Multi-data mixed exploration method based on deep learning

Abstract

The invention provides a multi-data mixed exploration method based on deep learning, and the method comprises the steps: evaluating a data scale according to the sampling coverage and sample capacity of different data sources, and judging whether the data sources can support the construction of a model or not; if the data scale cannot meet the requirement, expanding sample distribution through a generative adversarial network; evaluating the quality of the data source based on a statistical test or priori knowledge, and determining a quality weight; constructing a sub-network capable of reflecting the quality of the data source, obtaining a quality score, and updating a quality weight; evaluating a data source matching degree through correlation analysis, and determining a sample and feature selection; extracting various features of the heterogeneous data source, and expressing a data relationship; representing a learning network by using part of marked data features, and evaluating the discrimination performance of the features; task subspaces corresponding to different data sources are learned, joint reasoning is achieved, and a final decision is obtained.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a multi-data mixed exploration method based on deep learning. Background Art

[0002] The uneven size and quality of data, coupled with limited annotation resources, make building data-driven decision-making systems extremely challenging. These issues pose a challenge to multi-source, heterogeneous data fusion applications, necessitating addressing key technical challenges such as data imbalance and quality insecurity. In real-world applications, data sources are often diverse, including data from various platforms, devices, and users. The size and quality of this data vary significantly, and some may contain noise, missing data, or errors. Furthermore, due to limited annotation resources, much data may not be accurately labeled, resulting in uncertain data quality. These issues pose challenges to data-driven decision-making systems. For example, in machine learning tasks, data imbalance can cause models to favor predictions for more populated categories while neglecting less populated ones. Furthermore, low-quality data can introduce noise, impacting model performance. Therefore, addressing these issues is crucial to building effective decision-making systems. Summary of the Invention

[0003] The present invention provides a multi-data mixed exploration method based on deep learning, which mainly includes:

[0004] According to the sampling coverage and sample capacity of different data sources, the data scale is evaluated to determine whether the data source can support model construction; if the data scale cannot meet the requirements, the sample distribution is expanded through generative adversarial networks; based on statistical tests or prior knowledge, the quality of the data source is evaluated and the quality weight is determined; a sub-network that can reflect the quality of the data source is constructed to obtain the quality score and update the quality weight; the matching degree of the data source is evaluated through association analysis to determine the sample and feature selection; the diverse features of heterogeneous data sources are extracted to express data relationships; the learning network is represented by partially labeled data features to evaluate the discriminative performance of the features; the task subspaces corresponding to different data sources are learned to achieve joint reasoning and obtain the final decision.

[0005] Preferably, the step of evaluating the data scale based on the sampling coverage and sample capacity of different data sources and determining whether the data source can support model building includes:

[0006] If different data source types include structured data, semi-structured data and unstructured data, data structure information is obtained according to the summary, semantics and representation of the data source to determine the sampling coverage; the number of samples is calculated to determine the sample capacity; the completeness rate of the sample labels is checked to evaluate the label quality; the feature dimensions in the samples are summarized to obtain the feature dimensions; the recurrence ratio of the samples is counted to obtain the sample repetition rate; a sample association graph is constructed to analyze the degree of association between the samples; the samples are input into the denoising autoencoder to obtain the reconstruction error to determine the degree of sample noise; if the number of samples is less than the minimum sample size required for model training, it is determined that the data scale cannot meet the model construction requirements; if the sample label completeness rate is lower than the preset completeness rate threshold, it is determined that the sample label quality cannot meet the model construction requirements; if the sample repetition rate is higher than the preset repetition rate threshold, it is determined that the sample scale can support model construction after expansion; if the sample correlation is lower than the preset correlation threshold and the sample noise level is higher than the preset noise level threshold, it is determined that the data scale can support model construction after expansion.

[0007] Preferably, if the data scale cannot meet the requirements, the sample distribution is expanded by generating an adversarial network, including:

[0008] If the existing sample size cannot meet the model training requirements, the category distribution of the samples is statistically analyzed to evaluate the degree of imbalance in the sample categories; the characteristic dimension of the samples is calculated to determine the sample complexity; a sample association graph is constructed to analyze the sample correlation; if the sample categories are evenly distributed and the characteristic dimension is lower than the preset characteristic dimension, a latent variable generative adversarial network is constructed to generate new samples through random noise; if the sample correlation is lower than the preset sample correlation, a conditional generative adversarial network is constructed to generate new samples through category conditions and some real samples; if the sample categories are unevenly distributed and the complexity is higher than the preset complexity, the sample size and number of augmentations required for each category are determined by the degree of category imbalance; the conditional entropy of the sample features is calculated to determine the characteristic dimension that needs to be augmented; The variable generative adversarial network or conditional generative adversarial network uses Wasserstein distance or Jensen-Shannon distance as the loss function of the generator and discriminator, and the optimization algorithm selects Adam or RMSProp adaptive learning rate algorithm; sets the parameter space of the latent variable generative adversarial network or conditional generative adversarial network, and gradually augments the samples through several rounds of iteration; after convergence, the generated samples are filtered based on the surrounding density and class cohesion; if the generated samples meet the requirements, the generated samples and real samples are integrated through joint training and distillation strategies to obtain an enhanced dataset; the data augmentation effect is evaluated through quantity, quality and distribution indicators; if the effect does not meet the requirements, the network structure and loss function are adjusted and the augmentation is performed again.

