Calculation network data utility quantification method, device, equipment, medium and program

Abstract

The invention belongs to the technical field of machine learning, and particularly relates to a computing network data utility quantification method and device, equipment, a medium and a program, and the method comprises the steps: obtaining a computing network data set; training a sample machine learning model used for executing the specified task based on a data alliance set to obtain trained model performance corresponding to the data alliance set; comparing the trained model performance corresponding to different data alliance sets to obtain marginal performance contributions of the target data unit; and determining data cost and training cost of the target data unit based on the data type, and determining a data utility quantized value of the target data unit according to the data cost, the training cost and the marginal performance contribution. According to the method, a data alliance set including a target data unit and a data alliance set not including the target data unit are used as training data for sample model training, so that energy efficiency quantification of small-scale data is realized in a data collaboration scene.

Description

Technical Field

[0001] This application belongs to the field of machine learning technology, and in particular relates to methods, devices, equipment, media and programs for quantifying the utility of computing network data. Background Technology

[0002] In the context of the digital economy, data has become a crucial resource for driving innovation and economic growth, especially within computing networks (CFNs), where the value of data elements is increasingly prominent. As a comprehensive trading ecosystem integrating data, computing power, and algorithms, the CSN aggregates diverse cross-domain data elements such as text, images, videos, and semantics. This data flows between suppliers, computing power providers, and demanders through the CSN trading platform, providing enterprises with market insights, driving intelligent decision-making, and automating services.

[0003] Typical uses of network data include supporting performance improvements for machine learning models (e.g., as training / fine-tuning / testing data). However, the diversity of data types (e.g., text, speech, images, videos, etc.) and machine learning task objectives (e.g., classification, regression, clustering, dimensionality reduction, etc.) makes it difficult to predict the utility of a specific data set or group. When the same type of data is applied to different tasks, there are objective differences in utility. When different types of data are applied to the same task, the degree of contribution of each type of data may be different. Even when the same type of data is applied to the same task, the degree of performance improvement to the machine learning model is still different due to the different content of the data.

[0004] These objective facts make it difficult for machine learning model developers to scientifically and accurately select appropriate types and content of data from the complex data provided by the internet when facing model performance optimization problems; and make it difficult for data providers to accurately assess the value of data, especially when facing different tasks.

[0005] Therefore, there is an urgent need to provide a method for quantifying the utility of computing network data. Summary of the Invention

[0006] This application provides a method, apparatus, device, medium, and program for quantifying the utility of computing network data, which can solve the problem of accurately quantifying the utility of computing network data for specific machine learning tasks.

[0007] In a first aspect, embodiments of this application provide a method for quantifying the utility of computing network data, characterized by comprising: Obtain a computing network data set; the data type of each data unit in the computing network data set is adapted to the same specified task; the data type includes text, image, video, or audio; A sample machine learning model for performing the specified task is trained based on the data consortium set, and the performance of the trained model corresponding to the data consortium set is obtained; wherein, the data consortium set is a subset of the computing network dataset; By comparing the performance of the trained model corresponding to different data consortium sets, the marginal performance contribution of the target data unit is obtained; wherein, the target data unit belongs to the first set, the first set is the complement of the second set in the third set, and the second set and the third set are the intersection of at least two different data consortium sets and the union of at least two different data consortium sets, respectively; The data cost and training cost of the target data unit are determined based on the data type, and the data utility quantification value of the target data unit is determined based on the data cost, the training cost, and the marginal performance contribution; the data utility quantification value is used to indicate the training utility of the target data unit for an unspecified machine learning model performing the specified task.

[0008] The beneficial effects of the embodiments in this application compared with the prior art are: Considering the diverse task characteristics of machine learning models, this paper discusses and analyzes the energy efficiency of multimodal data (text, images, videos, audio, etc.) for a specific task. The performance difference of the same model (i.e., the example machine learning model) after undergoing the same training operations (but with different training data) is used as the data basis for energy efficiency analysis, resulting in quantifiable and repeatable marginal performance contribution evaluation results. At the same time, considering the synergistic effect between multiple data units, the example model is trained using a data consortium set including the target data unit and a data consortium set excluding the target data unit, respectively. This enables the quantification of energy efficiency for small-scale data in a data collaboration scenario.

[0009] In one possible implementation of the first aspect, the step of comparing the performance of the trained model corresponding to different subsets to obtain the marginal performance contribution of the target data unit includes: Obtain the entire data consortium set including the target data unit. The corresponding post-training model performance ,in, It is the sequence number of the data consortium set that includes the target data unit, and ; Obtain the complete comparison set of the target data unit. The corresponding post-training model performance The comparison set From a set The set after deleting the target data unit; The marginal performance contribution of the target data unit is calculated. ,in, In the formula, , The computing network data set, , Sets , The number of elements in it.

[0010] In one possible implementation of the first aspect, the step of acquiring the computing network data set includes: Obtain the data type required for the specified task; The original multimodal data of the required task types in the scheduling network; The original multimodal data is subjected to adaptation operations for the example machine learning model to obtain data elements. The adaptation operations include word segmentation, stop word removal, and word vectorization of text data; size standardization, color space conversion, and feature extraction of image data; frame extraction, key frame filtering, and feature encoding of video data; and noise reduction, segmentation, and voiceprint feature extraction of speech data. Redundant features are eliminated based on the feature importance of the data elements to obtain the data units, and the computing network data set is constructed based on the data units.

