time feature vectors from the input vector:

[0017] h1 = LSTMs(x i )

[0018] The self-attention module is used to enable the model to pay attention to the correlation between different time points and to focus more on the key information in the input vector. The specific formula is as follows:

[0019] h2 = Att(h1)

[0020] Att(Q, K, V) = w(QK) T V

[0021] Where Q, K, and V are respectively the input sequence multiplied by the weight matrix W. q W k W v get,

[0022] QK T To calculate the similarity between sample features and information features,

[0023] w() is an activation function used to obtain the distribution of similarity or correlation between various samples and various pieces of information.

[0024] V is a value vector matrix;

[0025] The outputs of the two fully connected layers are as follows:

[0026]

[0027]

[0028] The For use in the next block operation, the Used to generate a stack for addition operations Output, the Used for superposition operations to obtain prediction results.

[0029] To better realize the present invention, further, in step S3, the optimal meta-learning method uses multi-objective optimization to balance multiple tasks in the meta-learning training process, assigning optimal weights to different tasks, so that the model training process iterates along the optimal path to obtain the optimal path and result; the optimization objective is as follows:

[0030]

[0031] Among them, L τi (θ t ) represents the model loss after fine-tuning for the t-th task.

[0032] The minimum sum of the losses from different tasks multiplied by their weights is transformed into the minimum norm of the matrix composed of the losses from different tasks.

[0033] To better realize the present invention, the optimal meta-learning-based method in step S3 further includes the following steps:

[0034] Step A1: Input the meta-training task Meta-testing task Task learning rate α, meta-learning rate β, memory learning rate γ, and parameter θ' are randomly initialized.

[0035] Step A2: During meta-training, for the meta-training task, calculate the loss using the support set data of the training task.

[0036] Update parameter θ' using this loss:

[0037]

[0038] Calculate the loss using the query set data from the training task.

[0039] Step A3: Calculate the gradient direction change for each task.

[0040] The optimal weight α for each task with respect to the initial parameters is calculated using an optimization algorithm. 1 ,…,α T ;

[0041] Update shared parameters using optimal weights:

[0042]

[0043] Step A4: For the meta-test task, calculate the loss using the support set data of that test task. Update parameter θ' using this loss:

[0044]

[0045] Use the updated parameter θ′ test-i Make predictions on the query set data of the test task.

[0046] To better implement this invention, step A3 further includes the following steps: Calculating the optimal weights for each task with respect to the initial parameters using an optimization algorithm.

[0047] 1) Initialize the weight vector α corresponding to the task.

[0048] 2) Calculate the gradient descent direction for each task. Relationship M between them:

[0049]

[0050]

[0051]

[0052]

[0053] Where e represents converting the original vector into a unit vector.

[0054] γ represents the optimized path direction of the task weight vector.

[0055] 3) When the maximum number of iterations is reached, output the optimal weight α. 1 ,…,α T .

[0056] To better realize this invention, the loss function is further defined as the mean squared error function, which is as follows:

[0057]

[0058] in, For the predicted value, y i This is the actual value.

[0059] This invention is mainly achieved through the following technical solutions:

[0060] A personalized recommendation system based on optimization meta-learning includes a data preprocessing module, a training module, and a personalization fine-tuning module. The data preprocessing module processes user behavior time-series data to obtain a training set, a validation set, and a test set. The training module trains a decomposition-based time-series prediction model using a Pareto optimal multi-objective optimization algorithm and meta-learning methods to obtain an initialization model. The personalization fine-tuning module uses data from specific users to personalize and fine-tune the initialization model to obtain a personalized user behavior habit model, thereby achieving personalized recommendations for specific users.

[0061] A storage medium storing computer program instructions that, when executed by a processor, implement the method described above.

[0062] The beneficial effects of this invention are as follows:

[0063] (1) This invention employs optimal meta-learning to train the model. By constructing different tasks from the data of multiple users, and then summarizing prior knowledge in different tasks, model parameters applicable to all users are obtained. Simultaneously, a multi-objective optimization algorithm is used to improve upon the traditional meta-learning algorithm, balancing the influence of different tasks on prior knowledge and obtaining the optimal path and optimal solution in the solution process. The trained personalized model can be fine-tuned with only a small amount of data when new users join, effectively providing new users with effective and personalized recommendation strategies, greatly improving the user's living experience.

