Enterprise multi-type data labeling method and system based on feature engineering

Abstract

The invention belongs to the technical field of data processing, and discloses an enterprise multi-type data labeling method and system based on feature engineering, and the method comprises the steps: collecting initial data of an enterprise internal information system, and dividing the initial data to obtain a text data set, a table data set and an image data set; feature vectors of different types of data are fused after features are extracted and screened, and a multi-dimensional feature space is constructed; constructing an annotation model according to the feature information of the feature space, constructing similar sample pairs as positive sample pairs and irrelevant sample pairs as negative sample pairs for unannotated new data, and comparing, learning and optimizing the similarity between the similar sample pairs; and carrying out label inference based on the feature vectors obtained by comparative learning, and generating a labeling result. According to the method, efficient and automatic labeling of various types of data can be achieved, the method is suitable for various heterogeneous data such as texts, tables and images, and high labeling precision and expansibility are achieved.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and specifically relates to a method and system for labeling multi-type enterprise data based on feature engineering. Background Art

[0002] In the course of daily operations, enterprises generate a large amount of multi-type data, such as text data (contracts, reports), table data (financial statements, sales data), and image data (scanned documents, photos). These data are usually not uniformly labeled in their original state, making them difficult to use directly for subsequent analysis and modeling. Traditional manual labeling methods are inefficient and cannot cope with large-scale data processing needs. Therefore, automated data labeling methods are of great significance for efficient data processing in enterprises.

[0003] Some existing automated data annotation methods usually only process a single type of data, and it is difficult to process multiple types of data at the same time, and there are deficiencies in annotation accuracy and adaptability. Therefore, how to efficiently and accurately automate the multi-type data generated by enterprises in their daily operations to better meet the actual multi-type data processing needs of enterprises has become one of the problems that need to be solved urgently. Summary of the invention

[0004] In response to the above problems, the present invention provides a method and system for enterprise multi-type data labeling based on feature engineering, which can realize efficient and automatic labeling of various types of data, is suitable for various heterogeneous data such as text, tables, images, etc., and has high labeling accuracy and scalability.

[0005] The present invention provides a method for labeling enterprise multi-type data based on feature engineering, comprising the following steps: S10, collecting initial data from the enterprise's internal information system, and classifying it according to data type to obtain a text data set, a table data set, and an image data set; S20, extracting and screening features of different data types of the initial data respectively, fusing feature vectors of different types of data, and constructing a multi-dimensional feature space; S30, constructing a labeling model according to the feature information of the feature space, training the labeling model through self-supervised learning to preliminarily learn the intrinsic structure and feature representation of the data, and fine-tuning the model in combination with the labeled data; S40, construct similar sample pairs as positive sample pairs for unlabeled new data, and irrelevant sample pairs as negative sample pairs, and optimize the similarity between similar sample pairs through comparative learning; S50, performing label inference based on the feature vector obtained by contrastive learning to generate a labeling result.

[0006] As a further solution of the present invention, features are extracted and screened for different data types of initial data respectively. When extracting and screening features of text data, the text is segmented into words, a vocabulary is constructed, each text is represented as a word frequency vector, the TF-IDF weight of each word is calculated, and the semantic and content features reflecting the text are screened out based on the word frequency statistics; when extracting and screening features of tabular data, the numerical data is standardized, the statistical features are calculated to generate new interactive features, and the numerical features are screened out based on correlation analysis; when extracting and screening features of image data, the image size is adjusted and grayscale and normalized are performed, and the scale-invariant feature transformation is used to extract and maximize the image information based on feature importance.

[0007] As a further solution of the present invention, when constructing a multidimensional feature space, the text, table and image feature vectors are normalized, and based on the importance of each feature in the task, weight coefficients are allocated through weight optimization, the weighted feature vectors are spliced ​​into a unified feature vector by column, and the spliced ​​comprehensive feature vector is mapped to the multidimensional feature space, different types of data are aligned in the same feature space, and a multi-level model framework is constructed in the multidimensional feature space.

[0008] As a further solution of the present invention, when the annotation model is initially learned through self-supervised learning, a mask language model is constructed for the text dataset, part of the data columns are used to predict other data columns for the table dataset, and image restoration is performed on the image dataset.

[0009] As a further solution of the present invention, when constructing the annotation model, an input layer is constructed according to the feature information of the feature space, a feature extraction layer of the model is added and an output layer is added, and a suitable output structure is selected according to the task; when training the annotation model by self-supervised learning, a pre-training task is designed for text data using a mask language model, column masking prediction is used for table data, and image block masking prediction is used for image data to generate training data for the self-supervised task, the model is trained using the self-supervised task, the model weights are updated, and the feature extraction layer is adjusted according to the loss function value of the model.

[0010] As a further solution of the present invention, when fine-tuning the model, it includes: obtaining a labeled data set, selecting a feature layer obtained by self-supervised pre-training, fine-tuning the model using the labeled data set, adjusting the loss function to a task-specific goal, and performing supervised learning on the model; evaluating the performance of the fine-tuned model on a validation set, optimizing the model by adjusting hyperparameters, building a pipeline from self-supervised learning to supervised fine-tuning, performing final verification on a test set, and deploying the final model.

[0011] As a further solution of the present invention, the enterprise multi-type data labeling method based on feature engineering also includes generating adversarial network data expansion, wherein a generative adversarial network is constructed, a generator is used to generate pseudo data with labels, and a discriminator distinguishes between real data and pseudo data; the generated pseudo-label data is combined with part of the labeled data to perform semi-supervised learning and optimize the labeling model.

[0012] As a further solution of the present invention, the enterprise multi-type data labeling method based on feature engineering also includes modeling sequence data based on an attention mechanism, using a multi-head attention mechanism to allocate different attention weights in the labeling of different data types, verifying the generated labels, and optimizing the labeling model through verification feedback.

