Abstract: As an emerging factor of production, the demand for trading and circulation of data resources has shown explosive growth. The problem of data quality has sparked widespread concern along with the exponential growth of data scale, and a lot of low-quality data is flooding into different types of data resource trade platforms. How to identify high-quality data resources in the massive resources has become the key for data trading platforms to gain competitive advantages and improve the efficiency of factor allocation. Existing research has provided a basis for high-quality data identification in the platform trading context, but there are still two deficiencies: Firstly, it is challenging to meet the requirements for large-scale data resources’quality identification because the existing identification methods, which are only applicable to the quality evaluation of small-scale and homogeneous data resources, have more manual participation components and insufficient automation. Secondly, the existing identification methods ignore the problem of uneven distribution of data resources of different quality, which easily triggers the bias of classification results and is difficult to meet the robustness requirements of heterogeneous sample classification. This paper is to clarify the mechanism of high-quality data resource formation in the context of platform transactions, discover the key factors for high-quality data resource formation in the context of platform transactions, and propose a method for efficiently recognizing high-quality data from large-scale and heterogeneous data resources.
Data circulation and transaction are necessary for the realization of the value of data resources, and in a platform economy, data circulation is in the form of an open market with numerous participants. This paper studies the flow of data resources and the generation process of high-quality data within the platform environment and builds a high-quality data resource identification index system of “intrinsic quality-commodity characterization”. After that, it suggests a K-medoids-NCA-SMOTE-BSVM fusion model, handles the identification of high-quality data as a pattern recognition problem, and uses supervised machine learning to identify high-quality data and solve the issue. Four primary sections make up the model: (1)Synthesizing the number of views, collections and downloads of data resources as the basis of discrimination, using K-medoids to cluster the samples of data resources, automatically creating the classification labels for data resources, and calculating the ideal number of classification labels by combining them with profile coefficients. (2)The built metrics are downscaled using Nearest Neighbor Component Analysis (NCA) to come up with a new set of features, taking into account the possibility that the chosen metrics contain elements that are less important for the categorization of high-quality data resources and may therefore influence the model’s effectiveness. (3)After clustering division, the number of samples from different classes fluctuates widely, and therefore, to maintain the balance of data between classes, the few class oversampling method (SMOTE) is used to increase the amount of data from the few sample classes under the new feature set. (4)A nonlinear high-quality data resource identification model based on Bayesian optimization support vector machine (BSVM) is constructed, and it achieves classification prediction of high, medium, and low quality data of various calibers by using the data resource identification indexes after feature dimensionality reduction, the clustered data resources as input, labeling the clustered data resources with categories, and balancing the dataset as the model’s output. Finally, based on the API datasets of real data trading platform, Python is used to crawl the request parameters, return parameters, update frequency, data sources, data descriptions, labels, application scenarios, specifications, registered capital of service merchants, views, downloads, and favorites of data resources to carry out the empirical research.
The results show that: a)SMOTE balanced processing can improve the effect of data resource quality identification and improve the classification accuracy of the optimization model based on the comparison of unbalanced with balanced datasets. b) Whether based on imbalanced or balanced datasets, BSVM outperforms SVM, WOA-SVM, PSO-SVM, MLP, and CNN approaches in terms of prediction accuracy, and BSVM has higher algorithmic efficacy with less training time than other optimization algorithms. In summary, this paper, which is an innovative attempt and a significant addition to the theory of data resource quality assessment, clarifies the meaning of high-quality data resources, builds a high-quality data resource identification index system, and fully verifies the validity of the index system with the aid of trading platform data. It also builds a high-quality data resource identification model, which can effectively generate the quality labels of massively parallel data sets. It has significant guiding relevance for encouraging the active trading of data resources and can effectively develop quality labels for vast data resources. It can also increase the recognition accuracy of heterogeneous data resources.

Key words: data trading platform, high quality data, K-medoids-NCA-SMOTE-BSVM, multi model integration

摘要： 随着数据服务形态不断衍生,数据资源作为一种新兴生产要素,其交易流通需求呈现爆发式增长。如何从海量数据中识别高质量数据资源,挖掘要素价值,成为数据交易平台获取竞争优势以及提升要素配置效率的关键。本文旨在发现平台交易情境下高质量数据形成的关键因素,提出从大规模、异质数据资源中高效识别高质量数据的方法。首先,基于高质量数据形成过程,构建“固有品质-商品表征”二维识别指标体系;然后,提出K-medoids-NCA-SMOTE-BSVM融合模型,对高、中、低三类不同质量数据进行分类预测;最后,收集真实数据交易平台的API交易数据,开展实证研究。结果显示:相比SVM,WOA-SVM,PSO-SVM,MLP和CNN等方法,K-medoids-NCA-SMOTE-BSVM模型在预测准确率和训练时间方面,均有良好的性能表现。本文提出的识别指标及分类模型,为平台经济下数据质量判断与预测提供了依据,对产品视角下数据质量标准制定以及数据交易定价优化具有一定实践意义。

关键词: 数据交易平台, 高质量数据, K-medoids-NCA-SMOTE-BSVM, 多模型集成