EXPERIMENTAL DETERMINATION OF THE OPTIMAL PARAMETERS OF THE RECURRENT NEURAL NETWORK FOR THE TASKS OF PATENT CLASSIFICATION

UDC 004.032.26
DOI:10.26102/2310-6018/2019.25.2.027

A.G. Kravets, A.S. Burmistrov, P.A. Zadorozhny

Indicators of patent activity are now often used in technological forecasting and competitive intelligence. An important role is to predict the development of patent trends in individual countries and around the world, which allows to identify the main priority directions of technology development. Analysis of patents in the field of analog technology was fulfilled. The international patent classification is outdated, most studies are interdisciplinary. There is a need to select and create new classes. The purpose of this study is to analyze the parameters that affect the results of the recurrent neural network, designed for the thematic classification of the patent array. The analysis of the identified parameters affects the quality of the neural network and the selection of optimal values. The optimal parameters of the neural network were determined: the number of layers, the size of the layer, the value of the exclusion parameter, the batch size for training in the network, the choice of the Keras library optimizer was made. The reported study was funded by RFBR according to the research project № 19-07-01200.

Keywords: trend, classification, patent, recurrent neural network, exclusion layer, optimizer analysis, batch size

Full text:
KravetsSoavtors_2_19_1.pdf