[0009] Preferably, the step of assessing the quality of the data source based on statistical testing or prior knowledge and determining the quality weight includes:

[0010] Count the sample size of each data source and determine the sample capacity; calculate the duplicate sample ratio based on the sample ID to test for sample duplication; compare the representation consistency of the same sample in different data sources and determine the representation deviation; test the accuracy of the sample label and evaluate the label quality; if the quality problem is related to the sample itself, set the sample weight to reduce the impact of the sample; if the quality problem is related to the collection and representation process, set the data source weight to reduce the role of the data source; based on the evaluation results, determine the quality weight, including the sample weight vector and the data source weight parameter.

[0011] Preferably, the constructing of a subnetwork that can reflect the quality of a data source, obtaining a quality score, and updating a quality weight includes:

[0012] Obtain labeled data sets corresponding to different quality indicators as quality labels, including sample noise, sample repetition rate and feature correlation; construct a quality assessment neural network, with the input being the sample features of the data source and the output being the quality score; the quality assessment neural network structure adopts a convolutional neural network with an attention mechanism to obtain local patterns of sample quality; the quality assessment neural network training is supervised by quality labels, and the loss function incorporates cross entropy and smooth L1 regularization, and the parameters are updated iteratively through Adam optimization; construct several sub-networks, each predicting a quality indicator; track the quality score output by the quality assessment neural network, and test the correlation with the true quality label. If the correlation decreases, increase the number of quality assessment neural network layers; each time a new sample is added, use the quality assessment neural network to predict the quality of the new sample, dynamically adjust the quality weight, and provide feedback to adjust the quality assessment neural network parameters.

[0013] Preferably, the step of evaluating the matching degree of data sources by association analysis and determining sample and feature selection includes:

[0014] In different data sources, the correspondence between samples is identified based on the primary key field; for the feature representation of the same entity in different sources, the representation deviation is calculated through string comparison to verify the consistency of features; sentence vectors are used to express samples and features, and the cosine similarity between vectors is calculated to determine the degree of matching; according to the proportion of detected associated anomalies, the weight of problem samples is set to be reduced; according to the degree of feature redundancy, the weight is set to be reduced; a sample analysis diagram based on co-occurrence and association rules is constructed to analyze the matching and association patterns between samples from different sources; according to the matching confidence of the samples, samples whose matching effects meet the preset standards are selected to construct a cross-source training dataset; according to the importance of the features, several features with the highest mutual information are selected to form a feature space.

[0015] Preferably, the extraction of diverse features of heterogeneous data sources and expression of data relationships includes:

[0016] Analyze the summaries of different data sources, construct a cross-domain joint feature space, and express the heterogeneous attributes of the subjects between sources; use the autoencoder network to learn the nonlinear mapping of samples between data sources, and align the features of different data sources to a unified representation; establish the association between different source features based on the subject identification, construct the association graph of heterogeneous features, and represent the connection between data sources; use one-hot encoding category features and Word2Vec to express text and image features, and uniformly convert them into vector representations; calculate the mutual information and information gain of features, perform feature selection from different data sources, and obtain complementary feature subsets; use the multi-layer perceptron algorithm, input the obtained feature subset, and obtain the corresponding relationship between data sources.

[0017] Preferably, the method of using the partially labeled data feature representation to learn the network and evaluate the discriminative performance of the feature includes:

[0018] Construct a representation learning network, adopt the BERT structure with cross-layer attention mechanism, input heterogeneous features, and adjust parameters to fit downstream business; prepare some labeled samples as the label set, and the discriminative effect of the labeled sample features on the business results as supervision of the discriminator; during the training process of the representation learning network, use the label set supervision to adjust the network parameters; use the adjusted network output features to input the discriminator to evaluate the business prediction performance of different features and perform feature selection; select the feature subset with the best discriminative performance to retrain the representation learning network and discriminator to achieve joint representation learning and discrimination; deploy the representation learning network online, use new samples to iteratively update, and dynamically adjust feature extraction according to discriminant feedback; evaluate the business prediction performance of different source features, further cross-source sample and feature selection, and construct a discriminative performance feature space.

[0019] Preferably, the learning of task subspaces corresponding to different data sources to achieve joint reasoning and obtain a final decision includes:

[0020] Analyze the input and output requirements of the task and decompose the task into several subtasks in a fine-grained manner according to the characteristics of different data sources. For different subtasks, construct the corresponding subtask neural network as a sub-model to learn the corresponding data source to represent the corresponding task subspace. Sub-model training adopts a joint optimization strategy, adding collaborative constraints or associated regularization terms to the sub-model output to achieve collaborative representation learning of the subspace. Deploy the sub-model online and dynamically adjust the model according to new samples to achieve adaptive adjustment of the subspace. Construct a meta-model for the output of the sub-model and use attention to fuse the decisions of different sub-tasks to obtain the predicted output of the overall task. During the joint reasoning process, evaluate the contribution of each sub-model to the decision and dynamically adjust the decision weight of the sub-model. Implement knowledge transfer by weakening the task subspace with the lowest sub-model relevance. Perform multiple rounds of iterative optimization on each sub-model and meta-model to select the model combination with the best decision effect to complete joint decision-making based on heterogeneous sources.

