Fine-tuning a Large Language Model for Simulating Association Rule Mining Algorithms

The presented work deals with the topic of fine-tuning and subsequent usage of large language models for frequent itemset mining. The main goal of the work was to determine whether it is possible to fine-tune models for higher success rates in this specific task. Within the scope of the work, two models (Deepseek R1 Distill Qwen 7B and Qwen1.5 1.8B Chat) were fine-tuned on three datasets created by the author. For each model and dataset, fine-tuning was also carried out using two different training method settings. The fine-tuned models were then tested on four different datasets for frequent itemset mining. Each dataset was provided to each model in the prompt, and the prompt was subsequently generated using three different creativity settings of the model. The outputs of the models were compared against the results of the Apriori algorithm and subsequently against the results of the non-fine-tuned version of the respective model on the same task. The best results were achieved by the non-fine-tuned Deepseek R1, with a 14% success rate. This was followed by the fine-tuned Deepseek R1 models, with an average success rate approaching zero. Qwen1.5 consistently failed in the frequent itemset mining task. In the case of the fine-tuned models, most of the models broke down, and in the case of the non-fine-tuned model, the most common outcome was code generation instead of producing the expected output.