MASTER Assignment
Exploring and Evaluating Alternative Models for Cross- Selling Recommendations
Type : Master M-BIT
Period: Feb, 2023 - Aug, 2023
Student : Adineh, M. (Marzieh, Student M-BIT)
Date Final project: July 24, 2023
Supervisors:
Abstract:
Today, many companies in the financial services sector face clients who experience a variety of difficulties in selecting the right products that align with their unique needs and preferences. Conventional approaches to product recommendations have demonstrated their limitations in providing suggestions, often leaving customers feeling overwhelmed by the vast array of options or unaware of alternative products that may suit their needs better. The present thesis addresses this challenge. Zooming in on the context of one global financial services company, namely ING which is a global company in the financial services sector, this master project sets out to propose a solution for providing personalized recommendations to customers for cross-selling financial products and maximizing untapped revenue potential. To achieve this objective, we adopted a Design Science grounded research process that included problem analysis, solution design, and solution evaluation.
This thesis makes a valuable contribution to strengthening the Transaction Services (TS) sales cash advisory team at ING by empowering them to offer customized recommendations to customers. The developed recommender system utilizes customer interactions and metadata, bridging the gap in cross-selling supplementary products and stimulating engagement and sales. By tailoring recommendations to individual preferences, the system benefits both new customers in search of suitable products and existing customers aiming to diversify their product portfolio.
The primary objective of this master thesis is to analyze and evaluate the performance of recommender systems in the banking sector, specifically focusing on cross-selling recommendations.
In this thesis, the literature review examines studies conducted in the banking sector to gain insights into the technologies and techniques used for generating personalized recommendations for clients. The literature review conducted in this thesis follows a systematic literature review approach, employing established methodologies and procedures. Various databases were utilized to gather relevant literature. The literature review outlines the methodology used for the review, including the search strategy and data extraction strategy. Moreover, the findings from the literature review indicate that developing recommender systems tailored to the banking domain requires considerations. These requirements provide valuable insights for designing effective recommender systems in the banking industry.
Employing the design science-based research process, the three main parts of this research are as follows. The problem analysis first explores the difficulties that banking organizations encounter when trying to promote appropriate products to clients. It focuses on the significance of customized cross-selling suggestions to raise client satisfaction and boost sales.
Second, a comprehensive strategy leveraging recommender system techniques, like collaborative filtering and content-based models, is proposed by the solution design. The design intends to produce personalized product recommendations by leveraging consumer data and preferences. The LightFM library is utilized to build a hybrid recommendation system that utilizes user-item interactions and metadata to create embeddings that capture individual user profiles and product characteristics. The suitability of the LightFM model, as well as other models like XGBoost and Decision Tree, is explored.
Thirdly, the performance of the developed recommender is evaluated through a series of experiments and comprehensive data analysis. Two experiments are conducted to test the model’s performance: Experiment 1 involves training the model on randomly selected samples of user-item interactions and evaluating its performance on the remaining interactions. Experiment 2 explores the model’s ability to perform well after being trained on a more diverse set of user-item interactions. Evaluation metrics, particularly the area under the curve (AUC), are utilized to assess the recommender system’s performance. By assessing the AUC scores, the performance of the developed model in accurately recommending relevant products is determined. The results of the experiments demonstrate the model’s ability to provide accurate recommendations, thus addressing the challenge of product selection and cross-selling in the banking industry.