[0021] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0022] The present invention discloses a method for mixed exploration of multiple data based on deep learning. In order to reflect the quality of the data source, a subnetwork is constructed to obtain the quality score and update the quality weight. The matching degree of different data sources is evaluated by association analysis to determine the sample and feature selection. On this basis, the diverse features of the heterogeneous data sources are extracted, and the relationship between the data is expressed. At the same time, the learning network is represented by partially labeled data features, and the discriminant performance of the features is evaluated. By learning the task subspaces corresponding to different data sources, joint reasoning is achieved. Ultimately, all the technologies can be integrated according to the above content to obtain a method that can effectively build a model and generate accurate decisions. This method can make full use of information from multiple data sources to improve the performance and accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flowchart of a multi-data mixed exploration method based on deep learning of the present invention.

[0024] Figure 2 This is a schematic diagram of a multi-data mixed exploration method based on deep learning of the present invention. DETAILED DESCRIPTION

[0025] To further understand the content of the present invention, the present invention is described in detail with reference to the accompanying drawings and examples. The present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the invention are shown in the accompanying drawings.

[0026] In this embodiment, a multi-data mixed exploration method based on deep learning may specifically include:

[0027] Step S101: Evaluate the data scale based on the sampling coverage and sample capacity of different data sources to determine whether the data source can support model construction.

[0028] If the different data sources include structured, semi-structured, and unstructured data, the data structure information is obtained based on the data source's summary, semantics, and representation to determine sampling coverage; the number of samples is calculated to determine sample capacity; the completeness of sample labels is checked to assess label quality; the characteristic dimensions in the samples are summarized to obtain characteristic dimensions; the proportion of repeated samples is calculated to obtain the sample repetition rate; a sample association graph is constructed to analyze the degree of association between samples; the samples are input into a denoising autoencoder to obtain the reconstruction error and determine the degree of sample noise. If the number of samples is less than the minimum sample size required for model training, the data size is determined to be insufficient for model building; if the sample label completeness is lower than the preset completeness threshold, the sample label quality is determined to be insufficient for model building; if the sample repetition rate is higher than the preset repetition rate threshold, the sample size is determined to be sufficient for model building after expansion. If the sample correlation is lower than the preset correlation threshold and the sample noise level is higher than the preset noise threshold, the data size is determined to be sufficient for model building after expansion.

[0029] For example, consider user data from an e-commerce platform, including purchase records, browsing histories, and user personal information. Based on the data source's abstract, semantics, and representation, we can derive data structure information. For example, purchase records include structured data such as product name, price, and purchase date; browsing records may be semi-structured data, including information such as the product category and time of browsing; and user personal information may be unstructured data, including the user's name, address, and phone number. First, we can calculate the number of samples to determine sample capacity. For example, consider data from 1,000 users, with 800 purchase record samples, 900 browsing record samples, and 1,000 user personal information samples. Next, we can assess label quality by examining the completeness of sample labels. To analyze user purchasing behavior, the labels in purchase records are the categories of purchased items. We can calculate the number of complete label samples in the purchase records. If 700 samples have complete labels, the completeness ratio is 700 / 800 = 0.875. We can then aggregate the feature dimensions in the samples to obtain the feature dimension. The feature dimensions in purchase records include product name, price, and purchase time; the feature dimensions in browsing records include product category and time; and the feature dimensions in user personal information include name, address, and phone number. Therefore, the total number of feature dimensions is 3 + 2 + 3 = 8. We can also calculate the sample repetition rate by calculating the recurrence ratio of samples. If 100 samples in the purchase record are repeated, the recurrence ratio is 100 / 800 = 0.125. We can construct a sample association graph to analyze the degree of association between samples. We construct an association graph based on the product names in the purchase records. We then analyze the connections between nodes to assess the degree of association between samples. Finally, we can input the samples into a denoising autoencoder and obtain the reconstruction error to determine the degree of sample noise. Inputting the purchase records into the denoising autoencoder yields a reconstruction error of 0.05.

[0030] Step S102: If the data scale cannot meet the requirements, the sample distribution is expanded through the generative adversarial network.