[0011] In one possible implementation of the first aspect, the step of acquiring the computing network data set includes: If the set of all the data units satisfies the preset complex set condition, then the data units are grouped and constructed into at least two computing network data sets.

[0012] In one possible implementation of the first aspect, each of the computing network data sets includes a reference data unit, and the reference data units belonging to different computing network data sets are the same data units; The step of determining the data utility quantification value of the target data unit includes: The data utility quantification value of the target data unit is determined based on the data cost, the training cost, and the standardized marginal performance contribution; Wherein, the standardized marginal performance contribution is positively related to the ratio of the marginal performance contribution of the target data unit to the marginal performance contribution of the reference data unit in the computing network data set where the target data unit is located.

[0013] In one possible implementation of the first aspect, the post-trained model performance includes at least one of accuracy, precision, recall, F1 score, ROC curve, AUC value, confusion matrix, mean squared error, and root mean square error.

[0014] Secondly, embodiments of this application provide a device for quantifying the utility of computing network data, characterized in that it includes: The data acquisition module is used to acquire a computing network data set; the data type of each data unit in the computing network data set is adapted to the same specified task; the data type includes text, image, video or audio; The training example module is used to train example machine learning models for performing the specified task based on the data consortium set, and to obtain the performance of the trained model corresponding to the data consortium set; wherein, the data consortium set is a subset of the computing network dataset; The marginal contribution module is used to compare the performance of the trained model corresponding to different data consortium sets to obtain the marginal performance contribution of the target data unit; wherein, the target data unit belongs to the first set, the first set is the complement of the second set in the third set, and the second set and the third set are the intersection of at least two different data consortium sets and the union of at least two different data consortium sets, respectively; The utility quantification module is used to determine the data cost and training cost of the target data unit based on the data type, and to determine the data utility quantification value of the target data unit according to the data cost, the training cost and the marginal performance contribution; the data utility quantification value is used to indicate the training utility of the target data unit for performing the specified task on an unspecified machine learning model.

[0015] Thirdly, embodiments of this application provide a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the computer network data utility quantification method described in any of the first aspects above.

[0016] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the computer network data utility quantification method described in any one of the first aspects.

[0017] Fifthly, embodiments of this application provide a computer program product that, when run on a terminal device, causes the terminal device to execute the computer network data utility quantification method described in any of the first aspects above.

[0018] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the method for quantifying the utility of computing network data provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of the computing network data utility quantification device provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the terminal device provided in the embodiments of this application; Figure 4 This is a schematic flowchart of an optional implementation of step 106 of the data utility quantification method for computing networks provided in this application embodiment; Figure 5 This is a flowchart illustrating a method for pricing network data elements based on Shapley values, according to an embodiment of the present invention.

[0021] Figure 6 This is a schematic diagram illustrating the process of selecting an evaluation model in a data element pricing method based on Shapley values ​​according to an embodiment of the present invention.

[0022] Figure 7 This is a schematic diagram illustrating the process of traversing and evaluating the performance of data elements in a data element pricing method based on Shapley values ​​according to an embodiment of the present invention.

[0023] Figure 8 This is a schematic diagram illustrating the process of calculating the Shapley value of a data element in a data element pricing method based on the Shapley value according to an embodiment of the present invention.

[0024] Figure 9 This is a schematic diagram illustrating the process of forming data element pricing by combining Shapley value and market demand in a data element pricing method based on Shapley value according to an embodiment of the present invention.

[0025] Figure label: Terminal equipment 30; Processor 301; Memory 302; Computer program 303. Detailed Implementation

[0026] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0027] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0028] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0029] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0030] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0031] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0032] like Figure 1 As shown in the figure, this application provides a method for quantifying the utility of computing network data, including: Step 102: Obtain the computing network data set; the data type of each data unit in the computing network data set is adapted to the same specified task; the data type includes text, image, video or audio; Step 104: Train a sample machine learning model for performing the specified task based on the data consortium set, and obtain the performance of the trained model corresponding to the data consortium set; wherein, the data consortium set is a subset of the computing network data set; Step 106: Compare the performance of the trained model corresponding to different data consortium sets to obtain the marginal performance contribution of the target data unit; wherein, the target data unit belongs to the first set, the first set is the complement of the second set in the third set, and the second set and the third set are the intersection of at least two different data consortium sets and the union of at least two different data consortium sets, respectively. Step 108: Determine the data cost and training cost of the target data unit based on the data type, and determine the data utility quantification value of the target data unit according to the data cost, the training cost, and the marginal performance contribution; the data utility quantification value is used to indicate the training utility of the target data unit for an unspecified machine learning model performing the specified task.

[0033] In this embodiment, the data type of each data unit in the computing network dataset depends on: 1. What types of data can be applied to this specified task? 2. What types of data units in the computing network data need to be evaluated?

[0034] Let's take the perspective of the computing network platform to introduce the application of this application's embodiments and related matters. Suppose that a batch of data is updated in the computing network platform. The computing network platform service provider can obtain the specific content of this data, but does not know what role this data can play in which machine learning model tasks. At the same time, this data contains a variety of data types (text, images, videos, audio, etc.), and it is also difficult to compare different types of data.

[0035] Now, if a computing platform service provider considers applying the solution of this application, it needs to determine the discussion status of the specified tasks in advance, and execute steps 102 to 108 of this application for the specific tasks of interest to obtain the utility of each data unit under multiple specified tasks. The definition of the specified tasks can be relatively general—classification, regression, clustering, dimensionality reduction—or it can be defined for specific scenarios—such as natural language processing, human / vehicle / object recognition in autonomous driving / assisted driving scenarios, etc.

