Interpretable Embeddings for Geographic Transactional Activity Analysis

• Published in: Procedia computer science 2023, Vol.229, p.357-366
• Main Authors: Glukhov, Gleb, Zhdanov, Pavel, Shikov, Egor
• Format: Article
• Language: English
• Subjects: Client Representation Learning; Clustering; Customer Embeddings; Customer Feature Extraction; Customer Segmentation; Transactional Data
• ISSN: 1877-0509; 1877-0509
• DOI: 10.1016/j.procs.2023.12.038

Transactional customer embeddings have become one of the most popular methods for businesses to solve customer-related challenges. Despite the high quality, embeddings may lack interpretability, which makes it difficult to understand and extract the meaningful information from the received representations. To address this limitation, we propose an embedding framework that make use of users’ geographical transactional activity, which captures individuals’ spatial patterns and preferences. Our framework consists of three steps: calculating custom client geographic activity feature vectors, clustering these vectors, and calculating the resulting user low-dimensional vectors where each value represents the distance between the user geographic activity vector and the cluster center. Experiments on the partner bank dataset demonstrate the effectiveness of proposed method. Our approach achieves comparable performance to the existing methods in purchase category forecasting, task of missing category prediction and that of campaign targeting. Furthermore, the obtained model results can be visualized and interpreted, which makes it possible to understand the user behavior