[0064] (2) This invention can deeply mine the behavioral data of household users, construct user behavior habit models and interest models, and generate personalized recommendation strategies for users. Compared with traditional methods, this invention decomposes the data layer by layer to more deeply mine the implicit relationships within historical behaviors, and can better generate the trend and periodic characteristics of habits. By using the above methods to model user behavior more comprehensively, the accuracy of user recommendation strategies is greatly improved. Attached Figure Description

[0065] Figure 1 This is a flowchart of the personalized recommendation method of the present invention;

[0066] Figure 2 This is a schematic diagram of the structure of a time series prediction model based on decomposition. Detailed Implementation

[0067] Example 1:

[0068] This invention presents a personalized recommendation method based on optimization meta-learning. It primarily utilizes a decomposition-based time series prediction model to decompose user behavior time-series data, extracting trend and seasonal features to predict future behavior. To avoid overfitting when user data is limited, meta-learning is employed to summarize different users, resulting in an initial model with good generalization. Even with only a small amount of user data, the model performs well after fine-tuning. Furthermore, to balance the impact of different users on the meta-learner during meta-training, a Pareto optimal multi-objective optimization algorithm is used to find the optimal path and achieve the optimal solution for the model during training.

[0069] This invention can summarize and mine common features among multiple user data and effectively balance the impact of different user data on the model, ultimately enabling rapid model adjustment for new users. Specifically, it can accurately predict future user behavior with only a small amount of user behavior data, thereby greatly improving the level of intelligence in smart home scenarios.

[0070] Preferably, such as Figure 1As shown, this invention includes user behavior data preprocessing, data partitioning, a decomposition-based time series prediction model, and optimization-based meta-learning. This invention uses a decomposition-based time series prediction model to mine information such as interests and behavioral habits from users' historical behavior data. The model is then trained using meta-learning methods to obtain initial parameters with strong generalization capabilities. These initial parameters can be fine-tuned with limited data to achieve excellent model performance, thus enabling personalized prediction and recommendations. Furthermore, to address the task balance problem in the meta-learning process and find the Pareto optimality of the impact of different user data on the initial parameters during training, a multi-objective optimization approach is introduced. Weights are dynamically allocated to user data, achieving optimal training path and optimal results during model training.

[0071] This invention enables in-depth analysis of household user behavior data, constructing user behavior habit and interest models to generate personalized recommendation strategies. Compared to traditional methods, this invention decomposes data layer by layer to more deeply uncover implicit relationships within historical behavior, better generating trend and periodic characteristics of habits. By using these methods to more comprehensively model user behavior, the accuracy of user recommendation strategies is significantly improved.

[0072] Example 2:

[0073] A personalized recommendation method based on optimization meta-learning includes the following steps:

[0074] S1. First, we need to preprocess the user behavior data (including data normalization, handling default values, etc.). Then, we divide the dataset into training, validation, and test sets in a 6:2:2 ratio. The training set is used for model training, the validation set is used to test the network structure and adjust the model's hyperparameters to determine if it is overfitting, and the test set is used to test the model's generalization ability.

[0075] S2, we use the time series prediction model based on decomposition as the base model, train it with the training dataset, and decompose the time series data layer by layer through the residual structure to finally obtain the trend features and periodic features between different levels of user behavior data.

[0076] S3. To achieve rapid personalized recommendations for users, our time series prediction model based on decomposition does not employ conventional deep learning training methods. Instead, we train the model using an optimal meta-learning approach. By summarizing the features of multiple tasks (user data), we obtain robust prior knowledge (initialization parameters). This generated prior knowledge can be used for fine-tuning when new users join, meaning that accurate user behavior habits and interests can be obtained using only a small amount of historical data.

