Product return prediction Master's Thesis Projekt by Yielva Kunz

Fastlinks:
Introduction | The project | Results and findings | Programme Information | Contact | Info-Events 

Yielva Kunz, a graduate of the MSc in Applied Information and Data Science at Lucerne University of Applied Sciences and Arts, wrote her Master's thesis on artificial intelligence in e-commerce, focusing on the early prediction of product returns. 

First of all, please tell us something about yourself: What hashtags best describe you?

#EmbraceTheUnknown

Please tell us more.

What originally was my motto for learning during my studies has now become my approach to new ideas and adventures.

About your job: What do you do at Zurich Insurance?

I work at Zurich Insurance as a Business & Data Analyst, where I contribute to automation projects and implement innovative solutions for simplifying processes and making them more efficient.

What did you do before?

Before my current job, I worked as a data analyst in digital sales. During my Master's studies, I wanted to try something new, learn about a different industry and gain experience in a corporate environment. A Master's degree in Applied Information and Data Science opens up many career paths, and I'm still discovering what I enjoy the most.

Please tell us about your research project.

My Master's thesis focuses on ways of predicting product returns in real time, the rate of which continues to increase steadily. Many current models rely on historical user data, i.e. information about customers' past behaviour, to predict the likelihood of the item being returned. However, this type of data is not available to guest users or first-time visitors. My approach works exclusively with anonymised browsing data and aims to detect the risk of a product being returned already during the browsing session, thus making it possible to take proactive steps before the purchase is completed.

What data and method did you use, and what insights did you gain or do you hope to gain?

For my project, I worked with extensive anonymised clickstream log files from a European fashion retailer that covered 26 weeks of user browsing behaviour. From this data, I extracted information on clicks, scrolling activity, page views, time spent on individual pages, shopping cart interactions, product details, and many additional features, resulting in 29 variables across four categories.

I applied a long short-term memory (LSTM) model to capture time-dependent patterns that indicate the likelihood of a return. I then split the data into training and test sets using a rolling window approach to simulate real-world conditions. The next step involved feature permutations, accuracy dropout and integrated gradients with which I analysed the decisive factors.

Interaction and shopping-cart features proved particularly relevant. The signals indicating a return also differed partly between mobile and desktop users. Full returns were more frequent when shopping carts had a wide range of items, when discounts were applied, and when individual product prices were relatively low compared to the total cart value.

How can your insights help society?

The findings, based solely on anonymised browsing data, show strong potential to predict the product return risk in real time before the purchase is completed and thus offer economic as well as environmental benefits. Companies can proactively reduce product returns and lower their operating costs while decreasing their environmental footprint through fewer delivery trips and less packaging waste at the same time.

What are your goals for your project in future?

Future research could explore ways of further improving and scaling the predictions, for example by testing alternative models or incorporating additional features. The long-term goal would be to apply the research to practical applications so that companies can proactively develop strategies for reducing product returns. It would also make sense to examine how the approach can be used in different industries and product categories.

How did your studies in the Applied Information and Data Science programme influence the project?

The programme provided me with the ideal foundation and methods for completing the project successfully. In particular, the combination of data science, machine learning and practice-oriented case studies helped me to understand complex relationships and apply them directly in my research.

What advice would you give to others starting on similar projects?

Give yourself the time you need to understand your data fully. Only when you know your data well can you develop meaningful models, identify patterns, and generate valid results. At the same time, it is important to remain flexible and continuously question your findings throughout the project.

And finally: What new hashtag are you aiming for in future?

#Curiosity, because I hope to keep questioning the status quo with courage and remain curious about what lies ahead.

We want to thank Yielva Kunz for her dedication and for sharing these valuable insights.