[0013] As a further scheme of the present invention, when performing label inference based on the feature vector obtained by contrastive learning, the feature vector obtained by contrastive learning is used to cluster the unlabeled data, similar samples in the feature space are assigned to the same cluster according to a similarity metric, and a pseudo-label is generated for each cluster, wherein the labels of the labeled samples are propagated to the cluster clusters; based on the compactness of the clusters and the similarity of the samples within the clusters, the confidence of the pseudo-label is evaluated, and the samples are sorted by confidence for label inference, and the samples below the confidence threshold are retained as unlabeled for further verification.

[0014] As a further solution of the present invention, the enterprise multi-type data labeling method based on feature engineering also includes fine-tuning and verification of pseudo labels. When fine-tuning the labeled data, the model is fine-tuned based on the labeled data. After the pseudo labels are generated, the validation set is used to evaluate the model performance, and the pseudo labels are iteratively corrected according to the evaluation results. After the model passes the verification, the final labeling results are generated for all unlabeled data, and the automated labeling is output.

[0015] The present invention also provides an enterprise multi-type data annotation system based on feature engineering, comprising: The data collection module is used to collect the initial data of the enterprise's internal information system and classify it according to the data type to obtain text data sets, table data sets and image data sets; The feature fusion module is used to extract and filter features of different data types of the initial data, fuse the feature vectors of different types of data, and construct a multi-dimensional feature space; The model building module is used to build a labeling model based on the feature information of the feature space, train the labeling model through self-supervised learning to initially learn the intrinsic structure and feature representation of the data, and fine-tune the model in combination with the labeled data; The similarity analysis module is used to construct similar sample pairs as positive sample pairs and irrelevant sample pairs as negative sample pairs for unlabeled new data, and to optimize the similarity between similar sample pairs through comparative learning; The data annotation module is used to perform label inference based on the feature vector obtained by contrastive learning and generate annotation results.

[0016] The beneficial effects of the present invention are: 1. The present invention can process multiple types of data such as text, tables, and images at the same time, breaking the limitation of traditional methods that only process a single data type; by designing specific feature extraction and processing strategies for each data type, it ensures that different types of data can be integrated and analyzed in a unified feature space, achieving efficient labeling of cross-type data, improving the applicability of the system, and being able to be applied to the data labeling needs of various information systems within the enterprise, enhancing the flexibility and wide adaptability of the system; 2. The present invention adopts the method of feature engineering to achieve the high efficiency of feature engineering and multi-dimensional feature fusion. It extracts and screens features of different types of data and constructs a multi-dimensional feature space by fusing feature vectors. The multi-dimensional feature space can capture the key features of multi-type data, fully reflect the inherent structure and relevance of the data, and provide more accurate and efficient feature input for subsequent annotation models. Through the fusion of multi-dimensional features, the system can better characterize and distinguish data, avoid inaccurate annotation problems caused by insufficient feature information, and significantly improve the annotation quality and accuracy. 3. The present invention uses self-supervised learning to enable the labeling model to automatically learn the intrinsic structure and feature representation of data in the absence of labeled data; the introduction of self-supervised learning reduces the dependence on large-scale labeled data, so that the model can have a strong learning ability in the early stage of training; at the same time, combined with fine-tuning of a small amount of labeled data, the performance of the model in actual labeling tasks is further improved; through self-supervised learning, the system performs well when facing large-scale unlabeled data, can effectively deal with the bottleneck problem of data labeling, reduce the cost of manual labeling, and improve labeling efficiency and the generalization ability of the model; 4. The present invention constructs similar sample pairs and irrelevant sample pairs through contrastive learning, optimizes the ability of the annotation model to distinguish positive and negative sample pairs, and has the ability to distinguish between the contrastive learning optimization model; through contrastive learning, the similarity of similar samples is strengthened, while the similarity of irrelevant samples is significantly reduced, and the ability of the model to distinguish different samples in the feature space is improved, thereby improving the overall annotation accuracy of the model; through contrastive learning, the system of the present invention can achieve efficient learning without a large amount of annotated data, especially in complex scenarios with many types of data, ensuring the reliability and accuracy of the annotation results; 5. The present invention has an efficient label inference mechanism. Based on the feature vectors obtained by contrastive learning, the system generates pseudo labels through an efficient label inference mechanism, which further improves the labeling efficiency of unlabeled data. At the same time, through iterative optimization of label inference and model training, the gradual optimization and accuracy improvement of the labeling results are ensured. The present invention enables the system to quickly generate high-confidence labeling results through the label inference mechanism, avoiding the problems of long time consumption and low efficiency of traditional labeling methods, and is particularly suitable for labeling tasks of large-scale internal enterprise data. 6. The automated labeling process in the present invention combines self-supervised learning and iterative optimization strategies, which can continuously optimize the learning effect of the model during the labeling process; with the introduction and iterative update of more unlabeled data, the system can continuously improve the accuracy and consistency of the labeling results; through the automated and iteratively optimized labeling process, the system can continuously improve the labeling quality while reducing manual intervention, and realize an efficient and intelligent enterprise data labeling process; 7. The present invention reduces the labor cost of enterprise data annotation through the automatic annotation mechanism of self-supervised learning and contrastive learning, reduces the dependence on a large amount of manual annotation, and enables enterprises to quickly process large-scale internal data without a large amount of manual participation, significantly reducing the labor cost and time cost of data annotation; the system of the present invention is particularly suitable for the large-scale data annotation needs of enterprises, reduces the investment of human resources, reduces the annotation cost, and improves the efficiency of enterprise data processing; 8. The present invention improves the accuracy of data mining and analysis through precise data labeling, and also provides high-quality labeled data for subsequent data mining and analysis, ensuring that data mining models and business decision models can be trained and optimized based on accurate labels, thereby improving the reliability of enterprise data analysis and decision-making, providing enterprises with a more accurate data foundation, assisting enterprises in data-driven decision-making and optimization, and improving the utilization efficiency of enterprise data assets.

