Few-shot Irregular Time Series Prediction Method Based on Meta-learning Framework

Abstract

This invention discloses a few-shot irregular time series prediction method based on a meta-learning framework, comprising: acquiring data, segmenting time windows, random sampling, establishing a self-trained bidirectional recurrent model, training and fine-tuning the model using a model-independent meta-learning method, evaluating prediction results, and visualization. This invention utilizes a self-trained bidirectional recurrent neural network model strategy and generalization tuning through a model-independent meta-learning method to fine-tune the model by gradients, enabling its optimized weight configuration to be transferred to new tasks. This effectively improves the accuracy and efficiency of time series prediction in single-domain or single-group irregular multivariate few-shot scenarios while maintaining low computational cost.

Description

Technical Field

[0001] This invention pertains to self-trained time series prediction techniques, specifically a few-sample irregular time series prediction method based on a meta-learning framework. Background Technology

[0002] Predicting small-sample irregular multivariate time series (IMTS) is crucial for many fields, including healthcare, smart industry, the Internet of Things, biomechanics, and climate science. While existing research addresses irregularities in time series using ordinary differential equations, the challenge of modeling the correlations between asynchronous irregular multivariate time series remains largely unexplored.

[0003] Few-shot learning refers to learning the underlying patterns in data from only a small number of training samples (e.g., 50) to extract the most useful information, find the optimal model parameters that minimize the loss function, and generalize the optimized model parameters to a larger dataset. Few-shot learning can significantly shorten the time required to build machine learning applications and is a low-cost solution to the problem of unavailable large datasets in practical work. Although time series forecasting has been extensively studied, most research focuses on multivariate time series with regular sampling and sufficient observation. Unlike conventional multivariate time series, modeling and analyzing irregular multivariate time series based on irregular sampling intervals and missing data is more challenging. The current challenge lies in how to perform regression or effective derivation from a few-shot irregular time series dataset generated from a single group of research subjects or a single scenario. For example, the paper [Xiaolin Zhang, Yunchao Wei, Yi Yang, Thomas S. Huang. Similarity-guided networks for one-time semantic segmentation. IEEE Transactions on Networks and Control, 2020, Vol. 50, No. 9, pp. 3855-3865] proposes using attention mechanisms and relationships between variables to guide the prediction of missing data; the paper [Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick. Mask R-CNN. Proceedings of the 2017 IEEE International Conference on Computer Vision, pp. 2980-2988] proposes using masks and recurrent convolutional neural networks to extract multivariate time series features; and the paper [Geng Chen, Shenlong Wang, Tianhong Li, Hui Li, Ming-Hsuan Yang. Incremental few-sample instance segmentation. Proceedings of the 2021 IEEE / CVF Conference on Computer Vision and Pattern Recognition, pp. 1185-1194] proposes using discriminative embedding vectors to encode and decode irregular multivariate time series for pattern analysis.

[0004] However, these existing technologies have at least the following problems:

[0005] 1. Irregularities in time series dependency modeling. Different time intervals between adjacent observations disrupt the consistent flow of time series data, making it difficult for classic time series forecasting models to accurately capture underlying temporal dynamics and dependencies.

[0006] 2. Asynchronicity in Modeling Correlation Between Time Series. Although there is always a considerable correlation between time series of different variables, observations between IMTS may be significantly misaligned in time due to irregular sampling or missing data. This asynchronicity complicates direct comparisons and correlations at specific time points and may obscure or distort the actual relationship between time series, thus posing a significant challenge to modeling the correlation between time series.

[0007] 3. Existing methods mainly focus on handling irregularities in time series based on the constant differential equation (ODE), failing to explicitly consider the key correlations between multiple series.

[0008] 4. Real-world data distribution exhibits a long-tail effect, making it difficult to generalize the same model to different domains or groups, thus reducing the robustness of the generalized model. Summary of the Invention

[0009] In view of the above, the present invention provides a few-sample irregular time series prediction method based on a meta-learning framework, which can improve the accuracy and efficiency of time series prediction in single-domain or single-group irregular multivariate few-sample scenarios, while maintaining a low computational cost.

