Data value evaluation method, electronic equipment and medium

Abstract

The invention relates to a data value evaluation method, electronic equipment and a medium. The method comprises the following steps: carrying out dimension reduction processing on high-dimensional data to obtain low-dimensional feature data; calculating the representativeness of the low-dimensional feature data and carrying out clustering; selecting samples with the highest representativeness from clustering results to form a representative sample set; performing data value evaluation on the representative sample by using a preliminary evaluation method to obtain the data value of the representative sample; according to the evaluation result, in combination with the representativeness of the unevaluated samples, performing data value complementation on the unevaluated samples to form a complemented data value set; high-dimensional data and complete annotations are used as input, and a numerical value enrichment model is constructed and optimized; and completing the final value evaluation of the high-dimensional data by using the trained numerical value enrichment model. According to the method, by means of representative sample screening and introduction of a numerical value enrichment model, the calculation complexity and time cost of data value evaluation are reduced.

Description

Technical Field

[0001] The present invention relates to the field of machine learning, and in particular to a data value assessment method, electronic equipment and medium. Background Art

[0002] With the advent of the big data era, data has become an important production factor that drives social and economic development. As an important way to share and circulate data resources, data trading has played a vital role in promoting the development of the digital economy. However, in the process of data trading, how to scientifically and reasonably evaluate the value of data has become a core issue that needs to be solved in data trading.

[0003] At present, the mainstream data value assessment method is mainly based on the "Shapley value" of the data. This method combines a sample to be evaluated in the data set with other samples into a sub-dataset, and uses a machine learning model to calculate the impact of the sub-dataset containing the sample and the sub-dataset excluding the sample on the machine learning effect, thereby determining the value of the sample in the entire data set. This evaluation method can theoretically effectively reflect the contribution of the sample to the overall data set and is widely recognized in academia. However, the Shapley value evaluation method requires training and calculation of multiple sub-datasets. The computational complexity increases exponentially with the increase in the size of the data set, and the computing power requirements are extremely high, resulting in a long evaluation time and difficulty in meeting the demand for efficiency in data transactions.

[0004] In addition, with the rapid growth of data volume and diversity, data value assessment faces more practical challenges. Specifically, there are a large number of similar or redundant samples in the data set, and these samples usually have similar values ​​in data value assessment. If all samples are valued one by one, it will not only waste computing resources, but also significantly increase the evaluation time. Therefore, how to selectively evaluate representative samples and then infer the value of other similar samples is an important research direction to improve evaluation efficiency. Summary of the invention

[0005] The present invention provides a data value assessment method, an electronic device and a medium, the purpose of which is to solve the problems of high computational complexity and low efficiency existing in the current data value assessment method in practical applications.

[0006] To achieve the above object, the present invention provides a data value assessment method in a first aspect, comprising the following steps: Use pre-trained encoders to encode high-dimensional data into low-dimensional feature data; Based on the low-dimensional feature data, calculating the representativeness of each sample; Clustering the low-dimensional feature data based on the representativeness, and selecting representative samples from each cluster to form a representative sample set; Use a preliminary evaluation method to evaluate the data value of samples in the representative sample set to obtain the data value of the representative samples; Based on the preliminary evaluation results of representative samples and the representativeness of other unevaluated samples, the data value of unevaluated samples is supplemented to form a supplemented data value set; Constructing a numerical enrichment model, taking the high-dimensional data as input and the completed data value set as a target, and training the numerical enrichment model by optimizing a preset loss function; The final data value assessment of high-dimensional data is performed through the trained numerical enrichment model.

[0007] Further, methods for calculating representativeness of low-dimensional feature data include: For each low-dimensional feature data, calculate the similarity measure between it and all other low-dimensional feature data; The sum of the similarity measures is taken as the representative value of the sample; wherein the similarity measure is calculated by the exponential function of the Euclidean distance between samples, and the calculation formula is as follows:

[0008] in, Indicates The representativeness of the sample, and Respectively represent two samples in the low-dimensional feature data, Represents the total number of samples in the low-dimensional feature data set, Indicates that when calculating the representative value, The index of the other samples to compare with, From 1 to Traverse all samples in the low-dimensional feature data set.

[0009] Furthermore, the method of clustering the low-dimensional feature data based on the representativeness and selecting representative samples from each cluster to form a representative sample set includes: The low-dimensional feature data is divided into multiple clusters by a preset clustering algorithm, wherein the clustering algorithm is the KMeans method and the number of clusters is a preset value. ; In each cluster, the low-dimensional feature data are sorted according to their representativeness, and the sample with the highest representativeness is selected as the representative sample of the cluster; Aggregate the representative samples selected from all clusters to form A representative sample set of samples.