[0077] Preferably, the purpose of step S1 is that, for user behavior data, different behavioral features have influence in terms of dimensions and orders of magnitude. Therefore, we need to perform linear normalization (Min-Max Scaling) to transform the original user behavior data into the range [0,1] to balance the influence of different behaviors on the model. In addition, for user behavior data, there may be missing values ​​in some time periods. We uniformly fill in the missing values ​​by adding data to the previous and next rows to ensure the consistency of the time granularity of the input data.

[0078] Preferably, the purpose of step S2 is to, as Figure 2 As shown, we need to model the deep features of user behavior data. By decomposing the time series data layer by layer, we can obtain the trends and periodic changes of each decomposition layer step by step. The decomposition-based model has several stacks, which use forward and residual connections. Each stack has two outputs: one passed down and one passed to the right. The value passed down is for further decomposing the time series data, while the value passed to the right is for summarizing the prediction results of each stack, and based on this, we can make predictions about user behavior.

[0079] Preferably, we obtain the input data X = input ∈ R through step S1. T×N T represents the time duration of the user behavior data, and N represents the dimension of the user behavior data. The data is then fed into a decomposition-based prediction model. The model computation includes several Stack operations, each generating two values. Used for stack operations at the next level. The results are obtained through superposition operations. The model's calculation process is as follows:

[0080]

[0081] The calculation process of Stack is as follows:

[0082]

[0083] Each stack block consists of several blocks, and each block also generates two outputs. Used for the next block operation Used to generate a stack for addition operations The output and calculation process are as follows:

[0084]

[0085]

[0086] Where W represents the learnable model weights.

[0087] Each block consists of an encoder composed of several LTSMs, a self-attention module, and a fully connected layer. The LSTMs module contains several LSTM units that extract long and short time features from the input vector.

[0088] h1 = LSTMs(x i )

[0089] Next, the obtained temporal feature vector is fed into the self-attention module, enabling the model to focus on the correlation between different time points and pay more attention to the key information in the input vector. The specific formula is as follows:

[0090] h2 = Att(h1)

[0091] The formula for the self-attention module is as follows:

[0092] Att(Q, K, V) = w(QK) T V

[0093] Q, K, and V are the input sequences multiplied by the weight matrix W, respectively. q W k W v QK T The similarity between sample features and information features is calculated. The distribution of similarity or correlation between each sample and each piece of information is obtained through the activation function w(). Finally, the result of self-attention is obtained by multiplying it with the value vector matrix V.

[0094] The previous step was then computed using fully connected layers to obtain the corresponding outputs, as follows:

[0095]

[0096]

[0097] Preferably, the purpose of step S3 is to use optimization meta-learning to train the decomposition-based prediction model. This method summarizes the commonalities among different users, mines prior knowledge of the model, and allows for fine-tuning with a small amount of data based on prior knowledge when new users join, resulting in a well-performing model and enabling personalized recommendation strategies for specific users. Compared to traditional meta-learning algorithms, we introduce the idea of ​​optimization, using multi-objective optimization to balance multiple tasks during the meta-learning training process, assigning optimal weights to different tasks, so that the model iterates along the optimal path during training to obtain the optimal path and results.

[0098] Preferably, the optimization objective is expressed by the following formula:

[0099]

[0100] Among them, L τi (θ t ) represents the model loss after fine-tuning for the t-th task. The minimum value of the sum of the losses of different tasks multiplied by their weights is transformed into the minimum norm of the matrix composed of the losses of different tasks.

[0101] Preferably, the algorithm flow after incorporating multi-objective optimization algorithms into meta-learning is as follows:

[0102] Algorithm 1 is based on optimal meta-learning:

[0103] Input: Meta-training task Meta-testing task Task learning rate α, meta-learning rate β, memory learning rate γ

[0104] Output: Meta-test task The query set behavior recognition results.

[0105] 1. Randomly initialize parameter θ'

[0106] 2. while not done

[0107] 3. / * Meta-training process * /

[0108] 4. From Randomly select a batch of tasks

[0109] 5. For task in tasks:

[0110] 6. Calculate the loss using the support set data from this training task.