[0036] Therefore, for the same data unit, its utility under different tasks can reflect its value in different scenarios. For the same specified task, the utility of each multimodal data can reflect which data type(s) is more critical to improving the performance of the model corresponding to that task.

[0037] In an optional implementation, the post-trained model performance includes at least one of accuracy, precision, recall, F1 score, ROC curve, AUC value, confusion matrix, mean squared error (MSE), and root mean square error (RMSE).

[0038] In practical applications, the performance of the trained model will vary depending on the specific task. The aforementioned performance data can be used in a weighted combination.

[0039] It is worth noting that the data in the computing network is external data provided by the data provider. It is not artificially set or fabricated, but rather belongs to the category of data that conforms to objective natural laws and is independent of human will.

[0040] For video, image, or similar data, the nature of the data makes the characteristics of its "technical means" relatively clear; for text, audio, or similar data, the generation process may involve human participation (e.g., conversations between natural persons, articles or comments written by natural persons as authors), but as the generation process ends (e.g., recordings are uploaded to the internet, articles are published), these data are "frozen" and transformed into objective data that cannot be changed by human will, which can be used as training data for machine learning models, and the characteristics of "technical means" become clear.

[0041] In some alternative implementations, the data consortium set in step 104 is a proper subset of the computing network dataset. In other alternative implementations, the data consortium set may be the same as the computing network dataset. In both implementations, the data consortium set may be an empty set.

[0042] In an extreme example, the data consortium set may contain only the target data unit, i.e., step 104 is performed separately by the target data unit itself, and the data after training the sample model separately by the target data unit is compared with the empty set (i.e., the untrained sample model). In this example and some other optional examples, the sample machine learning model should be a model with certain specified task capabilities.

[0043] The aforementioned extreme example considers the performance improvement capability of the target data unit itself for a specific task model. In more widely used scenarios, data units do not play an independent role, and it is necessary to consider the performance improvement capability of the model after the data units are combined.

[0044] At this point, we can consider constructing a data consortium set that does not contain the target data unit, using it as the execution subject of step 104 to obtain its corresponding post-trained model performance, then adding the target data unit to this data consortium set to form another data consortium set, and executing step 104 again. By evaluating the combined effect of the target data unit and other data units in the aforementioned data consortium set through the post-trained model performance obtained from executing step 104 twice, we can assess the combined effect of the target data unit and other data units in the aforementioned data consortium set.

[0045] Furthermore, the performance of the trained model can be compared using more data consortium sets to more comprehensively quantify the marginal performance contribution of the target data unit. Based on this idea, this application provides another preferred embodiment as follows.

[0046] like Figure 4 As shown, based on the foregoing embodiments, the step of comparing the performance of the trained model corresponding to different subsets to obtain the marginal performance contribution of the target data unit includes: Step 1062: Obtain the entire data consortium set including the target data unit. The corresponding post-training model performance ,in, It is the sequence number of the data consortium set that includes the target data unit, and ; Step 1064: Obtain the complete comparison set of the target data unit. The corresponding post-training model performance The comparison set From a set The set after deleting the target data unit; Step 1066: Calculate the marginal performance contribution of the target data unit. ,in, In the formula, , The computing network data set, , Sets , The number of elements in it.

[0047] In this embodiment, the execution idea of ​​step 106 is similar to the benefit distribution strategy in cooperative game theory based on Shapley value. It also introduces weights and uses the concept of alliance to use the "cooperative" feature in the Shapley value algorithm to reflect the synergistic effect between different data units.

[0048] Generally, game theory models are rarely applied to the field of data processing because data is not a subject of action and does not possess the concept or ability of "game playing," which constitutes a technological bias. Specifically, in the Shapley value algorithm, data itself does not possess the concept or ability of "cooperation." However, the embodiments of this application construct a multi-element alliance analogous to collaborators by dividing data units. The performance of the model after training through combinations of data units and training of example models reflects the effect of collaborative training (cooperative output) between different data units (collaborators). Based on this, quantification is performed to reflect the collaborative utility of a specific data unit with other data units in the computing network dataset. This allows for a more accurate determination of the relatively fair quantitative value of the utility of the specific data unit, i.e., the target data unit, in unspecified data groups—that is, evaluating the collaborative ability of the target data unit based on the cooperative concept of Shapley value, in order to obtain a fair utility quantitative value that does not depend on specific collaborative data.

[0049] Under this approach, the more data units there are and the more comprehensive their types are in the computing network dataset, the higher the reliability of the utility quantification value of the target data unit. However, correspondingly, the number of data units in the computing network dataset is positively correlated with the computing power required for steps 1062 / 1064 on a factorial level. When the number of data units in the computing network dataset is too high, the excessive computing power requirement will cause the evaluation of data utility to lose its application significance.

[0050] Therefore, in a preferred embodiment, rather than increasing the number of data units in the computing network dataset, priority should be given to increasing the data type coverage in the computing network dataset, that is, ensuring that each computing network dataset contains as many possible data types as possible for the specified task.

[0051] In another preferred embodiment, the step of acquiring the computing network dataset includes: If the set of all the data units satisfies the preset complex set condition, then the data units are grouped and constructed into at least two computing network data sets.

