Improving the Discovery of Technological Opportunities Using Patent Classification Based on Explainable Neural Networks
Purpose: The paper aims to present an approach supporting the improvement of technological opportunities discovery using patent classification based on explainable neural networks. Design/Methodology/Approach: Empirical research was conducted applying a dataset containing U.S. patent documents. Firstly, this dataset was checked for the correctness of the saved patent data to be further analyzed. Then, a custom Bidirectional Encoder Representations from Transformers (BERT) Neural Network was developed and trained. Finally, the Local Interpretable Model-agnostic Explanations (LIME) method was applied for interpreting the results achieved with the BERT classifier. Findings: The studied classifier achieved high quality (precision of 80.6%), allowing correct classification of the technologies described in the patents. Such neural classifiers are easy to use in practice and highly versatile; however, there is an insufficient trust of managers in the decisions suggested by that black-box method. The proposed new approach may help overcome the lack of trust of the users of neural models towards the technological opportunities suggested by them. Practical Implications: Various patent databases are often used to discover innovative solutions, as well as economic and technological opportunities, because they contain vast resources of prosperous and extensive information recorded in patent documentation. Such analyzes are critical to businesses and public organizations as they help them make decisions about carrying out strategic investment projects. The presented approach, which supports the improvement of automated processes of technological opportunities discovery, may increase confidence in the results obtained using neural classifiers. Originality/Value: Earlier studies focused mainly on using more effective classifiers and better learning algorithms. Progress in this type of research did not help solve the problem of the lack of reliable justification for individual decisions indicated by machine learning models. In this study, a proposal for an approach enables the discovery of technological opportunities using patent classification based on explainable neural networks.