[0017] To sum up, the present invention has many advantages such as strong versatility, high labeling accuracy, high degree of automation, strong iterative optimization capability and low labor cost. It significantly improves the efficiency and quality of multi-type data labeling in enterprises, and can provide a solid foundation for enterprise data mining, data analysis and business optimization, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of a method for labeling multi-type enterprise data based on feature engineering according to an embodiment of the present invention; Figure 2Schematic diagram of the test results of feature extraction effectiveness of text datasets, where (a) is a schematic diagram of the change of feature extraction semantic similarity over time, (b) is a schematic diagram of the change of feature extraction semantic information retention over time, (c) is a comparison of classification accuracy before and after feature extraction, and (d) is a comparison of processing time before and after feature extraction; Figure 3 Schematic diagram of the test results of feature extraction and screening effect of tabular data sets, where (a) is a schematic diagram of the change of feature quantity over time, (b) is the result of correlation analysis, (c) is the comparison of prediction accuracy, and (d) is the comparison of training time; Figure 4 Schematic diagram of the test results of feature extraction and screening accuracy of image datasets, where (a) is the number of key points extracted by SIFT and SURF, (b) is the comparison of classifier accuracy, (c) is the comparison of feature extraction speed, and (d) is the robustness test result; Figure 5 Schematic diagram of the test results of multi-dimensional feature space construction and alignment effect, where (a) is the comparison of classification accuracy, (b) is the effect of feature space dimensionality reduction, (c) is the comparison of F1 scores, and (d) is the classification accuracy under noise interference; Figure 6 Schematic diagram of the performance test results of the self-supervised learning annotation model, where (a) is the classification accuracy comparison, (b) is the F1 score comparison, (c) is the feature representation cosine similarity comparison, and (d) is the model convergence speed comparison. DETAILED DESCRIPTION

[0019] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0020] In view of the problem that the existing active data labeling methods only process a single type of data and are difficult to be applied to multiple types of data at the same time, the present invention proposes a method and system for enterprise multi-type data labeling based on feature engineering, which can realize efficient and automatic labeling of multiple types of data, and is applicable to various heterogeneous data such as text, tables, images, etc., and has high labeling accuracy and scalability.

[0021] like Figure 1 As shown, an embodiment of the present invention provides a method for labeling enterprise multi-type data based on feature engineering, and the method includes the following steps:

[0022] S10. Collect initial data from the enterprise's internal information system, and classify it according to data type to obtain text data sets, table data sets, and image data sets.

[0023] In this step, features are extracted and screened for different data types of the initial data. When extracting and screening features of text data, the text is segmented into words, a vocabulary is constructed, each text is represented as a word frequency vector, the TF-IDF weight of each word is calculated, and the semantic and content features reflecting the text are screened out based on the word frequency statistics; when extracting and screening features of tabular data, the numerical data is standardized, the statistical features are calculated to generate new interactive features, and the numerical features are screened out based on correlation analysis; when extracting and screening features of image data, the image size is adjusted and grayscale and normalized are performed, and the scale-invariant feature transformation is used to extract the maximum image information based on feature importance.

[0024] Initial data is collected from the internal information system of the enterprise and classified according to its data type, and divided into text data sets, table data sets and image data sets. For example, the content of internal emails of the enterprise is collected as text data; monthly reports are exported from the financial system as table data; product images are collected and annotated as image data.

[0025] The present invention can simultaneously process multiple types of data, such as text, tables, and images, breaking the limitation of traditional methods that only process a single data type. By designing specific feature extraction and processing strategies for each data type, it ensures that different types of data can be integrated and analyzed in a unified feature space, achieving efficient labeling of cross-type data, improving the applicability of the system, and being able to be applied to the data labeling needs of various information systems within the enterprise, enhancing the flexibility and wide adaptability of the system.

[0026] S20, extracting and screening features of different data types of the initial data respectively, fusing feature vectors of different types of data, and constructing a multi-dimensional feature space.

[0027] In this step, when constructing a multi-dimensional feature space, the text, table and image feature vectors are normalized, and based on the importance of each feature in the task, the weight coefficients are allocated through weight optimization. The weighted feature vectors are spliced ​​into a unified feature vector by column, and the spliced ​​comprehensive feature vector is mapped to the multi-dimensional feature space. Different types of data are aligned in the same feature space, and a multi-level model framework is constructed in the multi-dimensional feature space.

[0028] In this embodiment, when extracting and filtering text data features, the text data is segmented, a vocabulary is generated, the text is represented as a word frequency vector, and the word frequency-inverse document frequency weight of each word is calculated, and representative words are screened out according to the weight. For example, when processing product descriptions, keywords such as "high efficiency" and "energy saving" are extracted, and important words reflecting product characteristics are screened out according to word frequency. When extracting and filtering tabular data features, the tabular data is numerically standardized, and statistical features are calculated to generate interactive features. The most valuable numerical features for the task are selected based on correlation analysis. For example, for financial statement data, statistical features such as the mean and standard deviation of different financial indicators are calculated, and the ratio of sales revenue to cost is generated as a new feature. When extracting and filtering image data features, the image size is adjusted to be consistent, and grayscale and normalized are performed. Scale-invariant feature transformation is used to extract representative image features, and they are screened based on feature importance. For example, when processing corporate logo images, edge features and key points are extracted, and these key point data are stored as feature descriptors of the image.

[0029] The present invention adopts the method of feature engineering to achieve the high efficiency of feature engineering and multi-dimensional feature fusion, by extracting and screening features of different types of data, and constructing a multi-dimensional feature space through the fusion of feature vectors. The multi-dimensional feature space can capture the key features of multi-type data, comprehensively reflect the inherent structure and correlation of the data, and provide more accurate and efficient feature input for subsequent annotation models. Through the fusion of multi-dimensional features, the present invention enables the system to better characterize and distinguish data, avoid inaccurate annotation problems caused by insufficient feature information, and significantly improve the annotation quality and accuracy.

[0030] S30. Construct a labeling model based on the feature information of the feature space, train the labeling model through self-supervised learning to preliminarily learn the intrinsic structure and feature representation of the data, and fine-tune the model in combination with the labeled data.

[0031] In this step, when the annotation model is initially trained through self-supervised learning to learn data, a mask language model is built for the text dataset, some data columns are used to predict other data columns for the table dataset, and image restoration is performed for the image dataset.