[0010] Furthermore, the method of using the preliminary evaluation method to evaluate the data value of samples in the representative sample set includes: For each sample in the representative sample set, generating a sub-dataset including the sample and a sub-dataset excluding the sample; By training the machine learning model, the difference in the impact of the sub-dataset containing the sample and the sub-dataset excluding the sample on the model performance is calculated respectively; Calculate the Shapley value of the sample based on the impact difference as its data value for preliminary evaluation; The Shapley values ​​of all representative samples are summarized to form a preliminary evaluation result.

[0011] Furthermore, based on the preliminary evaluation results of the representative samples and in combination with the representativeness of other unevaluated samples, the data value of the unevaluated samples is supplemented to form a method of supplementing the data value set, including: aligning the preliminary evaluation results of the representative samples with the index of the low-dimensional feature data; Initialize the values ​​of data not selected as representative samples to zero; According to the aligned index relationship, the preliminary evaluation result is merged with the initialized zero value to obtain a completed data value set.

[0012] Furthermore, the training method of the numerical enrichment model includes: Taking the completed data value set as the goal, inputting high-dimensional data into the numerical enrichment model; The parameters of the numerical enrichment model are adjusted by optimizing a preset loss function so that the data value output by the numerical enrichment model is aligned with the non-zero value in the complementary data value set, and the zero value is dynamically corrected based on the similarity of low-dimensional features.

[0013] Furthermore, the loss function is defined as:

[0014] in, Represents the output of the numerical enrichment model The predicted data value of high-dimensional data, Indicates the corresponding completion data value.

[0015] Furthermore, the numerical enrichment model is a deep neural network, whose input is high-dimensional data and output is a single-channel numerical value, and the single-channel numerical value directly represents the value of the data; the deep neural network adopts a Densenet structure.

[0016] To achieve the above objectives, the first aspect of the present invention provides an electronic device, including a processor and a memory, wherein the processor is used to implement the steps of the data value assessment method when executing a computer program stored in the memory.

[0017] To achieve the above objectives, the first aspect of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the data value assessment method are executed.

[0018] Beneficial effects of the present invention: Compared with the prior art, the data value assessment method, electronic device and medium provided by the present invention significantly reduce the computational complexity and time consumption of data value assessment by introducing representative sample screening and numerical enrichment models. Specifically, the present invention first uses a pre-trained encoder to reduce the dimensionality of high-dimensional data and extract low-dimensional features; then calculates the representative value based on the similarity between samples, and selects the most representative samples from each cluster cluster through a clustering method to form a representative sample set; then, by performing a preliminary evaluation of the representative samples and combining the numerical enrichment model, the value of the remaining samples is quickly inferred. This method avoids the computational burden of evaluating all samples one by one by giving priority to key data points, and at the same time uses the similarity principle to ensure the accuracy of value inference of unevaluated samples, thereby effectively solving the problems of time-consuming and high computing power requirements of the evaluation methods in the prior art, and is suitable for efficient data value assessment in large-scale data trading scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments.

[0020] Figure 1 It is a flow chart of a data value assessment method disclosed in an embodiment of the present invention.

[0021] Figure 2 It is a process diagram of a data value assessment method disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0023] According to an embodiment of the present invention, it should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the following method, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0024] The traditional Shapley value method needs to traverse a large number of sample combinations, which is extremely computationally expensive, and simply using clustering to select representative samples can only simplify some of the calculations. This method introduces a numerical enrichment model, combines the preliminary evaluation results with deep learning, and designs a specific loss function, so that the model can learn the global value distribution from sparse annotations (only some samples have evaluation values); the following will explain this method in detail.

[0025] like Figure 1 , Figure 2 As shown, the present invention provides a data value assessment method, which prioritizes the assessment of representative data values, and then utilizes the relationship between data similarity and value similarity to achieve fast and efficient data value assessment, specifically comprising the following steps: Step S100, using a pre-trained encoder to encode high-dimensional data into low-dimensional feature data; Step S200: Calculate the representativeness of each sample based on the low-dimensional feature data; Step S300: clustering the low-dimensional feature data based on the representativeness, and selecting representative samples from each cluster to form a representative sample set; Step S400: Use a preliminary evaluation method to evaluate the data value of samples in the representative sample set to obtain the data value of the representative samples; Step S500: Based on the preliminary evaluation results of the representative samples and in combination with the representativeness of other unevaluated samples, the data value of the unevaluated samples is supplemented to form a supplemented data value set; Step S600: constructing a numerical enrichment model, taking the high-dimensional data as input and the completed data value set as a target, and training the numerical enrichment model by optimizing a preset loss function; Step S700: Perform a final data value assessment on the high-dimensional data using the trained numerical enrichment model.