[0031] If the existing sample size does not meet the model training requirements, the sample category distribution is statistically analyzed to assess the degree of class imbalance; the sample feature dimensions are calculated to determine sample complexity; and a sample association graph is constructed to analyze sample correlation. If the sample category distribution is uniform but the feature dimensions are lower than the preset feature dimensions, a latent variable generative adversarial network is constructed to generate new samples using random noise. If the sample correlation is lower than the preset sample correlation, a conditional generative adversarial network is constructed to generate new samples using category conditions and some real samples. If the sample category distribution is uneven and the complexity is higher than the preset complexity, the degree of class imbalance is used to determine the number of samples to be augmented for each category and the number of augmentations. The conditional entropy of the sample features is calculated to determine the feature dimensions to be augmented. For the constructed latent variable generative adversarial network or conditional generative adversarial network, the Wasserstein distance or Jensen-Shannon distance is used as the loss function for the generator and discriminator, and the Adam or RMSProp adaptive learning rate algorithm is used as the optimization algorithm. The parameter space of a latent variable generative adversarial network or a conditional generative adversarial network is set, and samples are gradually augmented through several rounds of iteration. After convergence, the generated samples are filtered based on surrounding density and cluster cohesion. If the generated samples meet the requirements, the generated samples and real samples are integrated through joint training and distillation strategies to obtain an augmented dataset. The effectiveness of data augmentation is evaluated using quantity, quality, and distribution metrics. If the results are not satisfactory, the network structure and loss function are adjusted and augmentation is repeated.

[0032] For example, consider a classification task requiring training a model to distinguish between spam and ham emails. However, there are only 1,000 samples, insufficient for training. Statistical analysis of the sample class distribution reveals 700 ham email samples and 300 spam email samples, indicating class imbalance. Calculating the sample feature dimensions reveals 100 features per sample, determining the sample complexity. A sample association graph is constructed to analyze the correlation between samples. Based on these statistical and analytical results, we can conclude that the sample class distribution is uneven and the feature dimensions exceed the preset feature dimensions. Therefore, we need to increase the number of spam samples, increase the number of augmentations, and increase the feature dimensions. A latent variable generative adversarial network can be constructed to generate new samples, generating spam samples using random noise. Since the sample correlation is low, a conditional generative adversarial network can be constructed to generate new samples using class conditions and a subset of real samples. Next, the Wasserstein distance is selected as the loss function for the generator and discriminator, and the Adam optimization algorithm is used to train the latent variable generative adversarial network or the conditional generative adversarial network. The network parameter space is set, and the sample augmentation is performed over several rounds of iterations. After convergence, the generated samples can be filtered based on surrounding density and class cohesion to remove low-quality samples. If the generated samples meet the requirements, they can be combined with real samples through joint training and distillation strategies to obtain an augmented dataset. Finally, the effectiveness of data augmentation can be evaluated using quantity, quality, and distribution metrics. If the results are not as expected, the network structure and loss function can be adjusted and augmentation can be repeated.

[0033] Step S103: Evaluate the quality of the data source based on statistical tests or prior knowledge and determine the quality weight.

[0034] Count the sample size of each data source to determine sample capacity; calculate the duplicate sample ratio based on the sample ID to test for sample duplication. Compare the representation consistency of the same sample across different data sources to identify representation bias; verify the accuracy of sample labels and assess label quality. If quality issues are related to the sample itself, assign sample weights to mitigate their impact; if quality issues are related to the collection and representation process, assign data source weights to mitigate the impact of the data source. Based on the assessment results, determine quality weights, including the sample weight vector and data source weight parameters.

[0035] For example, there are three data sources, A, B, and C, each containing a certain number of samples. We want to count the sample size of each data source and determine the sample capacity. Data source A has 100 samples, data source B has 200 samples, and data source C has 150 samples. The sample capacity is 100 + 200 + 150 = 450 samples. Next, we want to calculate the duplicate sample ratio using the sample ID to test for sample duplication. We found 10 duplicate samples in the three data sources. The duplicate sample ratio is 10 / 450 ≈ 0.022, or approximately 2%. We then compare the expression of the same sample across different data sources to determine representation consistency and identify representation bias. We selected 100 samples for comparison and found that the expression of 90 samples in data source A was consistent with that of the other data sources, while the expression of 85 and 88 samples in data sources B and C, respectively, was consistent with that of the other data sources. The consistency is (90+85+88) / (3*100)≈0.877, which is about 88%. Next, check the accuracy of the sample labels and evaluate the quality of the labels. There are 100 sample labels, of which 10 labels are determined to be wrong. The accuracy of the labels is (100-10) / 100=0.9, which is about 90%. If it is found that the quality problem is related to the sample itself, the sample weight can be set to reduce the impact of these samples; if it is found that the quality problem is related to the acquisition and representation process, the data source weight can be set to reduce the role of these data sources. Finally, based on the evaluation results, the sample weight vector and data source weight parameters can be determined to perform weighted processing on the samples and data sources to reduce the impact of quality problems.

[0036] Step S104: construct a sub-network that can reflect the quality of the data source, obtain the quality score, and update the quality weight.

[0037] A labeled dataset corresponding to different quality metrics is obtained as quality labels. These metrics include sample noise, sample repetition rate, and feature correlation. A quality assessment neural network is constructed, with the input being the sample features of the data source and the output being a quality score. The quality assessment neural network architecture uses a convolutional neural network with an attention mechanism to capture local patterns of sample quality. Quality labels are used to supervise the training of the quality assessment neural network. The loss function incorporates cross entropy and smoothed L1 regularization, and the parameters are updated iteratively through Adam optimization. Several subnetworks are constructed, each predicting a different quality metric. The quality scores output by the quality assessment neural network are tracked and their correlation with the true quality labels is tested. If the correlation decreases, the number of quality assessment neural network layers is increased. Each time a new sample is added, the quality of the new sample is predicted using the quality assessment neural network, and the quality weight is dynamically adjusted, with feedback provided to adjust the quality assessment neural network parameters.