[0032] Among them, when building the annotation model, the input layer is constructed according to the feature information of the feature space, the feature extraction layer of the model is increased and the output layer is added, and the appropriate output structure is selected according to the task; when training the annotation model by self-supervised learning, the text data uses the mask language model to design pre-training tasks, the table data uses column masking prediction, and the image data uses image block masking prediction to generate training data for the self-supervised task, use the self-supervised task to train the model, update the model weights, and adjust the feature extraction layer according to the loss function value of the model.

[0033] In this embodiment, fine-tuning the model includes: obtaining a labeled data set, selecting a feature layer obtained by self-supervised pre-training, fine-tuning the model using the labeled data set, adjusting the loss function to a task-specific goal, and performing supervised learning on the model; evaluating the performance of the fine-tuned model on the validation set, optimizing the model by adjusting hyperparameters, building a pipeline from self-supervised learning to supervised fine-tuning, performing final verification on the test set, and deploying the final model.

[0034] In some embodiments, the enterprise multi-type data labeling method based on feature engineering also includes modeling sequence data based on an attention mechanism, using a multi-head attention mechanism to assign different attention weights in the labeling of different data types, verifying the generated labels, and optimizing the labeling model through verification feedback.

[0035] When performing feature vector fusion, the three types of feature vectors, text, table, and image, are normalized separately to unify the feature scale. The weight coefficient is adjusted according to the task requirements, and the corresponding weight is assigned to each feature vector. The weighted feature vectors are spliced ​​into a unified feature vector by column. For product descriptions (text), product specifications (tables), and product pictures (images), the feature vectors are calculated and normalized separately, and the text features are weighted to highlight the semantic content. Then, each vector is spliced ​​by column to form a complete product feature vector.

[0036] When constructing a multidimensional feature space, the spliced ​​comprehensive feature vector is mapped to the multidimensional feature space to ensure that the text, table, and image features are aligned in the space. In the multidimensional feature space, a multi-level model framework is constructed to facilitate data annotation and analysis in subsequent steps. By comprehensively considering the text description, performance indicators, and image features of the company's products, a multidimensional feature space is formed for subsequent data annotation and analysis.

[0037] The present invention uses self-supervised learning to enable the labeling model to automatically learn the intrinsic structure and feature representation of data in the absence of labeled data. The introduction of self-supervised learning reduces the dependence on large-scale labeled data, allowing the model to have a strong learning ability in the early stages of training. At the same time, combined with fine-tuning of a small amount of labeled data, the performance of the model in actual labeling tasks is further improved. Through self-supervised learning, the system performs well when facing large-scale unlabeled data, can effectively deal with the bottleneck problem of data labeling, reduce the cost of manual labeling, and improve labeling efficiency and the generalization ability of the model.

[0038] The automated labeling process in the present invention combines self-supervised learning and iterative optimization strategies, which can continuously optimize the learning effect of the model during the labeling process. With the introduction and iterative update of more unlabeled data, the system can continuously improve the accuracy and consistency of the labeling results. Through the automated and iteratively optimized labeling process, the system can continuously improve the labeling quality while reducing manual intervention, and realize an efficient and intelligent enterprise data labeling process.

[0039] S40. For the unlabeled new data, similar sample pairs are constructed as positive sample pairs, and irrelevant sample pairs are constructed as negative sample pairs. Comparative learning is used to optimize the similarity between similar sample pairs.

[0040] In this step, when label inference is performed based on the feature vector obtained by contrastive learning, the feature vector obtained by contrastive learning is used to cluster the unlabeled data, similar samples in the feature space are assigned to the same cluster according to the similarity metric, and pseudo labels are generated for each cluster, where the labels of the labeled samples are propagated to the clusters; the confidence of the pseudo labels is evaluated according to the compactness of the clusters and the similarity of the samples within the clusters, and the samples are sorted by confidence for label inference, and the samples below the confidence threshold are retained as unlabeled for further verification.

[0041] In the constructed multidimensional feature space, self-supervised learning is used for preliminary labeling. The model learns the intrinsic structure of the data in an unsupervised way, generates pseudo labels for unlabeled data, and then fine-tunes with a small amount of labeled data. For example, self-supervised learning is performed on product data, and the "high-end" or "basic" categories are preliminarily labeled. Then, fine-tuning is performed based on actual data to improve the labeling accuracy. The labeling model is trained through comparative learning, and similar sample pairs and dissimilar sample pairs are constructed to optimize the model's ability to distinguish. The similarity of similar samples is strengthened, and the similarity of dissimilar samples is reduced, thereby optimizing the labeling accuracy. Comparative learning is performed on similar product descriptions to strengthen the model's ability to recognize similar product categories, so as to achieve efficient labeling of product categories.

[0042] The present invention constructs similar sample pairs and irrelevant sample pairs through contrastive learning, optimizes the ability of the annotation model to distinguish positive and negative sample pairs, and has the ability to distinguish between the contrastive learning optimization model. Through contrastive learning, the similarity of similar samples is strengthened, while the similarity of irrelevant samples is significantly reduced, and the ability of the model to distinguish different samples in the feature space is improved, thereby improving the overall annotation accuracy of the model. Through contrastive learning, the system of the present invention can achieve efficient learning without a large amount of annotated data, especially in complex scenarios with many types of data, ensuring the reliability and accuracy of the annotation results.

[0043] S50, performing label inference based on the feature vector obtained by contrastive learning to generate a labeling result.

[0044] In this step, the enterprise multi-type data labeling method based on feature engineering also includes fine-tuning and verification of pseudo labels. When fine-tuning the labeled data, the model is fine-tuned based on the labeled data. After the pseudo labels are generated, the validation set is used to evaluate the model performance, and the pseudo labels are iteratively corrected based on the evaluation results. After the model passes the verification, the final labeling results are generated for all unlabeled data, and the automated labeling is output.

[0045] In this embodiment, when the label inference mechanism is performed, clustering is applied to infer labels based on the feature vectors obtained by contrastive learning, and pseudo-label generation and annotation results are iteratively optimized to improve the annotation quality. In the process of classifying products, pseudo-labels are automatically generated through clustering algorithms, and data with inconsistent labels are re-annotated to improve label accuracy.

