University of Twente Student Theses


Exploring the applicability of a data-driven approach on recommending in-store replenishments for fashion retailers

Meinderts, S. (2021) Exploring the applicability of a data-driven approach on recommending in-store replenishments for fashion retailers.

[img] PDF
Abstract:Having the right articles on the sales floor is essential for maximizing the profit of a fashion retailer. Identifying size gaps on the sales floor using RFID counts leads to recommendations on what articles should be refilled from the in-store stockroom to the sales floor. Although effective, not all size gaps should be resolved whereas some articles should be refilled even when they are still present on the sales floor. This research investigates the applicability of a data-driven approach on predicting what articles should be refilled. In order to determine this applicability, a dataset was formed from scratch for a single fashion retailer, the features and labels of which were based on historic data. Different machine learning models were tested on this dataset, of which a mixture of a Wide and Deep classification model and a residual regression network turned out to be the best performing, achieving 0.48 and 0.44 on the test set for the F1 score and Matthews Correlation Coefficient (MCC) respectively. This was an improvement of the current refill implementation, which achieved 0.30 and 0.24. The feature attributions of the model and the differences between outputs for specific categories and stores imply that the model is able to learn how to differentiate between these different categories or stores. Measuring the performance in actual stores turned out to not deliver statistical significant result. The evaluation of the model in actual stores could have been more exhaustive, were it not that difficulties were introduced by the global pandemic. The performance of applying the trained model on the data of a different retailer was considerably worse, resulting in an F1 score and MCC of only 0.24 and 0.26 respectively. In hindsight the chosen retailer used for training the model turned out to be unfortunate given the noise introduced by their frequency of counting. All in all the data-driven approach showed promising results and is definitely worth pursuing once data can be collected with less noise and online testing can be performed on a larger scale.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science MSc (60300)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page