[0026] In this embodiment, as described in step S100 above, the high-dimensional data is subjected to dimensionality reduction processing, specifically, the high-dimensional data is converted into low-dimensional feature data by using a pre-trained encoder model. The encoder model can select a mainstream deep learning structure, such as an autoencoder or a variational autoencoder (VAE), which can effectively extract low-dimensional potential feature representations of high-dimensional data. For example, for image data, the encoding part of a pre-trained DenseNet or ResNet model can be used to convert the original high-dimensional image data (such as ) is encoded into a low-dimensional feature vector (such as The specific process includes: inputting high-dimensional data into the encoder model, the encoder model extracts key features through forward propagation, and outputs low-dimensional feature representation; all low-dimensional feature data form a set , providing input for subsequent steps. This dimensionality reduction process can effectively reduce the complexity of the data while retaining the main structure and semantic information of the data, improving the efficiency of subsequent processing steps.

[0027] In this embodiment, as described in the above steps S200 to S300, the representativeness of the low-dimensional feature data is calculated, and the low-dimensional feature data is clustered based on the representativeness. The representative value of each sample in is calculated. The representative value is calculated by the similarity between samples. The similarity is represented by the Euclidean distance between samples. That is, the closer the distance between two samples, the higher the credibility of the two samples in representing each other. The formula is as follows:

[0028] in, Indicates The representativeness of the sample, and Respectively represent two samples in the low-dimensional feature data, Represents the total number of samples in the low-dimensional feature data set, Indicates that when calculating the representative value, The index of the other samples to compare with, From 1 to Traverse all samples in the low-dimensional feature data set.

[0029] Then, we use clustering algorithm to Clustering is performed to classify samples with high similarity into the same category. The commonly used KMeans clustering method can be selected as the clustering algorithm. For example, for a low-dimensional feature data set containing 1,000 samples, if each sample is a 128-dimensional feature vector, the KMeans method can be used to cluster it into 10 clusters, and the samples in each cluster have high similarity. After the calculation is completed, the representative value of each sample and the clustering result provide input for the subsequent steps to ensure that the most representative samples can be screened out.

[0030] It should be noted that the "representativeness" in this scheme refers to the ability of a data sample in a specific data set to fully reflect the main characteristics of the set. Representativeness is measured by the similarity between samples, and the similarity is reflected by the distance between sample features (such as Euclidean distance). Specifically, the higher the representativeness of a sample, the more similar the sample is to other samples in the data set, or it can more broadly represent the characteristics of a group of samples. In this scheme, representativeness is quantified by calculating the similarity weight (such as the inverse of the distance) between the sample and other samples. Samples with higher representative values ​​are considered to be the most representative samples in the cluster, which can minimize the amount of calculation for subsequent data value assessment while ensuring the accuracy and comprehensiveness of the assessment.

[0031] The most representative sample is selected from the clustering results of step S200 to form a representative sample set. Specifically, for each cluster, the representative value calculated in step S200 is used. , find the sample with the highest representative value in each cluster. For each cluster, the sample with the largest representative value is regarded as the sample that best represents the characteristics of the cluster, and these samples are combined into a representative sample set For example, suppose 1000 low-dimensional feature samples are divided into 10 clusters by KMeans clustering, each cluster contains 100 samples, then in each cluster, ,choose The largest sample is used as the representative sample of the cluster. Finally, the samples selected from the 10 clusters form a representative sample set containing 10 samples. This method effectively reduces the sample size while ensuring that the selected samples can represent the main characteristics of the entire data set.

[0032] In this embodiment, as described in step S400 above, the representative sample set constructed in step S300 The data value of the samples in the dataset is evaluated to obtain the data value of the representative samples. Specifically, the Shapley value method is used to evaluate the representative samples, that is, by combining each representative sample with other samples into different sub-datasets, the contribution of the sample to the performance of the machine learning model is evaluated to determine its data value. For example, for a classification task, suppose There are 10 representative samples in the dataset. Each sample can be combined with the rest of the samples to train the model and calculate the classification accuracy change of the model. The Shapley value of each sample is determined by the accuracy change to obtain its data value. After the evaluation is completed, the Shapley values ​​of all representative samples are stored as data value annotations. This method significantly reduces the computational complexity of data value evaluation by giving priority to the evaluation of representative samples, while retaining the accuracy of the evaluation results.