[0046] During the iterative optimization of the multi-level annotation model framework, a multi-level model framework is constructed, covering model modules such as text classification, table clustering, and image recognition, providing highly adaptable annotation solutions for different types of data. For enterprise data annotation, text data is annotated through the text classification model, table data is processed through the clustering algorithm, and image data is annotated through the recognition model to form a multi-level annotation system. Through the automated iterative optimization process, more unlabeled data is introduced in each round of annotation to continuously optimize the annotation accuracy of the model. The system optimizes the annotation results in real time based on feedback. After annotating the new product categories, the model weights are continuously updated based on user feedback to improve the accuracy and consistency of the system when annotating new product features.

[0047] The enterprise multi-type data labeling method of the present invention effectively solves the efficiency and accuracy problems of multi-type data labeling through the steps of collection, feature extraction, feature fusion, self-supervised learning, comparative learning, etc. The method described in this embodiment provides an automated and efficient labeling solution for different types of data within the enterprise, significantly reduces labor costs, improves labeling quality, and lays a solid foundation for subsequent data mining and business analysis.

[0048] The present invention has an efficient label inference mechanism. Based on the feature vectors obtained by contrastive learning, the system generates pseudo labels through an efficient label inference mechanism, further improving the labeling efficiency of unlabeled data. At the same time, by iteratively optimizing label inference and model training, the gradual optimization and accuracy improvement of the labeling results are ensured. The present invention enables the system to quickly generate high-confidence labeling results through the label inference mechanism, avoiding the problems of long time consumption and low efficiency of traditional labeling methods, and is particularly suitable for labeling tasks for large-scale internal enterprise data.

[0049] The present invention reduces the labor cost of enterprise data annotation through the automated annotation mechanism of self-supervised learning and contrastive learning, reduces the reliance on a large amount of manual annotation, and enables enterprises to quickly process large-scale internal data without a large amount of manual participation, significantly reducing the labor cost and time cost of data annotation. The system of the present invention is particularly suitable for the large-scale data annotation needs of enterprises, reduces the investment of human resources, reduces the annotation cost, and improves the efficiency of enterprise data processing.

[0050] The present invention improves the accuracy of data mining and analysis through precise data labeling. The present invention also provides high-quality labeled data for subsequent data mining and analysis, ensuring that data mining models and business decision models can be trained and optimized based on accurate labels, thereby improving the reliability of enterprise data analysis and decision-making, providing enterprises with a more accurate data foundation, assisting enterprises in data-driven decision-making and optimization, and improving the utilization efficiency of enterprise data assets.

[0051] It should be noted that the above figures are only schematic illustrations of the processes included in the method according to an exemplary embodiment of the present invention, and are not intended to be limiting. It is easy to understand that the processes shown in the above figures do not indicate or limit the time sequence of these processes. In addition, it is also easy to understand that these processes can be performed synchronously or asynchronously, for example, in multiple modules.

[0052] It should be understood that, although described in a certain order, these steps are not necessarily performed in sequence in the above order. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be performed in other orders. Moreover, a part of the steps of the present embodiment may include a plurality of steps or a plurality of stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a part of the steps or stages in other steps or other steps.

[0053] In a second aspect of the embodiment of the present invention, the present invention further provides an enterprise multi-type data labeling system based on feature engineering, comprising:

[0054] The data collection module is used to collect the initial data of the enterprise's internal information system and classify it according to the data type to obtain text data sets, table data sets and image data sets;

[0055] The feature fusion module is used to extract and filter features of different data types of the initial data, fuse the feature vectors of different types of data, and construct a multi-dimensional feature space;

[0056] The model building module is used to build a labeling model based on the feature information of the feature space, train the labeling model through self-supervised learning to initially learn the intrinsic structure and feature representation of the data, and fine-tune the model in combination with the labeled data;

[0057] The similarity analysis module is used to construct similar sample pairs as positive sample pairs and irrelevant sample pairs as negative sample pairs for unlabeled new data, and to optimize the similarity between similar sample pairs through comparative learning;

[0058] The data annotation module is used to perform label inference based on the feature vector obtained by contrastive learning and generate annotation results.

[0059] Through the above detailed steps and examples, the enterprise multi-type data labeling system based on feature engineering of the present invention is used to execute the steps of the enterprise multi-type data labeling method based on feature engineering in the above embodiment, which will not be repeated here. The enterprise multi-type data labeling system based on feature engineering of the present invention has many advantages such as strong versatility, high labeling accuracy, high degree of automation, strong iterative optimization capability, and low labor cost. It significantly improves the efficiency and quality of enterprise multi-type data labeling, can provide a solid foundation for enterprise data mining, data analysis and business optimization, and has broad application prospects.

[0060] In order to verify the performance of the method and system of the present invention, the following experiments were conducted.

[0061] (1) Test the effectiveness of feature extraction of text datasets;

[0062] Experimental background: In modern enterprise information systems, the automated processing and analysis of text data is one of the key tasks in the field of data science. As the first step in text mining, the effectiveness of feature extraction directly affects the accuracy and efficiency of subsequent analysis. This experiment aims to evaluate the effectiveness of feature extraction in text datasets, with a special focus on the role of term frequency-inverse document frequency (TF-IDF) weights in extracting text semantics and content features. The TF-IDF algorithm can quantify the importance of a word to a document set or one of the documents in a corpus. Therefore, this experiment will use the TF-IDF algorithm to identify and extract key features in text data in order to accurately represent the text content.

[0063] Experimental method: The experiment selected 1,000 internal corporate text data, which covered a variety of text types such as corporate reports, contracts, and emails. First, these text data were preprocessed, including word segmentation, stop word removal, and stem extraction to reduce noise and redundancy in subsequent processing. Subsequently, the TF-IDF algorithm was applied to extract features from the preprocessed text data and calculate the TF-IDF weight of each word, which reflects the importance of a word in the document collection. Based on the calculated TF-IDF weight, the top 10% of the words with the highest weight were selected as key features. These features are considered to be able to effectively capture the core content and semantic information of the text data.