[0010] A few-shot irregular time series prediction method based on a meta-learning framework includes the following steps:

[0011] (1) Obtain multivariate irregular time series data for model training and testing;

[0012] (2) Construct a bidirectional recursive time series prediction model based on BRNN (Bidirectional Recurrent Neural Network);

[0013] (3) Train the bidirectional recursive time series prediction model using multivariate irregular time series data;

[0014] (4) Randomly set a position mask for a small portion of the input data and perform self-training learning on the model.

[0015] Further, the specific implementation of step (1) is as follows: initialize a time window with a fixed unit span, randomly extract a small amount of multivariate irregular time series data within the time window. The time dimension and variable dimension of these time series data are not the same, that is, different data have different numbers of variables and different sampling intervals; then, after standardizing the extracted time series data, it is divided into training data, test data and validation data.

[0016] Furthermore, the specific implementation of step (2) is as follows:

[0017] S21: Divide the neuron state in a conventional recurrent neural network into two parts: a forward state and a backward state. There is no direct connection between the forward state and the backward state.

[0018] S22: Keep all previously input temporal data in the hidden state of the recurrent neural network. Process the temporal data input with irregular time intervals through the recursive hidden state. The activation of the hidden state at each time depends on the hidden state of the previous time.

[0019] S23: To reduce the gradient vanishing problem in recurrent neural network training, the recurrent units of gated recurrent neural networks are used for sequence modeling, and ReLU (corrected linear unit) is used as the recurrent activation function.

[0020] Furthermore, in step (3), a model-independent meta-learning method is used to train the bidirectional recursive time series prediction model. The specific implementation method is as follows:

[0021] S31: Create training and testing tasks, and split the training and testing data into vertically stacked data of the same size;

[0022] S32: The meta-learning method employs a dual-loop iterative fine-tuning of the model's weight parameters, including meta-learning model optimization and single-task optimization. The meta-learning model optimization part uses ADAM (Adaptive Moment Estimation Function), which calculates the current gradient weights based on the dataset and gradient step size of each task to fine-tune the weight parameters, and iterates within the number of steps.

[0023] S33: In the training task, obtain the model weight parameters to be optimized, iterate the loss function and the gradient of all weights and loss function changes in the batch multivariate learning loop, find the sensitive weight parameters, and optimize the model by minimizing the loss function.

[0024] S34: In the inner loop (single task) iterative computation of the model-independent meta-learning method, the stochastic gradient descent method is used to fine-tune all the newly configured weight parameters on the training task data. After fine-tuning, the model is reloaded and the loss function of the next batch is calculated.

[0025] S35: In the outer loop (cross-task) iterative computation of the model-independent meta-learning method, ADAM is used to adjust all newly configured weight parameters of the optimized model using the meta-learning optimization method on the test task data. After fine-tuning, the model is reloaded and the loss function of the next batch is calculated, and finally a new optimized model is generated.

[0026] Furthermore, in step S32, the learning rate used in the model weight parameter optimization process is 0.001, the gradient step size is 0.25, and the number of steps is 400.

[0027] Furthermore, the specific implementation of step (4) is as follows:

[0028] S41: Randomly set a mask for less than 40% of the training data, divide the unmasked data and the masked data into two datasets, and calculate the estimated value of the masked data according to the position of the masked data, with weights of 0.4 and 0.6 for the unmasked data before and after the mask, respectively.

[0029] S42: Use a bidirectional recursive time series prediction model to infer and generate predicted values ​​for the unmasked data. Compare the actual values ​​and predicted values ​​of the unmasked part to adjust and optimize the model weights.

[0030] S43: Use the optimized model to infer all training data, generate predicted values, compare the predicted values ​​and estimated values ​​of the mask part, and if the difference is less than the preset threshold, replace the estimated values ​​of the mask part with the predicted values ​​to form new training data. Repeat this process 5 times.

[0031] S44: Use the cyclically tuned model to generate predicted values ​​for the test data.

[0032] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method for predicting few-sample irregular time series based on a meta-learning framework.

[0033] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned method for predicting few-sample irregular time series based on a meta-learning framework.

