Application of Machine Learning Algorithms for Identification of Key Criteria Groups in Public Tendering Proceedings in Poland

Identifying a consistent set of criteria is essential to structuralize decision support models. It is especially relevant for the transparency of such models in public proceedings, where identifying a consistent family of criteria is a key factor in their objectification. It is worth pointing out that the criteria in public proceedings in Poland have arbitrary content and weighting determined each time by the contracting authority. There are no dictionaries that would make it easier for experts to analyze the criteria used, their categories, their dependence on the type of purchase, or their relevance to the subject of the contract. Following these shortcomings, this research uses natural language processing (NLP) to classify the criteria involved. The dataset was obtained from the Supplement to Tenders Electronic Daily (TED). The classification tasks used lemmatization, embeddings, and cosine similarity between embeddings where distances of 0.6 and 0.8 were checked. For comparison, the Lingo clustering algorithm was used. The work analyzed 113 373 proceedings in which 25 535 unique criterion names were used. Using NLP made it possible to group the criteria into 473 to 1 414 groups with a precision coefficient of 0.74 obtained.