[0111] 7. Use this loss to update parameter θ'.

[0112] 8. Calculate the loss using the query set data from the training task.

[0113] 10.end

[0114] 11. Calculate the gradient direction change for each task.

[0115] 12. Using Algorithm 2, calculate the optimal weight α for each task with respect to the initial parameters. 1 ,…,α T

[0116] 13. Update shared parameters using weights.

[0117] 12.end

[0118] 13. for each task from

[0119] 14. / * Meta-testing process * /

[0120] 15. Calculate the loss using the support set data from this test task.

[0121] 16. Use this loss to update the parameter θ′.

[0122] 17. Use the updated parameter θ′ test-i Predict the query set data for the test task.

[0123] 18. End.

[0124] Preferably, the optimal weights for each task are calculated based on the gradient direction and loss for different tasks. The specific calculation process for the optimal weights is as follows:

[0125] Algorithm 2: Optimization Algorithm

[0126] Input: Gradient update directions for multiple tasks

[0127] 1. Initialization

[0128] 2. Calculate the relationship

[0129] 3.do

[0130] 4.

[0131] 5.

[0132] 6.

[0133] 7.while or reached the maximum number of iterations

[0134] 8. returnα 1 ,…,α T

[0135] Preferably, the model in this paper is a regression model, and the loss function used is the mean squared error function, which is as follows:

[0136]

[0137] in, For the predicted value, y i This is the actual value.

[0138] This invention employs optimization meta-learning to train the model. By constructing different tasks from data from multiple users, and then summarizing prior knowledge across these tasks, model parameters applicable to all users are obtained. Simultaneously, a multi-objective optimization algorithm is used to improve upon traditional meta-learning algorithms, balancing the influence of different tasks on prior knowledge and obtaining the optimal path and solution in the solution process. The trained personalized model can be fine-tuned with only a small amount of data when new users join, effectively providing them with efficient and personalized recommendation strategies, greatly improving the user's living experience.

[0139] This invention enables in-depth analysis of household user behavior data, constructing user behavior habit and interest models to generate personalized recommendation strategies. Compared to traditional methods, this invention decomposes data layer by layer to more deeply uncover implicit relationships within historical behavior, better generating trend and periodic characteristics of habits. By using these methods to more comprehensively model user behavior, the accuracy of user recommendation strategies is significantly improved.

[0140] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A personalized recommendation method based on optimization meta-learning, characterized in that, A decomposition-based time series prediction model is used to decompose user behavior time series data and extract trend and seasonal features to predict future behavior. A meta-learning approach is used to train a time series prediction model based on decomposition. By summarizing different users, an initial model with good generalization is obtained. A Pareto optimal multi-objective optimization algorithm is used to find the optimal path during the model training process and obtain the optimal solution of the model. The initial model is fine-tuned to obtain a personalized user behavior habit model, which is used to realize personalized recommendations for different users. Includes the following steps: Step S1: Perform data preprocessing on the user behavior time series data, and then divide the dataset into training set, validation set, and test set according to a 6:2:2 ratio; the training set is used for model training; the validation set is used to test the network structure and adjust the model hyperparameters to determine whether overfitting occurs; the test set is used to test the generalization ability of the model. Step S2: Use the decomposition-based time series prediction model as the base model and train the base model using the training set. Decompose the time series data layer by layer through the residual structure to finally obtain the trend features and periodic features between different levels of user behavior time series data. Step S3: Train the basic model using the optimal meta-learning method. By summarizing the commonalities among different users, mining the model's prior knowledge, and obtaining robust initialization parameters, an initialization model applicable to all users is obtained. Step S4: When a new user joins, the initial model is fine-tuned using a small amount of data to obtain a user behavior habit model for the new user, which is used to achieve personalized recommendations for different users.

2. The personalized recommendation method based on optimization meta-learning according to claim 1, characterized in that, In step S1, the data preprocessing includes: performing linear normalization on the user behavior time series data to transform the original user behavior data to the range of [0,1] in order to balance the impact of different behaviors on the model; for user behavior time series data with default values ​​in some time periods, uniformly filling in the upper and lower data to ensure the consistency of the time granularity of the input data.