[0052] When there is more than one dataset involved in the computing network, the following issues need to be considered: 1. Does it involve quantifying the utility of multiple simultaneous computing network datasets? 2. Does it involve the synchronization and comparison of data utility of prior computing network datasets?

[0053] The reason for these problems is that steps 102 to 108 essentially provide relative rather than absolute values ​​of the utility of each data unit in the computing network dataset. For data units belonging to different computing network datasets, their utility values ​​cannot be directly compared.

[0054] Therefore, in order to solve these problems, this application provides another preferred embodiment based on the foregoing embodiments. In this preferred embodiment, each of the computing network data sets includes a reference data unit, and the reference data units belonging to different computing network data sets are the same data units. The step of determining the data utility quantification value of the target data unit includes: The data utility quantification value of the target data unit is determined based on the data cost, the training cost, and the standardized marginal performance contribution; Wherein, the standardized marginal performance contribution is positively related to the ratio of the marginal performance contribution of the target data unit to the marginal performance contribution of the reference data unit in the computing network data set where the target data unit is located.

[0055] Based on this embodiment, data in the same batch containing a large number of data units can be evaluated for relative computing power saving (by splitting into multiple computing network data sets), while simultaneously comparing the utility of data units in different computing network data sets within the batch based on reference data units.

[0056] From a timeline perspective, cross-temporal comparison of data unit utility can be achieved by adding the same reference data units to different batches of data.

[0057] Furthermore, multiple reference data units can be set to correct the relative utility of data units in different computing network datasets; multiple reference data units can be reference data of different data types to better optimize the relative utility of cross-computing network datasets for data types; multiple reference data units can also be reference data of the same data type, and errors can be reduced by means or similar statistical methods to obtain a more accurate relative utility of cross-computing network datasets.

[0058] According to any of the foregoing embodiments, in yet another embodiment, the step of acquiring the computing network data set includes: Obtain the data type required for the specified task; The original multimodal data of the required task types in the scheduling network; The original multimodal data is subjected to adaptation operations for the example machine learning model to obtain data elements. The adaptation operations include word segmentation, stop word removal, and word vectorization of text data; size standardization, color space conversion, and feature extraction of image data; frame extraction, key frame filtering, and feature encoding of video data; and noise reduction, segmentation, and voiceprint feature extraction of speech data. Redundant features are eliminated based on the feature importance of the data elements to obtain the data units, and the computing network data set is constructed based on the data units.

[0059] Based on the foregoing embodiments, if data utility is regarded as the value of data and associated with the transaction price of computing networks, then the foregoing embodiments can be combined. The following will introduce this combined embodiment (a data pricing method based on Shapley value).

[0060] First, we will explain some existing technologies and their problems from the perspective of computer network transaction prices.

[0061] The diversity of data modalities and the uncertainty of value within computing networks make the rational pricing of these elements a pressing challenge. A reasonable pricing method for data elements is crucial for the development of computing networks; it not only promotes the effective utilization and circulation of data but also enhances market competitiveness and compliance, driving the healthy development of the entire computing network system. Furthermore, with increasingly stringent data privacy regulations, users must be more cautious when using or trading data within computing networks, further increasing the necessity of a reasonable pricing mechanism. Establishing an effective pricing mechanism can accurately reflect the true value of data and promote the sharing and circulation of data elements within computing networks, thereby improving the resource utilization efficiency of computing networks. Therefore, exploring reasonable pricing methods for data elements within computing networks has significant practical implications and market demand.

[0062] The existing data-oriented pricing methods mainly include the following: 1) Cost-plus pricing is a traditional pricing strategy widely used for goods and services. In the context of data pricing, this method first requires determining the production cost of the data, including the costs of data collection, storage, processing, and analysis. Companies then add a predetermined profit margin to this to set the final price. The advantage of this method lies in its transparency and operability; companies can formulate relatively stable prices through explicit cost analysis, thus avoiding uncertainty in the pricing process. However, in computing networks, cost-plus pricing has significant limitations. First, the production cost of data is often difficult to define accurately, especially when real-time data streams and big data analytics are involved, as costs can fluctuate significantly due to the diversity of data sources. Second, this method ignores the market demand and potential value of data, potentially leading to the undervaluation of high-value data. For example, some data can generate significant economic benefits in specific application scenarios, but cost-based pricing may not reflect this. Furthermore, as market competition intensifies, companies using market-oriented or value-based pricing methods may gain an advantage, putting companies using cost-plus pricing at a disadvantage. 2) Market-based pricing methods draw on auctions, pricing models, and supply and demand principles. Common methods include auction-based pricing, fixed pricing, and dynamic pricing. In auction-based pricing, data sellers allow buyers to determine the price through open bidding; fixed pricing allows data suppliers to set prices based on data type, quality, and scarcity; dynamic pricing adjusts prices according to real-time changes in market demand. These methods are suitable for data markets and data trading platforms. However, in computing network scenarios, data supply and demand may be dynamic, requiring market mechanisms to adjust prices in real time, which demands robust infrastructure and rapid response capabilities. Furthermore, the asymmetry in the perceived value of data between buyers and sellers, especially when data quality cannot be directly assessed, can easily lead to market failures or information asymmetry.

[0063] 3) Data quality-based pricing methods price data based on quality metrics such as completeness, accuracy, timeliness, scarcity, and availability. For example, real-time data (such as sensor data or frequently updated transaction data) is typically more expensive than historical static data. Data quality can be assessed through data preprocessing and quality inspection methods. However, in computing networks, data quality evaluation can vary depending on the context; for instance, some applications prioritize real-time data, while others focus more on data completeness. Furthermore, data streams in dynamic environments can be affected by network latency and jitter, introducing uncertainty into quality-based pricing. Quantifying the correlation between quality metrics and price is also a challenge.