[0033] Preferably, the method of using the preliminary evaluation method to evaluate the data value of samples in the representative sample set includes: Step S401: for each sample in the representative sample set, generate a sub-dataset including the sample and a sub-dataset excluding the sample; Step S402: Calculate the difference in the impact of the sub-dataset containing the sample and the sub-dataset not containing the sample on the model performance by training the machine learning model; Step S403, calculating the Shapley value of the sample according to the impact difference as its data value for preliminary evaluation; Step S404: Summarize the Shapley values ​​of all representative samples to form a preliminary evaluation result.

[0034] In this embodiment, as described in step S500 above, according to the representative sample set obtained in step S400 The data value evaluation results of the unevaluated samples are combined with the representativeness of other unevaluated samples to complete the data value of the unevaluated samples and form a complete data value set. Specifically, the initial data value of the unevaluated samples is set to 0, and then the Shapley value evaluation results of the representative samples are used to complete the value by combining the similarity between the unevaluated samples and the representative samples (such as the inverse of the Euclidean distance based on low-dimensional features). For example, assuming that there are 1,000 samples in the unevaluated sample set, each sample calculates the similarity weight with the nearest representative sample through low-dimensional features, according to the formula:

[0035] in, is the value of the completed data of the unevaluated sample, Representation sample With representative samples similarity (such as the inverse of the Euclidean distance), is the evaluation data value of the representative sample. By calculating the similarity between each unevaluated sample and the representative sample and taking the weighted sum, the complementary data value of the unevaluated sample is obtained. Finally, a complementary data value set is formed.

[0036] In this embodiment, as described in step S600 above, a deep learning model suitable for the data value assessment task is selected as the value enrichment model, such as a network structure based on DenseNet. The input of the model is high-dimensional data , the output is the predicted data value of a single channel During the training process, the data value set is completed in step S500 As the target value, the optimization objective is defined as minimizing the following loss function:

[0037] in, The model predicts The data value of each sample is is the corresponding value of the completed data. Through the optimization of the loss function, the model gradually learns the mapping relationship between high-dimensional data and data value. Taking image data as an example, the high-dimensional input can be the original image data (such as 224*224*3). Through the feature extraction module and single-channel output layer of DenseNet, the data value of each image is finally predicted. The training process is iteratively optimized and the model parameters are adjusted using the gradient descent method until the loss function converges to obtain a numerical enrichment model that can accurately predict the data value. This step provides a powerful learning ability for the final data value assessment, while further improving the evaluation efficiency and accuracy.

[0038] In this embodiment, as described in step S700 above, the original high-dimensional data Input to the trained numerical enrichment model In the example, the numerical enrichment model generates the predicted data value of each high-dimensional data according to the mapping relationship between high-dimensional data and data value learned previously. For example, for a set of high-dimensional image data (such as a set of images with a size of 224*224*3), the images are input into the trained DenseNet model one by one. The DenseNet model extracts image features and generates corresponding predicted data value scores through a single-channel output layer. For 1,000 images, the high-dimensional input will generate 1,000 predicted data value scores. , ,… These prediction scores, as the final evaluation results, can accurately reflect the value of each high-dimensional data sample. This process fully utilizes the reasoning ability of the numerical enrichment model and realizes fast and efficient high-dimensional data value evaluation.

[0039] Preferably, the training method of the numerical enrichment model includes: Step S601, taking the completed data value set as the target, inputting high-dimensional data into the numerical enrichment model; Step S602: adjust the numerical enrichment model parameters by optimizing the preset loss function so that the data value output by the numerical enrichment model is aligned with the non-zero value in the complementary data value set, and dynamically correct the zero value based on the similarity of low-dimensional features.

[0040] The present invention screens representative samples through feature clustering for preliminary value assessment, and realizes rapid value promotion based on the numerical enrichment model. Specifically, the method uses an encoder to reduce the dimension of high-dimensional data and then clusters it, selects representative samples to calculate their Shapley values ​​as preliminary evaluation results, and then uses a deep learning model (such as Densenet) combined with a designed loss function to spread the value of representative samples to similar unevaluated samples. This process significantly reduces the computational complexity by reducing the number of samples for directly calculating the Shapley value, while using data similarity to ensure the accuracy of value assessment.