3. The personalized recommendation method based on optimization meta-learning according to claim 1, characterized in that, In step S2, the time series prediction model based on decomposition includes several components set sequentially from top to bottom. Stack The Stack Consisting of several arranged sequentially from top to bottom Block Composition, each Block It includes an encoder, a self-attention module, and two fully connected layers arranged in parallel from front to back; The encoder consists of several LSTM Unit composition, used to extract long and short time feature vectors from the input vector: ； The self-attention module is used to enable the model to pay attention to the correlation between different time points and to focus more on the key information in the input vector. The specific formula is as follows: ； ； in, Q , K , V These are the input sequences multiplied by the weight matrix, respectively. W q , W k , W v get, QK T To calculate the similarity between sample features and information features, w ( ) is the activation function, used to obtain the distribution of similarity or correlation between each sample and each piece of information. V It is a value vector matrix; The outputs of the two fully connected layers are as follows: ； The For the next Block The operation, the Used for addition operations Stack of Output, the Used for superposition operations to obtain prediction results.

4. The personalized recommendation method based on optimization meta-learning according to claim 1, characterized in that, In step S3, the optimal meta-learning method uses multi-objective optimization to balance multiple tasks during the meta-learning training process, assigning optimal weights to different tasks so that the model training process iterates along the optimal path to obtain the optimal path and result; the optimization objective is as follows: ； in, Indicates the first t The model loss after fine-tuning for each task The minimum sum of the losses from different tasks multiplied by their weights is transformed into the minimum norm of the matrix composed of the losses from different tasks.

5. The personalized recommendation method based on optimization meta-learning according to claim 4, characterized in that, The optimal meta-learning-based method in step S3 includes the following steps: Step A1: Input the meta-training task Meta-testing task Task learning rate α, meta-learning rate β, and memory learning rate γ are randomly initialized parameters. θ' ; Step A2: During meta-training, for the meta-training task, calculate the loss using the support set data of the training task. , Update parameter θ' using this loss: ； Calculate the loss using the query set data from the training task. ; Step A3: Calculate the gradient direction change for each task. ; The optimal weights for each task with respect to the initial parameters are calculated using an optimization algorithm. α 1 ,…, α T ; Update shared parameters using optimal weights: ； Step A4: For the meta-test task, calculate the loss using the support set data of that test task. Update parameters using this loss. θ' : ； Use the updated parameters Make predictions on the query set data of the test task.

6. The personalized recommendation method based on optimization meta-learning according to claim 5, characterized in that, Step A3 uses an optimization algorithm to calculate the optimal weights for each task with respect to the initial parameters, including the following steps: 1) Initialize the weight vector corresponding to the task. α 2) Calculate the relationship between the gradient descent directions ∇ for each task. M : ； ； ， ； in, e This means converting the original vector into a unit vector. γ represents the optimized path direction of the task weight vector. 3) When the maximum number of iterations is reached, output the optimal weights. α 1 ,…, α T .

7. The personalized recommendation method based on optimization meta-learning according to claim 4, characterized in that, The loss function used is the mean squared error function, which is as follows: ； in, For predicted values, y i This is the actual value.

8. A personalized recommendation system based on optimization meta-learning, used to implement the personalized recommendation method based on optimization meta-learning as described in claim 1; characterized in that, The system includes a data preprocessing module, a training module, and a personalization fine-tuning module. The data preprocessing module processes user behavior time-series data to obtain a training set, a validation set, and a test set. The training module trains a decomposition-based time-series prediction model using a Pareto optimal multi-objective optimization algorithm and meta-learning methods to obtain an initialization model. The personalization fine-tuning module uses data from specific users to personalize and fine-tune the initialization model to obtain a personalized user behavior habit model, thereby enabling personalized recommendations for specific users.

9. A storage medium storing computer program instructions, characterized in that, When the program instructions are executed by the processor, they implement the method described in any one of claims 1-7.

Images