[0064] 4) Pricing based on data access methods: This method prices data based on the access method (e.g., data download, API call, data stream transmission). Typically, pay-per-use or pay-as-you-go pricing is the primary approach. For example, data APIs are charged based on the number of calls, and data stream transmissions are charged based on traffic. This method is suitable for small-scale scenarios requiring frequent data access. However, in computing network scenarios, billing based on data access methods relies on efficient and secure recording and tracking mechanisms. In existing computing network environments, due to the distrust between multiple parties and the dynamically changing network environment, the tracking and verification of data access may be delayed or erroneous. Furthermore, pricing solely based on access methods may ignore the actual value of the data itself, leading to unreasonable pricing.

[0065] 5) Dynamic pricing is a pricing strategy based on factors such as real-time data streams, changes in market demand, and user behavior. Computing network trading platforms can adjust data prices in real time through big data analytics and machine learning technologies. This method enables flexible pricing, increases revenue potential, and adapts to rapidly changing market environments. However, in computing networks, the limitations of dynamic pricing primarily lie in the need for robust data analytics capabilities and technical support. Enterprises must possess the ability to monitor in real time and react quickly to adjust prices promptly in response to market changes. Furthermore, dynamic pricing can lead to customer dissatisfaction, especially when customers discover price differences for the same data purchased at different times, potentially impacting customer loyalty and trust. Simultaneously, dynamic pricing may also raise compliance risks, particularly given increasingly stringent data privacy regulations; enterprises need to ensure their pricing strategies comply with relevant laws and regulations.

[0066] While existing pricing methods for data elements in computing networks each have their own characteristics, they generally share some common limitations. In computing network scenarios, many traditional pricing methods often fail to accurately assess the actual value of data. This is because data value depends not only on its production cost or market price but also on multiple factors such as data quality, usage scenarios, and market demand. Secondly, the data environment in computing networks changes rapidly, and the value and demand for data are often dynamically evolving. However, many pricing methods lack sufficient flexibility to adjust prices in real time to reflect changes in the computing network trading market. Finally, different tasks may perceive significantly different values ​​for the same data, and existing pricing methods often fail to adequately consider this, making it difficult to meet the needs of different users and thus impacting customer satisfaction.

[0067] The data pricing method based on Shapley value provided in this embodiment aims to overcome these limitations. Shapley value, derived from cooperative game theory, can rationally allocate the value of data in different application scenarios, ensuring a fair distribution of benefits between data providers and users. This is particularly important for multi-party cooperation in computing networks. Furthermore, the Shapley value method possesses dynamic adaptability, reflecting the contribution of data in a specific context in real time, thereby enhancing the flexibility and adaptability of pricing. Simultaneously, through Shapley value, users in computing networks can more accurately assess the actual utility of data elements in specific applications, maximizing the exploitation of data potential and increasing returns. Therefore, a Shapley value-based pricing method for computing network data elements provides a more reasonable and flexible solution for pricing computing network data elements, promoting the effective circulation and utilization of data.

[0068] The technical problem this embodiment aims to solve is the pricing of data elements in computing networks. Addressing the difficulty of existing pricing systems accurately reflecting the actual value of data in complex computing network environments, this embodiment introduces the Shapley value to provide a scientific evaluation mechanism. This mechanism can calculate the utility of data elements in different tasks based on the complex scenarios of the computing network, effectively resolving the unfairness inherent in the distribution of benefits among multiple parties in computing network transactions. A Shapley value-based pricing method for computing network data elements ensures the reasonable reflection of data element value and helps enhance the transparency and trust in data element transactions within the computing network.

[0069] In computing networks, the value and demand for data often change rapidly. The Shapley value-based approach has the ability to dynamically adjust pricing, responding in real-time to changes in market demand and data quality. This method addresses the issue of unequal distribution of benefits in data transactions by reasonably allocating the interests of all parties involved in data use. The calculation of the Shapley value ensures a fair distribution of benefits among data providers, users, and computing network platforms, enhancing the stability and sustainability of cooperation.

[0070] Please see Figure 5 The present invention discloses a method for pricing network data elements based on Shapley values, which includes the following steps: S1: In the computing network transaction scenario, clearly define the specific task type, collect and preprocess cross-domain multimodal data elements in the computing network, filter out data that meets the task characteristics, and standardize it; the steps include: S11: Clearly define the specific problems to be solved in the computing network transaction scenario, define task types such as intelligent customer service, content review, smart city analysis, financial risk control, etc., and identify the required multimodal data element types such as text, image, video, voice, etc. S12: Identify relevant multimodal data sources in the computing network, including databases, APIs, public datasets, and private datasets. Use direct scheduling or request scheduling to extract raw multimodal data related to the task, ensuring data diversity to meet the computing requirements of specific computing network transaction scenarios. S13: Specialized preprocessing is performed for different modalities of data, including removing duplicate text, filling in or deleting missing values, handling missing values ​​using mean, median or interpolation methods, segmenting text data, removing stop words, and vectorizing words; standardizing image data, converting color spaces, and extracting features; extracting frames, selecting keyframes, and encoding features for video data; and denoising, segmenting, and extracting voiceprint features for speech data. S14: Use correlation analysis in statistical analysis and random forest feature importance in machine learning methods to evaluate the importance of each feature, screen out the features most relevant to the given computing network transaction task and eliminate redundant features; S15: Standardize or normalize numerical features, use one-hot encoding or label encoding for categorical features, and convert categorical data into numerical format; S16: Assess the integrity and consistency of data elements, detect outliers and data distribution through visualization methods, ensure that data elements meet the requirements of the computing network transaction scenario, and record each step of data processing for subsequent review and traceability.