[0038] For example, a neural network combining a CNN and an attention mechanism was constructed to assess data quality. The network inputs statistical features of data source samples, such as the proportion of null values ​​in each column, string length distribution, and value range cardinality, and outputs a quality score between 0 and 1. Three quality labels were assigned: high (0.8-1), medium (0.6-0.8), and low (0-0.6). 5,000 manually annotated high, medium, and low-quality samples were collected, with 4,000 used as the training set and 1,000 as the test set. Binary cross entropy was chosen as the loss function, and an L2 regularization term was added to control model complexity by using the square of the sum of parameter weights. The Adam optimization algorithm was used, with an initial learning rate of 0.001, which was gradually decayed through regression testing. After 1,000 epochs of training, the model achieved the best quality classification accuracy of 0.86 on the test set. To evaluate different quality metrics, subnetworks for sample duplication rate and missing rate were also constructed. The three sub-networks were fused through an integrated approach, averaging the quality confidence scores of each sub-network's output to generate a quality score for the fused model, improving classification accuracy to 0.91. A new quality correlation verification module was also added to track the Pearson correlation coefficient between the network's quality output and the true quality labels. When the correlation coefficient falls below 0.8, the number of neural units in the network's intermediate layers is increased, and retraining is performed to improve quality judgment. During the model prediction process, if the quality score of a new sample deviates significantly from the previous result, the sample is flagged to verify the stability of the network evaluation. Subsequently, data source weights are adjusted in real time based on prediction quality, selecting high-quality data sources to improve subsequent modeling results.

[0039] Step S105: Evaluate the matching degree of data sources through association analysis and determine sample and feature selection.

[0040] In different data sources, sample correspondences are identified based on primary key fields. For feature representations of the same entity across different sources, string comparison is used to calculate representation deviations and verify feature consistency. Sentence vectors are used to represent samples and features, and cosine similarity between vectors is calculated to determine the degree of matching. Based on the proportion of detected association anomalies, problem samples are weighted down; based on the degree of feature redundancy, weight reduction is also set. A sample analysis diagram based on co-occurrence and association rules is constructed to analyze matching and association patterns between samples from different sources. Based on the sample matching confidence, samples whose matching results meet preset standards are selected to construct a cross-source training dataset. Based on feature importance, several features with the highest mutual information are selected to form a feature space.

[0041] For example, consider two data sources, A and B, each with a primary key field used to identify sample correspondences. By comparing the feature representations of the same entity in the two data sources, representation deviation can be calculated and feature consistency can be verified. A feature named "Feature 1" from the two data sources was compared, and the string alignment result was 0.8, indicating a certain degree of deviation. Samples and features were converted into sentence vectors, and the cosine similarity between the vectors was calculated to determine their degree of match. The cosine similarity between a sample and a feature was 0.9, indicating a high degree of match. Based on the detected correlation anomaly ratio and feature redundancy, the weight of problematic samples can be reduced. If the correlation anomaly ratio is 10%, the weight of the problematic sample can be reduced to 90%. If the feature redundancy is 20%, the weight of the feature can be reduced to 80%. To analyze the matching and association patterns between samples from different sources, a sample analysis graph based on co-occurrence and association rules can be constructed. This graph reveals high co-occurrence between certain samples, indicating a certain degree of correlation. Based on the sample matching confidence, we select samples that meet the preset matching criteria to construct a cross-source training dataset. Finally, based on feature importance, we select several features with the highest mutual information to form a feature space. Three features, "Feature 4," "Feature 5," and "Feature 6," were selected, with mutual information values ​​of 6, 7, and 8, respectively, indicating increasing importance for the sample.

[0042] Step S106: extract diverse features of heterogeneous data sources and express data relationships.

[0043] Analyze the summaries of different data sources and construct a cross-domain joint feature space to express the heterogeneous attributes of entities across sources. Use an autoencoder network to learn nonlinear mappings between samples from different data sources, aligning features from different data sources into a unified representation. Establish associations between features from different sources based on entity identification, and construct a correlation graph of heterogeneous features to represent the connections between data sources. Use one-hot encoding for category features and Word2Vec for text and image features, converting them into a unified vector representation. Calculate mutual information and information gain for features, select features from different data sources, and obtain complementary feature subsets. Use a multi-layer perceptron algorithm as input for the obtained feature subsets to determine the corresponding relationships between data sources.