[0034] This invention employs a self-trained bidirectional recurrent neural network (BRN) model strategy and a model-independent meta-learning method for generalization tuning. This allows for gradient-based fine-tuning of the model, enabling its optimized weight configuration to be transferred to new tasks. On one hand, the self-trained BRN model can iteratively tune training data with random partial masks. The BRN model receives data simultaneously from both the forward and reverse directions of time, minimizing the target loss function without requiring a delay layer to store future information of delayed outputs, unlike unidirectional recurrent neural networks. On the other hand, the model-independent meta-learning method splits the data into stacks of equal size according to the task. In the training task, adaptive moment estimation and stochastic gradient descent are used to optimize all trainable weights in a double-loop iterative computation. In each iteration, the loss function of the predicted value versus the actual value due to weight changes is calculated to find a weight configuration sensitive to the target loss function. This new weight configuration is then generalized to new tasks, achieving accurate, efficient, and low-cost training for small-sample, multivariate, irregular time series analysis. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the flow of the few-sample irregular time series prediction method based on the meta-learning framework of the present invention. Detailed Implementation

[0036] To describe the present invention in more detail, the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] This implementation provides a few-sample irregular time series prediction method based on a meta-learning framework, the specific implementation process of which is as follows: Figure 1 As shown:

[0038] Step S11: Obtain a small sample of time series data. Randomly sample irregular multivariate data within a time window. Irregular multivariate time series data are characterized by irregular time intervals within each variable and misaligned observations across the variable variables. These characteristics make it challenging to provide accurate predictions of irregular time series data for making accurate decisions and plans.

[0039] We consider a... n A small-sample, multivariate, irregular time series dataset consisting of [number] samples. , n <200, of which It is the first i One sample. In T A timestamp and U Given the sum of variables, each sample can be represented as containing ... M A set of observation tuples ,in , , These represent the timestamp, observation, and variable index, respectively. For multivariate irregular time series forecasting tasks, given the split timestamps... These samples were divided into backtracking windows. and prediction window Predictive queries By combining in the prediction window of j Based on the observation tuples, our goal is to learn a predictive model. So that a backtracking window is given and predictive queries As input, it accurately predicts the corresponding observations. :

[0040]

[0041] In this implementation, we used daily Beijing weather quality data from 2018 for testing, employing five variables: temperature, air pressure, humidity, wind direction, and wind speed. The training and test datasets had sample sizes of 50 and 10 respectively, providing continuous predictions for 5 days. The overall data missing rate was 18%, and the data was analyzed using interval... Standardized processing between them.

[0042] Step S12: Establish a bidirectional recurrent neural network.

[0043] Bidirectional recurrent neural networks (NRNNs), as an extension of recurrent neural networks, aim to overcome the limitations of conventional recurrent models. Conventional recurrent models use input information in a limited way, only including pre-set future frames; bidirectional recurrent models improve their structure by including not only pre-set future information but also estimations of historical information. This structure allows them to be trained using input information bidirectionally in both positive and negative time directions simultaneously.

[0044] S21: Input a small number of time series data and establish a bidirectional recursive time series prediction model. Specifically, the neuron states learned by the conventional recursive neural network are divided into two parts: one part is used for the positive time direction (positive state), and the other part is used for the negative part (backward state). There is no direct relationship between the forward state and the backward state.

[0045] Specifically, recurrent neural networks maintain representations of all previous inputs in their hidden states and can handle variable-length sequential inputs with recursive hidden states, where activation at each time step depends on the hidden state of the previous time step. This applies to a series of paired input / output vectors. The basic recurrent neural network activation is calculated as follows:

[0046]

[0047]

[0048] in: It is the hidden layer vector. It is the input weight matrix to the hidden layer. It is the hidden layer weight matrix. It is the output weight matrix of the hidden layer. and These are the hidden layer bias and the output layer bias. The formula shows that at the current time... T The hidden state depends on the previous time. T Hidden state on -1 T Input and hidden layers T The bias value in.

[0049] A bidirectional recurrent neural network consists of backward and forward states. Without the backward state, this structure becomes a unidirectional forward recurrent neural network; conversely, if the forward state is removed, the structure is oriented to a simple recurrent neural network with a backward time axis. Therefore, in a bidirectional recurrent neural network, both the forward and backward time directions can be input into the computation to minimize the objective function without delaying output generation.