[0071] S2: Select a suitable learning model for initial training, and use a portion of the dataset for model training to capture the relationships between the data; Figure 6 This invention illustrates the process of selecting an evaluation model; the steps include: S21: Based on the task to be solved and the characteristics of the data, select an appropriate learning model for initial training. When selecting and evaluating a model, consider its complexity, interpretability, and ability to capture feature interactions. First, for specific task types in the computing network transaction scenario, analyze the data modality types and feature dimensions required by the task. For example, intelligent customer service tasks need to process multimodal data such as text, voice, and images, so a model that can effectively fuse these modalities should be selected. Next, for scenarios with high real-time requirements, such as intelligent customer service, choose a model with lower computational complexity. For scenarios with high accuracy requirements, such as financial risk control, a model with higher computational complexity can be selected. Then, in scenarios requiring transparent decision-making, choose a model with high interpretability. In scenarios requiring high accuracy, interpretability can be appropriately sacrificed. Finally, evaluate the model's ability to capture multimodal feature interactions and prioritize model architectures that can effectively handle multimodal feature interactions.

[0072] S22: Divide the prepared modal data feature set into training and test sets to ensure the training set covers data diversity. Use random sampling to avoid sample bias and ensure fairness in training and validation. Select an appropriate radius parameter. First, divide the data according to different modal types to ensure the prepared modal data feature set is complete and clear. Remove missing and outlier values ​​in each modality to improve the model training effect. Next, use random sampling to divide each modality dataset into training and test sets, with the training set accounting for 70%-80% of the total data and the test set accounting for 20%-30%. Then, ensure the training set covers data diversity, i.e., the proportion of each feature in the training set should be similar to that of the overall dataset. This can be achieved through stratified sampling. Use random sampling to reduce sample bias and ensure fairness in model training and validation. Select an appropriate radius parameter to optimize the data distribution in subsequent model training.

[0073] S23: Initially train the model using the selected training set, input feature data, optimize model parameters through the defined loss function, monitor changes in the loss value during training, and focus on how the model handles the influence of different modal data elements to accurately reflect the contribution of each feature in the subsequent pricing process; first, input the segmented training set feature data into the selected learning model, ensuring that the data format is consistent with the model requirements; then, select an appropriate loss function (such as mean squared error, cross-entropy, etc.) according to the task type to evaluate the gap between the model output and the true value; then, adjust the model parameters through optimization algorithms such as backpropagation (such as gradient descent) to reduce the value of the loss function, monitor changes in the loss value during training, and ensure that the model gradually converges during training; during training, pay particular attention to the impact of the model on different modal data elements, analyze how features affect the prediction results, and accurately reflect the contribution of each feature in the subsequent pricing process.

[0074] S3: Traverse all subset spaces composed of all data elements and calculate the model performance under different subset combinations; Figure 7 This invention illustrates the process of traversing and evaluating the performance of data elements; the steps include: S31: For all modal data elements involved in the relevant task, use combinatorial mathematics to generate all possible subset combinations, iterating through the subsets formed by different combinations of each individual modal data element; first, identify all single modal data elements related to the task, which may include different types such as numerical and categorical features strongly correlated with the task; use combinatorial mathematics to generate all possible subset combinations, which can be done by calculating the number of combinations. This is achieved by: where ... S32: For each generated subset combination, evaluate the performance of that subset on the test set using the selected learning model, and quantify the model performance using accuracy, mean squared error, R² value, or other relevant metrics. First, evaluate the test set using the selected learning model to calculate the model performance for that subset. Then, use the model to make predictions by inputting feature data. When selecting performance metrics, accuracy can reflect the proportion of correct predictions by the classification model, mean squared error measures the difference between predicted and actual values, and R² value measures the model's ability to explain data variations; the closer the value is to 1, the better the model fit. Finally, record the model performance metrics for each subset combination for subsequent analysis and comparison.

[0075] S4: Evaluate the marginal contribution of each data element through performance changes and calculate the Shapley value; Figure 8 This invention illustrates the process of calculating the Shapley value of data elements; the steps include: S41: For each modal data feature, compare the performance difference between the subset containing the feature and the subset not containing the feature to obtain the marginal contribution value; First, for each data feature, generate subsets that include the feature and subsets that do not include the feature. These subsets can be specific instances of all combinations generated in the previous steps; Use the selected learning model to evaluate the performance of these two subsets respectively, and calculate the model's performance metrics (such as mean squared error, accuracy, etc.) through the test set; Calculate the marginal contribution value, which is equal to the performance of the subset containing the feature minus the performance of the subset not containing the feature, reflecting the unique contribution of the data feature to the model performance; S42: In all subset combinations, count the frequency of the marginal contribution of each modal data element to provide the necessary weight information for subsequent Shapley value calculation; S43: Based on marginal contribution and frequency information, calculate the weighted average of the marginal contribution of each modal data element in the possible subset according to the Shapley value formula; S44: For each data feature, iterate through all possible subset combinations and calculate its marginal contribution in different combinations; the calculation formula is as follows: , in i For the first task in a given task i Each data element n The total number of data elements. D For the overall data elements, the global perspective S For the currently traversed subset of data, U() The model being trained; S45: Based on the calculated Shapley values, evaluate the current data element pricing strategy. If the Shapley values ​​of some data elements are significantly lower than those of other elements, it indicates that the pricing of these data elements should be lower for a given task, and the pricing of these elements can be reduced. Based on the evaluation results, optimize the data element pricing strategy.