[0044] For example, consider two data sources: one containing a user's purchase history and the other containing their browsing history. To build a unified user representation by learning the relationship between these two data sources, we first need to convert the purchase and browsing history into feature vectors. For purchase history, we can use one-hot encoding to represent the categorical characteristics of each item. For browsing history, we can use the Word2Vec model to convert text descriptions into feature vectors. Next, we can use an autoencoder network to learn a nonlinear mapping between purchase and browsing history. This autoencoder network maps the high-dimensional feature space into a low-dimensional representation, preserving the structure of the original data as much as possible. Next, we can establish associations between purchase and browsing history based on the user's ID. We can use the user's ID as the subject identifier to build a correlation graph between purchase and browsing history. Nodes in the graph represent users, and edges represent connections between users across different data sources. To select the most representative feature subset, we can calculate mutual information and information gain. Mutual information measures the correlation between two features, while information gain measures the contribution of a feature to the target variable. By calculating these metrics, we can select the feature subset with the highest correlation and contribution. Finally, we can use a multilayer perceptron algorithm to establish the correspondence between purchase and browsing history. By inputting a feature subset, a multi-layer perceptron model can be used to predict the user's behavioral relationships between different data sources.

[0045] Step S107: Using the partially labeled data features to represent the learning network, the discriminative performance of the features is evaluated.

[0046] Build a representation learning network using the BERT architecture with a cross-layer attention mechanism, input heterogeneous features, and adjust parameters to fit downstream business needs. Prepare a set of labeled samples as a tag set. The discriminative effect of the labeled sample features on business outcomes serves as supervision for the discriminator. During the training of the representation learning network, use the tag set supervision to adjust the network parameters. Use the adjusted network output features as input to the discriminator to evaluate the business prediction performance of different features and perform feature selection. Select the feature subset with the best discriminative performance to retrain the representation learning network and discriminator, achieving joint representation learning and discrimination. Deploy the representation learning network online, iteratively update it with new samples, and dynamically adjust feature extraction based on discriminative feedback. Evaluate the business prediction performance of different source features, further cross-source sample and feature selection, and construct a discriminative performance feature space.

[0047] For example, a representation learning network is being built using the BERT architecture and a cross-layer attention mechanism. The input is heterogeneous features, and the network parameters are adjusted to fit the downstream business. To supervise the discriminator, a set of labeled samples is prepared, and the features of these samples are used to discriminate business outcomes. During the training of the representation learning network, the labeled set is used for supervision to adjust the network parameters. Then, the discriminator uses the adjusted network output features as input to evaluate the business prediction performance of different features and perform feature selection. The feature subset with the best discriminative performance is selected, and the representation learning network and discriminator are retrained to achieve joint representation learning and discrimination. When the representation learning network is deployed online, it is iteratively updated using new samples, and feature extraction is dynamically adjusted based on discriminative feedback. The business prediction performance of different source features is evaluated, and further cross-source sample and feature selection are performed to construct a discriminative feature space. The labeled set contains 100 samples, 80 of which are used to supervise the adjustment of network parameters. The remaining 20 samples are used to evaluate the business prediction performance of different features, and 10 features are selected as the best feature subset. The representation learning network and discriminator were then retrained, and the representation learning network was deployed online. Iterative updates were performed using 100 new samples, and feature extraction was dynamically adjusted based on discriminant feedback. Finally, the business prediction performance of different source features was evaluated, and the optimal feature subset was selected to construct a discriminative performance feature space.

[0048] Step S108: Learn the task subspaces corresponding to different data sources, implement joint reasoning, and obtain the final decision.

[0049] The input and output requirements of the task are analyzed and the task is fine-grainedly decomposed into several subtasks based on the characteristics of different data sources. For each subtask, a neural network is constructed as a sub-model to learn the task subspace corresponding to the corresponding data source representation. Sub-model training utilizes a joint optimization strategy, adding collaborative constraints or associated regularization terms to the sub-model outputs to achieve collaborative representation learning of the subspace. The sub-models are deployed online and dynamically adjusted based on new samples to achieve adaptive adjustment of the subspace. A meta-model is constructed based on the sub-model outputs, using attention to fuse the decisions of different sub-tasks to obtain the predicted output for the overall task. During the joint inference process, the contribution of each sub-model to the decision is evaluated, and the decision weights of the sub-models are dynamically adjusted. Knowledge transfer is achieved by weakening the task subspace with the lowest sub-model relevance. Multiple rounds of iterative optimization are performed on each sub-model and meta-model to select the model combination with the best decision-making effect, thus completing joint decision-making based on heterogeneous sources.

