An Unified Definition of Data Mining

Since many years, theoretical concepts of Data Mining have been developed and improved. Data Mining has become applied to many academic and industrial situations, and recently, soundings of public opinion about privacy have been carried out. However, a consistent and standardized definition is still missing, and the initial explanation given by Frawley et al. has pragmatically often changed over the years. Furthermore, alternative terms like Knowledge Discovery have been conjured and forged, and a necessity of a Data Warehouse has been endeavoured to persuade the users. In this work, we pick up current definitions and introduce an unified definition that covers existing attempted explanations. For this, we appeal to the natural original of chemical states of aggregation.

💡 Research Summary

The paper begins by tracing the evolution of the term “data mining” from its early articulation by Frawley et al. (1999) as the automatic extraction of useful patterns from raw data, through the broader “Knowledge Discovery in Databases” (KDD) paradigm that later incorporated data preprocessing, modeling, evaluation, and deployment. While this expansion has made the concept more inclusive, it has also blurred the essential distinction between raw data, processed information, and the higher‑level knowledge or insight that practitioners ultimately seek.

A comprehensive literature review categorizes existing definitions into three families: (1) pattern‑focused definitions that emphasize statistical and machine‑learning techniques; (2) knowledge‑focused definitions that stress the role of domain experts and decision‑support; and (3) infrastructure‑focused definitions that treat data warehouses and ETL pipelines as indispensable. The authors point out that each family captures valuable aspects but none provides a unified, operationally clear framework, especially in light of contemporary privacy concerns that restrict access to the raw “data” layer.

To resolve this fragmentation, the authors introduce a novel metaphor drawn from the physical states of matter—solid, liquid, and gas—to model the transformation pipeline of data mining. In this model:

• Solid (Data) – The raw, unprocessed records collected from sensors, logs, or transactions. This stage emphasizes data acquisition, storage, and security, with an explicit focus on minimizing loss and preserving fidelity.
• Liquid (Information) – The result of cleaning, integration, normalization, and dimensionality reduction. Here, data warehouses, data lakes, and ETL processes act as the “container” that holds the fluid information, making it ready for analysis.
• Gas (Knowledge/Insight) – The outcome of modeling, pattern discovery, and interpretation. Machine‑learning algorithms, statistical tests, and visual analytics convert the liquid into diffuse, high‑level knowledge that can be disseminated and acted upon.

The authors propose quantitative metrics for each transition: transformation cost, information loss, and uncertainty reduction. By measuring these metrics, practitioners can evaluate the efficiency and effectiveness of the entire pipeline, rather than assessing isolated algorithmic components.

Two case studies illustrate the practical applicability of the unified definition. The first examines a retail scenario where click‑stream logs (solid) are cleaned into transaction tables (liquid) and then used to generate customer segmentation models and targeted‑marketing insights (gas). The second explores a healthcare setting where electronic health records (solid) are standardized into clinical variables (liquid) and subsequently fed into disease‑prediction models that produce treatment guidelines (gas). In both cases, the clear delineation of stages improves reproducibility, transparency, and the ability to audit privacy‑preserving transformations.

In the discussion, the authors argue that the solid‑liquid‑gas framework not only reconciles disparate definitions but also provides a scaffold for future research. They suggest that optimization techniques specific to each transition (e.g., lossless compression for solid‑to‑liquid, explainable AI for liquid‑to‑gas) and advanced privacy‑preserving mechanisms (differential privacy, federated learning) can be systematically integrated within this structure. Moreover, the framework can guide curriculum development, standardization efforts, and the creation of benchmark suites that evaluate end‑to‑end pipelines rather than isolated components.

The conclusion reaffirms that viewing data mining as a series of physical‑state transformations yields a coherent, extensible definition that captures the essence of the field while accommodating emerging challenges such as data governance, ethical AI, and cross‑domain interoperability. The paper calls for community adoption of this unified terminology to foster clearer communication, more robust methodology, and accelerated innovation in both academic and industrial contexts.