[0064] Result analysis: Through TF-IDF feature extraction of 1,000 text data, the experimental results are as follows Figure 2 As shown, Figure 2 (a) is a schematic diagram of the change of feature extraction semantic similarity over time. Figure 2 (b) is a schematic diagram of the change of feature extraction semantic information retention over time. Figure 2 (c) is the comparison of classification accuracy before and after feature extraction. Figure 2 (d) is the comparison of processing time before and after feature extraction. The dimension of the feature vector after screening is reduced to 30% of the original data. This result is in line with the expected goal of feature extraction, that is, to reduce the data dimension while retaining key information, so as to facilitate subsequent data processing and analysis. Further semantic information retention test shows that the feature vector after screening can retain more than 95% of the semantic information, which means that the feature extraction process not only effectively reduces the data dimension, but also retains the semantic content of the original text to the greatest extent. In terms of specific values, taking 100 text data as a test set, by comparing the feature vectors extracted by manual annotation and machine, it is found that the average accuracy of the feature vector in semantic similarity reaches 87.5%, which shows that the effectiveness of feature extraction has been verified. In addition, by comparing with the original feature vector, the accuracy of the feature vector after screening in the classification task only decreases by 5%, while the processing time is reduced by 40%. This result further confirms the advantages of TF-IDF feature extraction in efficiency and accuracy.

[0065] (2) Testing the feature extraction and screening effect of table data sets;

[0066] Experimental background: In enterprise decision support systems, feature extraction and screening of tabular data sets, especially financial data sets, are key steps in achieving data-driven decision making. Financial data sets usually contain a large number of numerical features, which are directly related to the company's operating conditions and market performance. Effective feature extraction and screening can not only reduce the dimension of the data and reduce the computational burden of subsequent analysis, but also highlight key features and improve the prediction accuracy of the model. This experiment aims to verify the effect of feature extraction and screening in tabular data sets, with a special focus on the importance of numerical data standardization and statistical feature calculation in revealing the intrinsic structure of financial data.

[0067] Experimental method: The experiment selected 1,000 pieces of corporate financial data, which covered multiple key financial indicators such as revenue, cost, and profit. First, these financial data were preprocessed, including missing value processing and outlier detection, to ensure the quality of the data. Subsequently, the data were standardized to eliminate the dimensional effects between different indicators so that the features were comparable. On this basis, statistical features such as mean and variance were calculated to describe the central tendency and dispersion of the data. Furthermore, based on correlation analysis, features that are highly correlated with the financial status of the enterprise were screened out, which can accurately reflect the operating efficiency and profitability of the enterprise. By setting the correlation threshold, features with a correlation coefficient greater than 0.7 were selected as key features, and a feature subset of financial data was constructed.

[0068] Result analysis: The experimental results are as follows Figure 3 As shown, Figure 3 (a) is a schematic diagram showing the change of the number of features over time. Figure 3 (b) is the result of correlation analysis. Figure 3 (c) is the comparison of prediction accuracy. Figure 3(d) is the training time comparison. After standardization and statistical feature calculation, the number of key features screened out was reduced to 50% of the original features, effectively reducing the dimension of the data. Specifically, the original data set contained 100 features, and the number of features after screening was reduced to 50, which not only reduced the complexity of model training, but also improved the generalization ability of the model. The results of correlation analysis further confirmed the effectiveness of the screening strategy. The correlation between the screened features and the financial indicators of the enterprise was significantly improved, and the average correlation coefficient increased from 0.6 to above 0.85. For example, the correlation coefficient between revenue and profit increased from 0.68 to 0.92, and the correlation coefficient between cost and profit increased from 0.55 to 0.81. These results show that the screened features can more accurately reflect the inherent structure of financial data and the operating conditions of the enterprise. In the subsequent financial forecasting model, the prediction accuracy of the model was improved by 10% compared with the original feature set, while the training time was shortened by 30%. In addition, the stability and robustness of the screened feature set were verified by the cross-validation method, and the performance of the model on different subsets was highly consistent.

[0069] (3) Image dataset feature extraction and screening accuracy test;

[0070] Experimental background: In the field of enterprise product management and quality control, the automated analysis of image datasets is crucial for tasks such as product identification, classification, and defect detection. As one of the core technologies of image processing, the accuracy of image feature extraction and screening directly affects the reliability of image analysis results. Scale-Invariant Feature Transform (SIFT) and Speeded Up Robust Features (SURF) are two widely used algorithms that can extract key points and descriptors from images. These features are robust to rotation, scaling, brightness changes, etc. This experiment aims to evaluate the accuracy of feature extraction and screening in image datasets, and especially explores the role of SIFT and SURF algorithms in extracting key information from enterprise product images.

[0071] Experimental method: The experiment selected 500 enterprise product images, which included products of different categories, such as electronic products, mechanical parts, etc., to ensure the breadth and representativeness of the experiment. First, the images were preprocessed, including grayscale and normalization, to reduce the computational complexity and improve the accuracy of feature extraction. Subsequently, the SIFT and SURF algorithms were applied to extract the key points and descriptors of the image, respectively. The SIFT algorithm determines the key points by finding extreme points in the scale space and generates a unique direction and scale-invariant descriptor for each key point. The SURF algorithm is an accelerated version of the SIFT algorithm, which quickly calculates key points and descriptors by integrating the image. The extracted features are used to train a support vector machine (SVM) classifier to identify product images of different categories.

[0072] Result analysis: The experimental results are as follows Figure 4 As shown, Figure 4 (a) is the number of key points extracted by SIFT and SURF. Figure 4 (b) is the comparison of classifier accuracy. Figure 4 (c) is the speed comparison of feature extraction. Figure 4 (d) is the robustness test result. After the SIFT and SURF algorithms extract features, the number of key points is 70% and 75% of the original image, respectively. This result shows that the feature extraction process can effectively capture the key visual information of the image and remove a lot of redundant information. In terms of specific values, the SIFT algorithm extracts an average of 120 key points per image, and the SURF algorithm extracts an average of 150 key points per image. In the image recognition task, the accuracy of the SVM classifier using SIFT features reached 85%, while the accuracy of the SVM classifier using SURF features reached 90%, showing the superiority of the SURF algorithm in processing corporate product images. In addition, through comparative experiments, it is found that the SURF algorithm is 30% faster than the SIFT algorithm in extracting features, which makes the SURF algorithm more suitable for processing large-scale image data sets. In the robustness test, when the image is interfered by 30% noise, the recognition accuracy of the two algorithms only drops by 5%, indicating that the extracted features have good robustness.