[0050] For example, one task involves predicting a user's purchase intention by analyzing their social media behavior. This task can be broken down into several subtasks. By analyzing a user's personal information, interests, hobbies, and friendships, a user profile can be constructed to represent the user's basic characteristics. By analyzing their social media posts, a user's emotional inclination can be determined, such as whether they express a preference for or dislike for a product. By analyzing the user's following and following relationships and social interactions on social media, a user's social influence and social circles can be inferred, thereby understanding the user's influence within that circle. By analyzing a user's ad click behavior on social media, a user's interest in the ad and purchase intention can be predicted. For each of these subtasks, corresponding neural network submodels can be constructed and trained. User profile analysis can use a multilayer perceptron model, text sentiment analysis can use a recurrent neural network model, social network analysis can use a graph neural network model, and ad click-through rate prediction can use a convolutional neural network model. When training the submodels, a joint optimization strategy can be employed, adding collaborative constraints or regularization terms to the submodel outputs to achieve collaborative representation learning in the subspace. A collaborative loss function can be introduced to minimize the difference between the different submodels so that they learn similar representations of the task subspace. When deploying sub-models online, they can be dynamically adjusted based on feedback from new samples to achieve adaptive subspace adjustments. When the feature distribution of new samples changes, the sub-model parameters can be updated through incremental learning to adapt to the new sample distribution. A meta-model can be constructed based on the sub-model outputs to fuse the decisions of different sub-tasks. An attention mechanism can be used to dynamically select the outputs of different sub-models and weight them according to their contributions to obtain the prediction output for the overall task. During joint inference, the contribution of each sub-model to the decision can be evaluated, and the decision weights of the sub-models can be dynamically adjusted based on their importance. A confidence assessment method can be used to measure the confidence of each sub-model in each decision, and trade-offs and adjustments can be made based on the confidence. Knowledge transfer can be achieved by weakening the task subspaces with the lowest relevance of the sub-models. Transfer learning can be used to select sub-models relevant only to the target task for parameter updates during training, thereby improving the model's generalization performance. Finally, through multiple rounds of iterative optimization of each sub-model and the meta-model, the optimal model combination can be selected to achieve joint decision-making based on heterogeneous sources. You can use the cross-validation method to conduct multiple experiments on the training set and validation set to select the optimal model combination.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-data mixed exploration method based on deep learning, characterized by: The method comprises: According to the sampling coverage and sample capacity of different data sources, the data scale is evaluated to determine whether the data source can support model construction; if the data scale cannot meet the requirements, the sample distribution is expanded through generative adversarial networks; based on statistical tests or prior knowledge, the quality of the data source is evaluated and the quality weight is determined; a sub-network that can reflect the quality of the data source is constructed to obtain the quality score and update the quality weight; the matching degree of the data source is evaluated through association analysis to determine the sample and feature selection; the diverse features of heterogeneous data sources are extracted to express data relationships; the learning network is represented by partially labeled data features to evaluate the discriminative performance of the features; the task subspaces corresponding to different data sources are learned to achieve joint reasoning and obtain the final decision.

2. The method according to claim 1, wherein The above mentioned process is to evaluate the data scale based on the sampling coverage and sample capacity of different data sources and determine whether the data source can support model construction, including: If different data sources include structured data, semi-structured data, and unstructured data, then the data structure information is obtained based on the summary, semantics, and representation of the data source to determine the sampling coverage; the number of samples is calculated to determine the sample capacity; the completeness of the sample labels is checked to evaluate the label quality; the feature dimensions in the samples are summarized to obtain the feature dimensions; Count the recurrence ratio of samples to obtain the sample repetition rate; construct a sample association graph to analyze the degree of association between samples; input samples into the denoising autoencoder to obtain the reconstruction error and judge the degree of sample noise; if the number of samples is less than the minimum sample size required for model training, it is judged that the data scale cannot meet the model construction requirements; if the sample label completeness rate is lower than the preset completeness rate threshold, it is judged that the sample label quality cannot meet the model construction requirements; if the sample repetition rate is higher than the preset repetition rate threshold, it is judged that the sample scale can support model construction after expansion; if the sample correlation is lower than the preset correlation threshold and the sample noise level is higher than the preset noise level threshold, it is judged that the data scale can support model construction after expansion.

3. The method according to claim 1, wherein If the data size cannot meet the requirements, the sample distribution is expanded through the generative adversarial network, including: If the existing sample size cannot meet the model training requirements, the category distribution of the samples is statistically analyzed to assess the degree of sample category imbalance; the characteristic dimension of the samples is calculated to determine the sample complexity; a sample association graph is constructed to analyze the sample correlation; if the sample category distribution is uniform but the characteristic dimension is lower than the preset characteristic dimension, a latent variable generative adversarial network is constructed to generate new samples through random noise; if the sample correlation is lower than the preset sample correlation, a conditional generative adversarial network is constructed to generate new samples through category conditions and some real samples; If the sample categories are unevenly distributed and the complexity is higher than the preset complexity, the number of samples that need to be augmented for each category and the number of augmentations are determined based on the degree of category imbalance; the conditional entropy of the sample features is calculated to determine the feature dimensions that need to be augmented; For the constructed latent variable generative adversarial network or conditional generative adversarial network, use Wasserstein distance or Jensen-Shannon distance as the loss function of the generator and discriminator, and select Adam or RMSProp adaptive learning rate algorithm as the optimization algorithm; The parameter space of a latent variable generative adversarial network or a conditional generative adversarial network is set, and samples are gradually augmented through several rounds of iteration. After convergence, the generated samples are filtered based on surrounding density and class cohesion. If the generated samples meet the requirements, the generated samples and real samples are integrated through joint training and distillation strategies to obtain an augmented dataset. The effectiveness of data augmentation is evaluated using quantity, quality, and distribution indicators. If the effect is not up to standard, adjust the network structure and loss function and perform augmentation again.