[0050] T Future information is also important for prediction; conventional recurrent neural networks can delay this by a certain amount of time. M This can be partially achieved by using the output of the time frame, and then the data can be recovered from the previous time frame. T arrive M Future information for prediction T The output at that point. Practice has shown that to achieve optimal performance, M The size has been optimized. In other words, the delay layer recovery... M The size cannot be too large, otherwise the prediction performance will degrade; however, bidirectional recurrent neural networks can make predictions by infinitely using future information in both the positive and negative time directions.

[0051] S22: To reduce the gradient vanishing problem in recurrent neural network training, this embodiment utilizes the recurrent units of a gated recurrent neural network for sequence modeling.

[0052] Specifically, T The hidden layer above is made by Calculated.

[0053] S23: Although the activation function of a gated recurrent neural network is the hyperbolic tangent function, the recurrent activation of a gated recurrent neural network is a modified linear unit to adjust the independent variable. xThe activation is limited to between the maximum value and the threshold; for optimized architectures, this activation is simple and effective, specifically:

[0054] for ;for ;for In this embodiment, the threshold d The default value is 0.

[0055] Step S13: Based on this model, use the "model-independent meta-learning" algorithm to optimize and tune the parameters of the recursive model in batches to minimize the difference between the predicted and actual values, and select the model parameters that minimize the objective loss function.

[0056] We identify model parameters that are sensitive to task variations, so that small changes in these parameters significantly improve the loss function for any task extracted from the task distribution. By meta-learning, the parameters of any standard model can be learned as the loss changes along its gradient direction, preparing the model for rapid adaptation. This gradient-based learning rule enables rapid progress on new tasks extracted from the task distribution without overfitting. In fact, this approach is based on the assumption that some internal representations are more transferable than others; for example, a neural network might learn internal features broadly applicable to all tasks in the task distribution, rather than a single task.

[0057] Compared to methods such as autoregressive moving average, decision trees, or logistic regression, this implementation optimizes local weights through gradient tuning for small-sample, multivariate, irregular time series data before generalizing to a global model. Random sample experiments demonstrate the accuracy, robustness, and efficiency of this model. The specific implementation process of this step is as follows:

[0058] S31: Establish training and testing tasks, and split the training and testing data into equal-sized subsets and stack them vertically. Specifically, split the training and testing data into subsets of equal size, stack the data according to the different labels of the training and testing task sets, and create an index.

[0059] S32: Adaptive moment estimation is used as the meta-learning optimization method. Specifically, the meta-learning method in this embodiment employs a double-loop iterative fine-tuning optimization of the bidirectional recursive learning model weights, including meta-learning model optimization and single-task optimization. The meta-learning model optimizer uses an adaptive moment estimation function, with an optimization learning rate of 0.001, a meta-learning gradient step size of 0.25, and a step count of 400. The current gradient weights are calculated and fine-tuned according to the task dataset and gradient step size, and the calculation is repeated within the number of steps.

[0060] S33: In the training task, obtain the model parameters to be optimized, iterate the loss function and the gradient of all trainable weights and the loss function in the batch multivariate learning loop, find the sensitive trainable weights, and optimize the model by minimizing the loss function.

[0061] Specifically, as shown in Algorithm 1, the gradient-based learning rule for the new task is fine-tuned using a "model-independent meta-learning" algorithm based on all trainable weights of the bidirectional recurrent neural network model. Therefore, our goal is to learn a model that allows this gradient-based learning rule to be applied to new tasks. Rapid progress is made on new tasks extracted from it without overfitting.

[0062] Algorithm 1:

[0063]

[0064] S34: In the inner loop (single task) iterative computation of the model-independent meta-learning method, all newly configured trainable weights are fine-tuned using stochastic gradient descent. After fine-tuning, the model is reloaded and the loss function for the next batch is calculated.

[0065] This step involves identifying parameters sensitive to task changes within the inner loop and fine-tuning them using the task optimizer's stochastic gradient descent method. This ensures that small changes in the parameters along the gradient direction of the loss will have a significant impact on the task. The loss function for any task being plotted yields significant improvements. We make no assumptions about the form of the model, only that it is derived from some parameter vector. i Parameterized, and the loss function is i The surface is smooth enough that we can use gradient-based learning techniques.

