Automating Product Categorization and Price Predictions Using Artificial Intelligence

Sammanfattning

This thesis examines how Artificial Intelligence (AI) can be utilized to facilitate workflows in commerce through automating categorization and price prediction. The goal is to simplify the process of structuring input data and generating predicted outcomes based on extracted features. The final product is a prototype that utilize two AI methods to classify text and perform price prediction. It allows free text input of new article data, where the article features are extracted by a fine-tuned Large Language Model (LLM), and forwarded to a price predicting model that suggest a price for the article, based on those features. The LLM is trained on structured and categorized data, to extract article features from free text through Named Entity Recognition (NER), such as brands, article types and colors, with the purpose of categorize and thereby structuring data. The model is trained on both a small and a large dataset. The results show that even though both models perform with an accuracy above 90%, the model trained on the larger dataset is the best performing one, with a 96% accuracy, and an F1 score of 0.95. The second AI method is classical Machine Learning (ML) that performs price predictions based on the structured data. Different ML algorithms are compared, both regression and classification. The results of price predictions show that the preferred algorithm is Decision Tree Classification. It is the least demanding algorithm and it performs better than both the regression model and the more complex classification models. The accuracy of all models are below 70%, the accuracy of the Decision Tree model being 67%. The low accuracy of price predictions is a result of a large number of classes and inconsistencies in pricing, causing confusion during model training.