[0076] S5: Set a base price for each data element based on the Shapley value, and make dynamic adjustments to the base price in combination with changes in market demand to ultimately form a complete pricing strategy; Figure 9 The following steps illustrate the process of combining Shapley value and market demand to form data element pricing according to the present invention. S51: Based on the calculated Shapley value, set a base price for each data element, and standardize the Shapley value of each feature to a specific range (e.g., 0 to 100) to facilitate pricing; this process can be expressed by the following formula: ; S52: After determining the base price, conduct market demand analysis, dynamically adjust the base price, assess demand changes at different price levels, and ensure that the pricing strategy can adapt to market changes; develop differentiated pricing benchmarks for different modal data, select typical data samples for price verification, and ensure that the pricing benchmarks are consistent with market expectations; based on the verification results, fine-tune the pricing coefficients for each modality (e.g., 0.1 for text data, 0.5 for image data, etc.); record the pricing benchmark adjustment process and basis to ensure the transparency of the pricing mechanism; dynamically adjust the base price, assess demand changes at different price levels, including using methods such as price elasticity analysis to determine the impact of price changes on demand; based on the demand analysis results, if it is found that the demand for certain data elements increases or decreases significantly at a specific price level, the base price should be adjusted in a timely manner. S53: Combine the base price obtained based on Shapley value with the dynamic adjustment mechanism to form a complete data element pricing strategy in the computing network; document the complete pricing strategy, including base price, dynamic adjustment rules, market demand analysis methods, etc., establish a monitoring and feedback mechanism, and regularly evaluate the effectiveness of the pricing strategy; adjust the pricing strategy in a timely manner according to market feedback and sales data to cope with market changes; It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0077] Corresponding to the data utility quantification method of the computing network described in the above embodiments, Figure 2 A structural block diagram of the data utility quantification device for computing networks provided in an embodiment of this application is shown. For ease of explanation, only the parts related to the embodiment of this application are shown.

[0078] Reference Figure 2 The device includes: The data acquisition module 201 is used to acquire a computing network data set; the data type of each data unit in the computing network data set is adapted to the same specified task; the data type includes text, image, video or audio; The training example module 202 is used to train example machine learning models for performing the specified task based on the data consortium set, and to obtain the performance of the trained model corresponding to the data consortium set; wherein, the data consortium set is a subset of the computing network data set; Marginal contribution module 203 is used to compare the performance of the trained model corresponding to different data consortium sets to obtain the marginal performance contribution of the target data unit; wherein, the target data unit belongs to the first set, the first set is the complement of the second set in the third set, and the second set and the third set are the intersection of at least two different data consortium sets and the union of at least two different data consortium sets, respectively. The utility quantification module 204 is used to determine the data cost and training cost of the target data unit based on the data type, and to determine the data utility quantification value of the target data unit according to the data cost, the training cost and the marginal performance contribution; the data utility quantification value is used to indicate the training utility of the target data unit for performing the specified task on an unspecified machine learning model.

[0079] In an optional implementation, the marginal contribution module 203 includes: The consortium set parameter submodule is used to obtain all the data consortium sets including the target data unit. The corresponding post-training model performance ,in, It is the sequence number of the data consortium set that includes the target data unit, and ; The comparison set parameter submodule is used to obtain the complete comparison set of the target data unit. The corresponding post-training model performance The comparison set From a set The set after deleting the target data unit; The contribution calculation submodule is used to calculate the marginal performance contribution of the target data unit. ,in, In the formula, , The computing network data set, , Sets , The number of elements in it.

[0080] In one optional implementation, the data acquisition module 201 includes: The type acquisition submodule is used to obtain the data type required by the specified task. The scheduling submodule is used to schedule the raw multimodal data of the required task types in the computing network; The adaptation submodule is used to perform adaptation operations on the original multimodal data for the example machine learning model to obtain data elements. The adaptation operations include word segmentation, stop word removal, and word vectorization of text data; size standardization, color space conversion, and feature extraction of image data; frame extraction, key frame filtering, and feature encoding of video data; and noise reduction, segmentation, and voiceprint feature extraction of speech data. The redundancy elimination submodule is used to eliminate redundant features based on the feature importance of the data elements, obtain the data units, and construct the computing network data set based on the data units.

[0081] In one optional implementation, the data acquisition module 201 includes: The splitting submodule is used to determine if the set of all the data units satisfies the preset complex set conditions, and then group the data units into at least two computing network data sets.

[0082] In an optional implementation, each of the computing network datasets includes a reference data unit, and the reference data units belonging to different computing network datasets are the same data units; Utility quantification module 204 includes: The standardized utility submodule is used to determine the data utility quantification value of the target data unit based on the data cost, the training cost, and the standardized marginal performance contribution. Wherein, the standardized marginal performance contribution is positively related to the ratio of the marginal performance contribution of the target data unit to the marginal performance contribution of the reference data unit in the computing network data set where the target data unit is located.

