Author ORCID Identifier

https://orcid.org/0000-0002-0015-040X

Date of Award

5-12-2024

Document Type

Thesis

School

School of Computing

Programme

Ph.D.-Doctoral of Philosophy

First Advisor

Dr.B.Santhi

Keywords

Reinforcement Learning, Recommender Systems, Collection Rate, Debt Collection Agency, Credit Risk Analysis

Abstract

Reducing the percentage of defaulters who often skip payments throughout the debt collection process might help minimize losses in the banking industry. The debt collection process should be optimized to reduce the rate of defaulters and improve collection rates. Traditional Machine Learning algorithms focused on credit risk analysis, defaulter prediction, and forecasting the recovery rate of debt collection. Researchers are not currently prioritizing the analysis of debt collectors’ performance. The debt collector’s primary responsibility is to retrieve outstanding debts from consumers on behalf of the debt collection firm.

Examining debt collectors’ performance is essential to enhance collection efficiency in the dynamic debt collection process. This research uses the continual learning approach to categorize the debt collector’s performance using the Enhanced Deep Q Network with Continual Learning (EDQN-CL). These techniques adapt to evolving data patterns over time and incrementally update new performance categories as needed. The metrics used for analyzing the debt collectors’ performance such as collection percentage, and visit patterns (date, time of visit, time interval between visits, and number of visits). The proposed algorithm achieves 13.56% higher accuracy than the existing system for employee performance.

If the customer reaches higher levels of overdue, then the customer will be allocated to the external collection agency. The proposed recommendation systems prevent the customer from being transferred to an external agency by preventing the customers from reaching higher levels of overdue. This thesis systematically optimized debt collection management by evaluating the debt collectors’ performance and categorizing the credit risk analysis.

A flexible recommendation system built using a hybrid actor-critic algorithm with transfer learning to enhance the debt collection agency’s collection rate. To adapt to these dynamic changes in the customer risk category, the proposed Reinforcement Learning (RL) methodology creates a customer allocation file that matches the risk of the customer with the appropriate debt collector. Moreover, it suggested recommendations such as visit count, actual statements, date, time of visits, and the time interval between the visits. After implementing this FRS-DRL for six months in a real-world situation, the collection rate increased to 20.34% and the number of visits decreased to 16.30%.

Text generation has advanced with the emergence of Large Language Models (LLM) in the field of artificial intelligence. This research explores the use of prompt engineering to enhance the text-based explainable AI of Debt Collection Management (XAI-DCM) using LLM. To get the intended results, several prompting techniques were used, including multiple-shot and zero-shot prompting. Conducted the comparative analysis between the existing system with LLM-based explanation generation. The proposed XAI-DCM model achieves a 16.03% higher BLEU score and a 12.80% ROUGE score than the existing knowledge graph systems.

The implementation of intelligent chatbots for debt collecting has the potential to significantly enhance the success rate of the debt recovery procedure. This research focuses on the development of chatbots by leveraging different prompting techniques. Occasionally LLM can generate some undesired or irrelevant response, called a hallucinated response. Additionally, this research investigated hallucinations through the use of Parameter Efficient Fine Tuning (PEFT) techniques, including Low-Rank adaptation (LoRa). This model sets automatic payment reminders regarding the due date and, based on the customer response, offers flexible payment, understanding the financial situation of the customers.

Moreover, Demonstrated the effectiveness of prompt engineering through a comparative analysis between two LLM such as GPT 3.5 and Llama 2 7B chat model in terms of ROUGE and BLEU score. In Zero-shot prompting, the Llama chatbot achieves a 0.46% higher BLEU-4 score. Whereas in Multiple shot prompting, the GPT 3.5 model achieves 42% improvement in the BLEU-4 score. This thesis made the comparative analysis with the existing work chatbot system, and the proposed DCM chatbot model achieves a higher BLEU-4 score of 1.88% for Kaggle’s chatbot dataset.

The fine-tuned model using LoRa achieves a 20.77% improvement in the BLEU - 4 score. Figure 1 gives the overall debt collection optimization process. The proposed FRS - DRL system provides the optimal recommendations without human intervention and increases the collection rate by 20.34% while reducing visits by 16.30%. Debt collectors achieved a higher collection rate when they adhered closely to the recommendations provided by the FRS-DRL.

Download

Share

COinS