APPLICATION OF TRANSFORMER ARCHITECTURE (VASWANI ET AL) FOR SEQUENTIAL SIMPLIFICATION OF FEATURE SPACE GEOMETRY AND SIMULTANEOUS CONSTRUCTION OF UNIQUE VECTOR REPRESENTATIONS OF PANEL DATA OBJECTS

The main problem with any approach to machine learning is the single-phase nature of the entire process. The original feature space is transformed so that all data corresponds to more or less standard normal distributions, after which the model is launched, and the result is output data. However, this article deals with the other side of the issue. And what happens if you convert the original feature space in a nonlinear way and apply some algorithm to the new space? More specifically, this paper presents a technical combination of correcting space geometry and then solving the problem of generating vector images for panel data objects. At the end of the study, the results are compared both with the previously proposed models and with the algorithm itself, taking into account modifications of the loss functions. The result shows the advantage of co-training the considered space simplification models and models used to solve some subsequent problem (for example, vector image construction and object clustering). The main emphasis is placed on the analytics of panel data, however, the ideology is quite amenable to generalization to any direction where there are many descriptive vectors characterizing a certain object (to preserve generality, what kind of “object” is not specified).