[0073] (4) Multi-dimensional feature space construction and alignment effect testing;

[0074] Experimental background: In the era of big data, the types of data generated by enterprise internal information systems are becoming increasingly diverse, including text, tables, and images. Each of these data carries different information and knowledge. How to effectively integrate these heterogeneous data and construct a unified feature representation has become a challenge in the field of data science. The construction of a multi-dimensional feature space aims to map data from different sources and types into the same space, achieve feature alignment and fusion, and facilitate cross-modal data analysis and machine learning tasks. This experiment aims to verify the effectiveness of the construction of a multi-dimensional feature space, especially the alignment and fusion effects of different types of data in the same feature space, and the impact of this process on the performance of data classification tasks.

[0075] Experimental method: The experiment selected a multimodal dataset containing text, tables and images, of which the text dataset contains 1,000 corporate reports, the table dataset contains 1,000 financial records, and the image dataset contains 1,000 product images. First, TF-IDF feature extraction is performed on the text data, statistical feature extraction is performed on the table data, and SIFT feature extraction is performed on the image data. Then, the three types of feature vectors are normalized to eliminate the dimension effect and spliced ​​into a unified feature vector by column. This process involves feature selection and dimensionality reduction techniques to ensure that the final feature vector can represent the key information of the original data and has an appropriate dimension. Finally, the spliced ​​feature vector is mapped to a multidimensional feature space, and the support vector machine (SVM) is used as a classifier to classify the data in this space.

[0076] Result analysis: The experimental results are as follows Figure 5 As shown, Figure 5 (a) is the comparison of classification accuracy. Figure 5 (b) is the effect of reducing the dimension of the feature space. Figure 5 (c) is the F1 score comparison. Figure 5(d) is the classification accuracy under noise interference. The construction of multidimensional feature space effectively reduces the dimension of data while retaining key information. Specifically, the dimension of the original feature space is 10,000. After feature selection and dimensionality reduction, the dimension of the multidimensional feature space is reduced to 4,000, which is reduced to 40% of the original data. In the classification task, the classification accuracy of data using multidimensional feature space is significantly improved compared with that using single modality data. In terms of specific values, the classification accuracy of single modality data is 78% for text data, 75% for table data, and 80% for image data. When the data of multidimensional feature space is classified, the accuracy reaches 85%, an increase of about 7 percentage points. In addition, the F1 score of the data in the multidimensional feature space in the classification task is also improved, from an average of 0.75 for single modality data to 0.82. In the robustness test, when the data is interfered by noise, the classification accuracy of the multidimensional feature space decreases less than that of the single modality data, showing better stability. For example, when 5% Gaussian noise was added, the classification accuracy of the multidimensional feature space only dropped by 2%, while the classification accuracy of single-modal data dropped by 5% on average. These results show that the construction of multidimensional feature space can not only effectively integrate key information of different data types, but also improve the performance of data classification tasks and enhance the generalization ability and robustness of the model.

[0077] (5) Performance test of self-supervised learning annotation model;

[0078] Experimental background: With the advent of the big data era, enterprises have accumulated a large amount of unlabeled data, which is difficult to be directly used for the training of supervised learning models due to the lack of labels. Self-supervised learning, as an emerging learning paradigm, can learn the potential representation of data on unlabeled data, providing an effective solution for labeled models. Self-supervised learning constructs prediction tasks and uses the structural information of the data itself as a supervisory signal to train the model to learn useful feature representations. This experiment aims to evaluate the performance of self-supervised learning in labeled models, especially the learning effect on unlabeled data, and compare it with supervised learning to verify the effectiveness and applicability of self-supervised learning.

[0079] Experimental method: The experiment selected 2,000 pieces of enterprise data, of which 1,000 pieces were labeled data for training supervised learning models, and the other 1,000 pieces were unlabeled data for training self-supervised learning models and performing performance tests. The self-supervised learning model adopts a contrastive learning framework, constructs positive sample pairs and negative sample pairs, and uses the similarity between features as a supervisory signal. Specifically, for text data, a mask language model is constructed as a self-supervised task; for tabular data, a self-supervised task of predicting other data columns from some data columns is used; for image data, a self-supervised task of image block mask prediction is used. The self-supervised learning model is trained on unlabeled data to learn the feature representation of the data, and then fine-tuned on the labeled data to adapt to specific labeling tasks. The supervised learning model is trained directly on the labeled data. The performance of the two models is evaluated on the same test set, including indicators such as classification accuracy and F1 score.

[0080] Result analysis: The experimental results are as follows Figure 6 As shown, Figure 6 (a) is the comparison of classification accuracy. Figure 6 (b) is the F1 score comparison. Figure 6 (c) is the feature representation cosine similarity comparison, Figure 6 (d) is a comparison of model convergence speed. The classification accuracy of the self-supervised learning model on unlabeled data reached 72%, which was 13% lower than the 85% of the supervised learning model, lower than the expected 15%, showing the effectiveness of self-supervised learning. In terms of specific values, the classification accuracy of the self-supervised learning model on text data was 70%, on table data was 75%, and on image data was 73%, all close to or slightly lower than the performance of the supervised learning model. In terms of F1 score, the average F1 score of the self-supervised learning model was 0.71, while the average F1 score of the supervised learning model was 0.82, with a gap of 11 percentage points. In addition, the self-supervised learning model also showed high consistency in the quality of feature representation. The average difference between the cosine similarity of the feature representation and the similarity of the supervised learning model was only 0.05. In terms of model convergence speed, due to pre-training on unlabeled data, the self-supervised learning model converges 20% faster on labeled data than the supervised learning model. These results show that self-supervised learning can effectively utilize unlabeled data, learn useful feature representations, and make up for the lack of labeled information to a certain extent. Although the performance of the self-supervised learning model is slightly lower than that of the supervised learning model, it shows significant advantages in the utilization of unlabeled data and feature learning, providing an effective solution for the training of labeled models.

