Article Preview
TopIntroduction
In the era of big data, data analysis is everywhere. Data Mining (DM) technique allows extracting knowledge from real-life data in various domains. Real-life data are complicated and messy so that different algorithms are suitable for different situations (Korzun, D., et al. (2019)). A critical problem is that there is no guide to help users for algorithm selection. Because of the complexity of the DM knowledge, the formulation of precise guidelines is often difficult or even impossible. Specialists rely on years of accumulated tacit experience, which is hard to express explicitly. Consequently, novice analysts are typically entirely overwhelmed. They have no idea which methods can be confidently used, and often resort to trial and error.
Aggravating the problem is the ever-increasing number of operators that few experts can keep up with (Serban, F., et al. (2013)). There is a dire need to support both novices and experts in their data analysis tasks. Traditionally, this kind of support is provided by experts/consultants. These are, however, often not available and also mired by the increasing number of methods. To address this issue, meta learning and semantic meta mining were proposed.
Meta learning (Jankowski, N., et al. (2011)) is defined as the application of machine learning (ML) techniques to past ML experiments. Its purpose is to modify certain aspects of the learning process to improve the performance of the results. Traditional meta-learning treats the learning algorithm as a black box, correlating the observed performance of the output model with the characteristics of the input data. However, the internal characteristics of algorithms with the same input/output type may vary.
Different from the traditional meta-learning, semantic meta mining (Hilario, et al. (2009)) was proposed for DM optimization via algorithm/model selection. It provides a suitable description framework to clarify the complex relationships between tasks, data, and algorithms at different stages in the DM process. Semantic meta mining mines DM metadata through querying DM expertise in knowledge bases. This expertise should be stored in a machine-interpretable format, so it can be automatically extracted and applied to new problems. Existing DM knowledge systems based on semantic meta mining provide useful help regarding the sequence of operators in the processes of data analyses as well as choosing their parameters (Ristoski, P. (2019)). The majority of existing systems support only separate steps of the overall DM process or specific DM algorithm annotation (Ristoski, P., & Paulheim, H. (2016)); (Benali, K., & Rahal, S. A. (2017)). This limits the capabilities of such systems. It is necessary to provide help in addressing the overall data mining process from collecting raw data to extracting actionable knowledge. To be able to support the whole process, the system must know each step in correlation to the other ones.
In semantic meta mining, ontology as a description language is used for the representation of the knowledge base. Ontology is a computer-understandable description language. Ontology-based systems are widely used in IoT domain (Fathy, N., et al. (2019, December); Lebedev, S., & Panteleyev, M. (2020); Smirnov, A., et al. (2020); Vodyaho, A., et al. (2020)). Naturally, it is the best choice for building DM knowledge systems (Korzun, D. G., et al. (2016)) and human-computer interaction systems (Kashevnik, A., et al. (2018)).
Several DM ontologies are currently under development. They contain prior knowledge represented as meta data. Some of them are complimentary, some overlap. These ontologies are oriented on expressing one or several stages of the DM process in detail. There is no unified operating environment for them. To support the overall process of data analyses, a newly established DM ontology that contains knowledge about all steps of data analyses is required.
Another challenge in developing DM systems based on Semantic meta mining is the problem of raw data representation. To know which techniques can be applied, the system needs information about the received input data. This information should be some abstract complex information. It should be high-level properties, such as the entropy of the value distributions or the signal-to-noise ratio, but not basic properties, such as the number of attributes in tabular data or the number of missing values. Most of the existing DM knowledge systems tend to store simple information. Simplification of properties can have a negative impact on the results of the data analyses.