Abstract:
The study examines the specifics of training machine learning algorithms on small datasets and addresses the task of forming a training set with high representativeness. It is known that class imbalance in objects, typical for small datasets, negatively affects the performance of algorithms. To mitigate this issue, various data synthesis methods have been developed in machine learning to supplement existing datasets and equalize the number of objects per class. However, these methods do not solve the problem of insufficient representativeness. This article proposes a method for constructing a representative training dataset by specifying the distribution that best corresponds to reality. The distribution is formed for each feature within the informative areas. Informative areas contain characteristic values of features that are most significant for distinguishing classes of objects. The proposed method of constructing areas is based on the idea of gradual expansion, accompanied by an increase in the informativeness of the areas. At the same time, informativeness is understood as a measure reflecting how well objects of different classes can be separated using the considered area. To form a complementary dataset, a generation method has been developed. As a result of its application, the complementary dataset is combined with the original one and forms the specified distribution in the informative area. This distribution can be determined either based on expert knowledge about the subject area, if the true distribution is known, or obtained as a result of computational experiments aimed at finding the most effective option. The applicability of the method is demonstrated by solving the problem of determining the level of temperature anomalies of the mammary glands. It is shown that the considered temperature features are characterized by a normal distribution. Increasing the representativeness of the training set allowed training a classic classification algorithm – logistic regression – with an accuracy comparable to a multilayer neural network. This approach to the formation of a training dataset opens up the possibility of creating more transparent and interpretable artificial intelligence systems.
Keywords:machine learning, small datasets, data representativeness, data synthesis, neural networks, logistic regression.
Fulltext:
PDF file (2821 kB)