[0066] Specifically, we considered a parameter as i Parameterized functions The model is represented when it is adapted to a new task. At that time, for each iteration of the inner loop j Model parameters i Become In this embodiment, the updated parameter vector Is using the task The updates are computed using one or more gradient descent methods. For example, when using a gradient update:

[0067]

[0068] in: α It is the learning step size for the current task.

[0069] Through optimization The meta-objective for training model parameters is as follows:

[0070]

[0071] Meta-optimization is aimed at model parameters i The execution is performed, while the target model uses the updated model parameters. It is computationally induced. In fact, this implementation aims to optimize the model parameters so that one or a few gradient steps on a new task will produce the most effective behavior on that task.

[0072] S35: In the outer loop (cross-task) iterative computation of the model-independent meta-learning method, for the test task data, all newly configured trainable weights of the model after parameter optimization by adjusting the parameters using adaptive moment estimation are used; after fine-tuning, the model is reloaded and the loss function of the next batch is calculated, and finally a new optimized model is generated.

[0073] Specifically, cross-task meta-optimization is performed through stochastic gradient descent, which optimizes the model parameters. i The update is as follows:

[0074]

[0075] in: β It is the step length.

[0076] This invention employs a model-independent meta-learning method to train a bidirectional recurrent neural network model, enabling it to quickly adapt to new tasks using only a few data points and training iterations. This model can solve new recurrent learning tasks with minimal training samples. In this embodiment, the model parameters are explicitly trained; therefore, a small number of gradient steps and limited training data from the new task will produce good generalization performance on that task, accelerating the fine-tuning of the neural network policy gradient reinforcement learning model.

[0077] Step S14: Build a self-trained model.

[0078] For less than 40% of the input data, a position mask is randomly assigned. When processing time series data, this position mask is applied to the attention distribution. In the self-trained model, a given output in the sequence is computed using a subset of the input sequence that contributes to output prediction; that is, output prediction is improved by focusing on a subset of the input. This design alleviates the weakness of attention when modeling order information and fully leverages the advantages of parallel computing. The specific implementation process of this step is as follows:

[0079] S41: Initialize the learnable vectors, randomly set position masks for less than 40% of the vector data, and calculate the estimated value of the mask by weighted average method according to the sequence position.

[0080] Specifically, a mask is randomly set for less than 40% of the training data, and the unmasked data and the masked data are divided into two datasets. Based on the position of the masked data, the unmasked data before the masked data and the unmasked data after the masked data are weighted at 0.4 and 0.6 respectively to calculate the estimated value of the masked data.

[0081] S42: Use a bidirectional recurrent neural network model to train the unmasked vectors and generate predicted values. Compare the actual unmasked values ​​with the predicted values ​​and adjust and optimize the model weights.

[0082] S43: Use the trained model to train all training data (including masked and unmasked data), generate predicted values, compare the predicted values ​​of the masked part with the estimated values, and if the difference is less than the preset optimization threshold, replace the estimated values ​​of the masked part with the predicted values ​​to form new training data. Repeat this process 5 times.

[0083] S44: Use the iteratively tuned model to generate predicted values ​​for the test data.

[0084] Self-training learning, as a method of transferring knowledge from a labeled source domain to an unlabeled target, aims to progressively correct domain shifts by incorporating an increasing number of pseudo-labeled target examples into the source training set. This is particularly relevant for asymptotic domain adaptation, where unlabeled instances from intermediate domains can be used. Algorithm 2 shows a prediction model-based approach. The self-trained pseudocode algorithm first randomly accesses an unlabeled instance when making a selection; second, it trains a BRNN to generate pseudocode predictions (psvalues) while considering a confidence threshold and the test instance itself; third, it only makes a selection if the minimum difference between the pseudocode prediction and the estimate is less than the confidence threshold.

[0085] Algorithm 2:

[0086]

[0087] Verification example:

[0088] We conducted experiments using temperature data (unit: Fahrenheit) from a public database. Within a time window, we randomly selected 50 training data points, 10 validation data points, and 5 prediction data points as experimental data. For each algorithm, we used the same small-sample irregular time series data and ran it 50 times consecutively. As shown in Table 1, compared to autoregressive models, decision tree models, and random forest models, the method of this invention can reduce the mean absolute error (MAE) of the loss function by more than 50% and the mean squared error by more than 70%.