[0083] In one alternative implementation, the post-trained model performance includes at least one of accuracy, precision, recall, F1 score, ROC curve, AUC value, confusion matrix, mean squared error, and root mean square error.

[0084] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0085] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0086] This application also provides a terminal device, such as... Figure 3 As shown, the terminal device 30 includes: at least one processor 301, a memory 302, and a computer program 303 stored in the memory and executable on the at least one processor. When the processor executes the computer program, it implements the steps in any of the above-described method embodiments.

[0087] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0088] This application provides a computer program product that, when run on a mobile terminal, enables the mobile terminal to implement the steps described in the above-described method embodiments.

[0089] If the integrated unit is implemented as 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, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0090] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0091] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0092] In the embodiments provided in this application, it should be understood that the disclosed apparatus / network devices and methods can be implemented in other ways. For example, the apparatus / network device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0093] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0094] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for quantifying the utility of network data, the method comprising: receiving a plurality of network data; and quantifying the utility of the network data. The method comprises: obtaining a set of algorithm network data; each data unit in the set of algorithm network data is adapted to the same specified task; the data type includes text, image, video or audio; training a sample machine learning model for performing the specified task based on a data alliance set to obtain the performance of the trained model corresponding to the data alliance set; wherein the data alliance set is a subset of the set of algorithm network data; comparing the performances of the trained model corresponding to different data alliance sets to obtain the marginal performance contribution of a target data unit; wherein the target data unit belongs to a first set, the first set is the complement of a second set in a third set, and the second set and the third set are the intersection of at least two different data alliance sets and the union of at least two different data alliance sets, respectively; determining the data cost and training cost of the target data unit based on the data type, and determining the data utility quantitative value of the target data unit according to the data cost, the training cost and the marginal performance contribution; the data utility quantitative value is used to indicate the training utility of the target data unit to the unspecified machine learning model for performing the specified task.

2. The method of claim 1, wherein the quantizing the utility of the grid data comprises, The step of comparing the performances of the trained model corresponding to different subsets to obtain the marginal performance contribution of a target data unit comprises: acquiring all the data coalition sets including the target data unit corresponding trained model performance wherein, is the data coalition set sequence number including the target data unit, and ; acquiring all contrast sets of the target data unit corresponding to the trained model performance wherein the contrast set is a set after the target data unit is deleted from the set ​ computing a marginal performance contribution of the target data unit wherein, wherein, , is the set of net data, , is the number of elements in the set , , respectively.

3. The method of claim 1, wherein the quantizing the utility of the grid data comprises: quantizing the utility of the grid data based on a number of grid data points. The step of obtaining the set of algorithm network data comprises: obtaining the required data type of the specified task; scheduling the original multi-modal data of the required task type in the algorithm network; performing an adaptation operation on the original multi-modal data for the sample machine learning model to obtain data elements, the adaptation operation including word segmentation, stop word removal, and word vectorization for text data, size standardization, color space conversion, and feature extraction for image data, frame extraction, key frame selection, and feature encoding for video data, noise reduction, segmentation, and voiceprint feature extraction for voice data; based on the feature importance of the data elements, excluding redundant features to obtain the data unit, and constructing the set of algorithm network data based on the data unit.

4. The method of claim 1-3, wherein, The step of obtaining the set of algorithm network data comprises: determining that the set of all data units satisfies a predetermined complex set condition, and grouping and constructing the data units into at least two sets of algorithm network data.

5. The method of claim 4, wherein the quantizing the utility of the grid data is performed by: Each set of algorithm network data includes a reference data unit, and the reference data units belonging to different sets of algorithm network data are the same data unit; ​ The step of determining the data utility quantitative value of the target data unit comprises: determining the data utility quantitative value of the target data unit according to the data cost, the training cost and the standardized marginal performance contribution; wherein the standardized marginal performance contribution is positively related to the ratio of the marginal performance contribution of the target data unit to the marginal performance contribution of the reference data unit in the set of algorithm network data where the target data unit is located.

6. The method of claim 1, wherein the quantizing the utility of the grid data comprises: quantizing the utility of the grid data based on a number of grid data points. The performance of the trained model includes at least one of accuracy, precision, recall, F1 score, ROC curve, AUC value, confusion matrix, mean square error and root mean square error.

7. A network data utility quantification device, comprising: The method comprises: The data acquisition module is configured to acquire an algorithm network data set; a data type of each data unit in the algorithm network data set is adapted to a same specified task; and the data type includes text, image, video, or audio; The training sample module is configured to train a sample machine learning model for performing the specified task based on a data union set, to obtain a performance of a trained model corresponding to the data union set; the data union set is a subset of the algorithm network data set; The marginal contribution module is configured to compare the performance of the trained model corresponding to different data union sets, to obtain a marginal performance contribution of a target data unit; the target data unit belongs to a first set, the first set is a complement of a second set in a third set, and the second set and the third set are an intersection of at least two different data union sets and a union of the at least two different data union sets, respectively; The utility quantification module is configured to determine a data cost and a training cost of the target data unit based on the data type, and to determine a data utility quantification value of the target data unit according to the data cost, the training cost, and the marginal performance contribution; the data utility quantification value is used to indicate a training utility of the target data unit to a non-specific machine learning model performing the specified task.

8. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the method of any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. The computer program is executed by the processor to implement the method of any one of claims 1 to 6.

10. A computer program, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1 to 6.

Images