[0041] According to another aspect of an embodiment of the present application, there is further provided an electronic device, including a processor and a memory, wherein the processor is configured to implement the steps of the method when executing a computer program stored in the memory.

[0042] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0043] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0044] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0045] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0046] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A data value assessment method, characterized in that: The steps include: Use pre-trained encoders to encode high-dimensional data into low-dimensional feature data; Based on the low-dimensional feature data, calculating the representativeness of each sample; Clustering the low-dimensional feature data based on the representativeness, and selecting representative samples from each cluster to form a representative sample set; Use a preliminary evaluation method to evaluate the data value of samples in the representative sample set to obtain the data value of the representative samples; Based on the preliminary evaluation results of representative samples and the representativeness of other unevaluated samples, the data value of unevaluated samples is supplemented to form a supplemented data value set; Constructing a numerical enrichment model, taking the high-dimensional data as input and the completed data value set as a target, and training the numerical enrichment model by optimizing a preset loss function; The final data value assessment of high-dimensional data is performed through the trained numerical enrichment model.

2. The data value assessment method according to claim 1, characterized in that: Methods for calculating the representativeness of each sample include: For each low-dimensional feature data, calculate the similarity measure between it and all other low-dimensional feature data; The sum of the similarity measures is taken as the representative value of the sample; wherein the similarity measure is calculated by the exponential function of the Euclidean distance between samples, and the calculation formula is as follows: in, Indicates The representativeness of the sample, and Respectively represent two samples in the low-dimensional feature data, Represents the total number of samples in the low-dimensional feature data set, Indicates that when calculating the representative value, The index of the other samples to compare with, From 1 to Traverse all samples in the low-dimensional feature data set.

3. The data value assessment method according to claim 1, characterized in that: The method of clustering the low-dimensional feature data based on the representativeness and selecting representative samples from each cluster to form a representative sample set includes: The low-dimensional feature data is divided into multiple clusters by a preset clustering algorithm, wherein the clustering algorithm is the KMeans method and the number of clusters is a preset value. ; In each cluster, the low-dimensional feature data are sorted according to their representativeness, and the sample with the highest representativeness is selected as the representative sample of the cluster; the representative samples selected from all clusters are aggregated to form a cluster containing A representative sample set of samples.

4. The data value assessment method according to claim 1, characterized in that: Methods for using preliminary assessment methods to assess the data value of samples in a representative sample set include: For each sample in the representative sample set, generating a sub-dataset including the sample and a sub-dataset excluding the sample; By training the machine learning model, the difference in the impact of the sub-dataset containing the sample and the sub-dataset excluding the sample on the model performance is calculated respectively; Calculate the Shapley value of the sample based on the impact difference as its data value for preliminary evaluation; The Shapley values ​​of all representative samples are summarized to form a preliminary evaluation result.

5. The data value assessment method according to claim 1, characterized in that: Based on the preliminary evaluation results of representative samples and combined with the representativeness of other unevaluated samples, the method of completing the data value of unevaluated samples to form a completed data value set includes: aligning the preliminary evaluation results of the representative samples with the index of the low-dimensional feature data; Initialize the values ​​of data not selected as representative samples to zero; According to the aligned index relationship, the preliminary evaluation result is merged with the initialized zero value to obtain a completed data value set.

6. The data value assessment method according to claim 1, characterized in that: The training method of the numerical enrichment model includes: Taking the completed data value set as the goal, inputting high-dimensional data into the numerical enrichment model; The parameters of the numerical enrichment model are adjusted by optimizing a preset loss function so that the data value output by the numerical enrichment model is aligned with the non-zero value in the complementary data value set, and the zero value is dynamically corrected based on the similarity of low-dimensional features.

7. The data value assessment method according to claim 6, characterized in that: The loss function is defined as: in, Represents the output of the numerical enrichment model The predicted data value of high-dimensional data, Indicates the corresponding completion data value.

8. The data value assessment method according to claim 1, characterized in that: The numerical enrichment model is a deep neural network, whose input is high-dimensional data and output is a single-channel value, and the single-channel value directly represents the value of the data; the deep neural network adopts a Densenet structure.

9. An electronic device, characterized in that: It comprises a processor and a memory, wherein the processor is used to implement the steps of the data value assessment method as claimed in any one of claims 1 to 8 when executing the computer program stored in the memory.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the data value assessment method according to any one of claims 1 to 8 are executed.

Images