[0089] Table 1

[0090]

[0091] The above description of the embodiments is provided to enable those skilled in the art to understand and apply the present invention. Those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made to the present invention by those skilled in the art based on the disclosure thereof should be within the scope of protection of the present invention.

Claims

1. A method for predicting irregular time series with few samples based on a meta-learning framework, characterized in that, Includes the following steps: (1) Obtain multivariate irregular time series data for model training and testing. The specific implementation method is as follows: initialize a time window with a fixed unit span, randomly extract a small amount of multivariate irregular time series data within the time window. The time dimension and variable dimension of these time series data are not the same, that is, different data have different numbers of variables and different sampling intervals. Then, after standardizing the extracted time series data, it is divided into training data, test data and validation data. The multivariate irregular time series data is weather quality data, and the variables are temperature, air pressure, humidity, wind direction and wind speed. (2) Construct a bidirectional recursive time series prediction model based on BRNN, and the specific implementation method is as follows: S21: Divide the neuron state in a conventional recurrent neural network into two parts: a forward state and a backward state. There is no direct connection between the forward state and the backward state. S22: Keep all previously input temporal data in the hidden state of the recurrent neural network. Process the temporal data input with irregular time intervals through the recursive hidden state. The activation of the hidden state at each time depends on the hidden state of the previous time. S23: To reduce the gradient vanishing problem in recurrent neural network training, the recurrent units of gated recurrent neural networks are used for sequence modeling, and ReLU is used as the recurrent activation function. (3) The bidirectional recursive time series prediction model is trained using multivariate irregular time series data and a model-independent meta-learning method. The specific implementation method is as follows: S31: Create training and testing tasks, and split the training and testing data into vertically stacked data of the same size; S32: The meta-learning method uses a dual-loop iterative fine-tuning of the model's weight parameters, including meta-learning model optimization and single-task optimization. The meta-learning model optimization part uses ADAM, which calculates the current gradient weights based on the dataset and gradient step size of each task to fine-tune the weight parameters, and iterates within the number of steps. S33: In the training task, obtain the model weight parameters to be optimized, iterate the loss function and the gradient of all weights and loss function changes in the batch multivariate learning loop, find the sensitive weight parameters, and optimize the model by minimizing the loss function. S34: In the inner loop iterative calculation of the model-independent meta-learning method, the stochastic gradient descent method is used to fine-tune all the newly configured weight parameters on the training task data. After fine-tuning, the model is reloaded and the loss function of the next batch is calculated. S35: In the outer loop iterative calculation of the model-independent meta-learning method, ADAM is used to adjust all the newly configured weight parameters of the optimized model using the meta-learning optimization method on the test task data. After fine-tuning, the model is reloaded and the loss function of the next batch is calculated, and finally a new optimized model is generated. (4) Randomly set a position mask for a small portion of the input data and perform self-training learning on the model. The specific implementation method is as follows: S41: Randomly set a mask for less than 40% of the training data, divide the unmasked data and the masked data into two datasets, and calculate the estimated value of the masked data according to the position of the masked data, with weights of 0.4 and 0.6 for the unmasked data before and after the mask, respectively. S42: Use a bidirectional recursive time series prediction model to infer and generate predicted values ​​for the unmasked data. Compare the actual values ​​and predicted values ​​of the unmasked part to adjust and optimize the model weights. S43: Use the optimized model to infer all training data, generate predicted values, compare the predicted values ​​and estimated values ​​of the mask part, and if the difference is less than the preset threshold, replace the estimated values ​​of the mask part with the predicted values ​​to form new training data. Repeat this process 5 times. S44: Use the cyclically tuned model to generate predicted values ​​for the test data.

2. The method for predicting irregular time series with few samples based on a meta-learning framework according to claim 1, characterized in that: In step S32, the learning rate used in the model weight parameter optimization process is 0.001, the gradient step size is 0.25, and the number of steps is 400.

3. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: The processor is used to execute the computer program to implement the few-sample irregular time series prediction method based on the meta-learning framework as described in any one of claims 1 to 2.

4. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the few-sample irregular time series prediction method based on the meta-learning framework as described in any one of claims 1 to 2.

Images