[0081] The above description is only a preferred embodiment of the present invention and is not intended to be a further limitation of the present invention. All equivalent changes made using the contents of the present specification and drawings are within the protection scope of the present invention.

Claims

1. The enterprise multi-type data labeling method based on feature engineering is characterized by: The following steps are involved: S10, collecting initial data from the enterprise's internal information system, and classifying it according to data type to obtain a text data set, a table data set, and an image data set; S20, extracting and screening features of different data types of the initial data respectively, fusing feature vectors of different types of data, and constructing a multi-dimensional feature space; S30, constructing a labeling model according to the feature information of the feature space, training the labeling model through self-supervised learning to preliminarily learn the intrinsic structure and feature representation of the data, and fine-tuning the model in combination with the labeled data; S40, construct similar sample pairs as positive sample pairs for unlabeled new data, and irrelevant sample pairs as negative sample pairs, and optimize the similarity between similar sample pairs through comparative learning; S50, performing label inference based on the feature vector obtained by contrastive learning to generate a labeling result.

2. The enterprise multi-type data labeling method based on feature engineering according to claim 1 is characterized in that: In S20, features are extracted and screened for different data types of the initial data, specifically: When extracting and filtering text data features, the text is divided into words, a vocabulary is constructed, each text is represented as a word frequency vector, the TF-IDF weight of each word is calculated, and the semantic and content features of the text are filtered out based on the word frequency statistics; When extracting and filtering features of table data, the numerical data is standardized, statistical features are calculated to generate new interactive features, and numerical features are filtered out based on correlation analysis; When extracting and screening image data features, the image size is adjusted and grayscale and normalized. Through scale-invariant feature transformation, the maximum image information is extracted based on feature importance.

3. The enterprise multi-type data labeling method based on feature engineering according to claim 2 is characterized in that: When constructing a multi-dimensional feature space, the text, table and image feature vectors are normalized. Based on the importance of each feature in the task, the weight coefficients are allocated through weight optimization. The weighted feature vectors are spliced ​​into a unified feature vector by column. The spliced ​​comprehensive feature vector is mapped to the multi-dimensional feature space. Different types of data are aligned in the same feature space, and a multi-level model framework is constructed in the multi-dimensional feature space.

4. The enterprise multi-type data labeling method based on feature engineering according to claim 3 is characterized in that: In S30, when the annotation model is initially learned through self-supervised learning, a mask language model is constructed for the text dataset, part of the data columns are used to predict other data columns for the table dataset, and image restoration is performed for the image dataset.

5. The enterprise multi-type data labeling method based on feature engineering according to claim 4 is characterized in that: When building the annotation model, the input layer is constructed according to the feature information of the feature space, the feature extraction layer of the model is increased and the output layer is added, and the appropriate output structure is selected according to the task; when training the annotation model with self-supervised learning, the mask language model is used to design pre-training tasks for text data, column masking prediction is used for table data, and image block masking prediction is used for image data to generate training data for self-supervised tasks, the model is trained using self-supervised tasks, the model weights are updated, and the feature extraction layer is adjusted according to the loss function value of the model.

6. The enterprise multi-type data labeling method based on feature engineering according to claim 5 is characterized in that: When fine-tuning the model, it includes: obtaining a labeled dataset, selecting the feature layer obtained by self-supervised pre-training, fine-tuning the model using the labeled dataset, adjusting the loss function to a task-specific goal, and performing supervised learning on the model; evaluating the performance of the fine-tuned model on the validation set, optimizing the model through hyperparameter adjustment, building a pipeline from self-supervised learning to supervised fine-tuning, performing final verification on the test set, and deploying the final model.

7. The enterprise multi-type data labeling method based on feature engineering according to claim 1 is characterized in that: It also includes generative adversarial network data expansion, in which a generative adversarial network is constructed, a generator is used to generate pseudo data with labels, and a discriminator distinguishes between real data and pseudo data; the generated pseudo-labeled data is combined with part of the labeled data for semi-supervised learning to optimize the labeling model.

8. The enterprise multi-type data labeling method based on feature engineering according to claim 7 is characterized in that: It also includes modeling sequence data based on the attention mechanism, using a multi-head attention mechanism to assign different attention weights in the annotation of different data types, verifying the generated labels, and optimizing the annotation model through verification feedback.

9. The enterprise multi-type data labeling method based on feature engineering according to claim 1 is characterized in that: In S50, when label inference is performed based on the feature vector obtained by contrastive learning, the feature vector obtained by contrastive learning is used to cluster the unlabeled data, similar samples in the feature space are assigned to the same cluster according to a similarity metric, and a pseudo-label is generated for each cluster, wherein the labels of the labeled samples are propagated to the cluster; based on the compactness of the clustering and the similarity of the samples within the cluster, the confidence of the pseudo-label is evaluated, and the samples are sorted by confidence for label inference, and samples below the confidence threshold are retained as unlabeled for further verification.

10. The enterprise multi-type data annotation system based on feature engineering is characterized by: The system is used to perform the enterprise multi-type data labeling method based on feature engineering as described in any one of claims 1 to 9, and comprises: The data collection module is used to collect the initial data of the enterprise's internal information system and classify it according to the data type to obtain text data sets, table data sets and image data sets; The feature fusion module is used to extract and filter features of different data types of the initial data, fuse the feature vectors of different types of data, and construct a multi-dimensional feature space; The model building module is used to build a labeling model based on the feature information of the feature space, train the labeling model through self-supervised learning to initially learn the intrinsic structure and feature representation of the data, and fine-tune the model in combination with the labeled data; The similarity analysis module is used to construct similar sample pairs as positive sample pairs and irrelevant sample pairs as negative sample pairs for unlabeled new data, and to optimize the similarity between similar sample pairs through comparative learning; The data annotation module is used to perform label inference based on the feature vector obtained by contrastive learning and generate annotation results.

Images