4. The method according to claim 1, wherein The method of assessing the quality of data sources based on statistical tests or prior knowledge and determining the quality weight includes: Count the sample size of each data source and determine the sample capacity; calculate the duplicate sample ratio based on the sample ID to test for sample duplication; compare the representation consistency of the same sample in different data sources and determine the representation deviation; test the accuracy of the sample label and evaluate the label quality; if the quality problem is related to the sample itself, set the sample weight to reduce the impact of the sample; if the quality problem is related to the collection and representation process, set the data source weight to reduce the role of the data source; based on the evaluation results, determine the quality weight, including the sample weight vector and the data source weight parameter.

5. The method according to claim 1, wherein The construction of a subnetwork that can reflect the quality of the data source, obtaining a quality score, and updating a quality weight include: Obtain labeled data sets corresponding to different quality indicators as quality labels, including sample noise, sample repetition rate and feature correlation; construct a quality assessment neural network, with the input being the sample features of the data source and the output being the quality score; the quality assessment neural network structure adopts a convolutional neural network with an attention mechanism to obtain local patterns of sample quality; the quality assessment neural network training is supervised by quality labels, and the loss function incorporates cross entropy and smooth L1 regularization, and the parameters are updated iteratively through Adam optimization; construct several sub-networks, each predicting a quality indicator; track the quality score output by the quality assessment neural network, and test the correlation with the true quality label. If the correlation decreases, increase the number of quality assessment neural network layers; each time a new sample is added, use the quality assessment neural network to predict the quality of the new sample, dynamically adjust the quality weight, and provide feedback to adjust the quality assessment neural network parameters.

6. The method according to claim 1, wherein The method of evaluating the matching degree of data sources through association analysis and determining sample and feature selection includes: In different data sources, the corresponding relationship between samples is identified based on the primary key field. For the feature representation of the same entity in different sources, the deviation is calculated through string comparison to verify the consistency of the feature. Sentence vectors are used to express samples and features, and the cosine similarity between vectors is calculated to determine the degree of matching. The weight of problem samples is set to be reduced based on the detected proportion of associated anomalies. According to the degree of feature redundancy, set the weight to be reduced; build a sample analysis diagram based on co-occurrence and association rules to analyze the matching and association patterns between samples from different sources; according to the matching confidence of the samples, select the samples whose matching effect meets the preset standards and build a cross-source training dataset; According to the importance of features, several features with the highest mutual information are selected to form a feature space.

7. The method according to claim 1, wherein The method of extracting diverse features of heterogeneous data sources and expressing data relationships includes: Analyze the summaries of different data sources, construct a cross-domain joint feature space, and express the heterogeneous attributes of the subjects between sources; use the autoencoder network to learn the nonlinear mapping of samples between data sources, and align the features of different data sources to a unified representation; establish the association between different source features based on the subject identification, construct the association graph of heterogeneous features, and represent the connection between data sources; use one-hot encoding category features and Word2Vec to express text and image features, and uniformly convert them into vector representations; calculate the mutual information and information gain of features, perform feature selection from different data sources, and obtain complementary feature subsets; use the multi-layer perceptron algorithm, input the obtained feature subset, and obtain the corresponding relationship between data sources.

8. The method according to claim 1, wherein The method of using the partially labeled data feature representation to learn the network and evaluate the discriminative performance of the feature includes: Build a representation learning network, adopt the BERT structure with a cross-layer attention mechanism, input heterogeneous features, and adjust parameters to fit downstream business; Prepare some labeled samples as a label set, and label the discriminative effect of the sample features on the business results as supervision for the discriminator; During the training process of the representation learning network, the labeled set supervision is used to adjust the network parameters; Use the adjusted network output features and input them into the discriminator to evaluate the business prediction performance of different features and perform feature selection; Select the feature subset with the best discriminative performance to retrain the representation learning network and discriminator to achieve joint representation learning and discrimination; Deploy the representation learning network online, iteratively update it with new samples, and dynamically adjust feature extraction based on discriminant feedback; evaluate the business prediction performance of different source features, further cross-source sample and feature selection, and construct a discriminant performance feature space.

9. The method according to claim 1, wherein The task subspaces corresponding to different data sources are learned to achieve joint reasoning and obtain the final decision, including: Analyze the input and output requirements of the task and decompose it into several subtasks based on the characteristics of different data sources. Construct a neural network for each subtask as a submodel to learn the task subspace corresponding to the data source representation. Submodel training uses a joint optimization strategy, adding collaborative constraints or associated regularization terms to the submodel output to achieve collaborative representation learning of the subspace. Deploy sub-models online and dynamically adjust them based on new samples to achieve adaptive adjustment of the subspace. Construct a meta-model based on the outputs of the sub-models, using attention to fuse the decisions of different sub-tasks to obtain the predicted output of the overall task. During the joint inference process, evaluate the contribution of each sub-model to the decision and dynamically adjust the decision weights of the sub-models. Knowledge transfer is achieved by weakening the task subspace with the lowest correlation among sub-models; Multiple rounds of iterative optimization of each sub-model and meta-model are performed to select the model combination with the best decision-making effect and complete joint decision-making based on heterogeneous sources.

Images