[2Win5] Poster session 2

[2Win5-90] PersonaForecast: Interpretable Feature Engineering via Vector Embeddings

〇Kensuke Takatani¹, Katsuya Ito¹, Yoshihiko Ichikawa² (1.INDX & Co. Ltd., 2.Insight Edge, Inc.)

Keywords:Recommendation, Machine Learning

This paper proposes PersonaForecast, a generative AI-based method that integrates and analyzes diverse customer data to automatically discover and visualize novel personas. The approach begins by vectorizing heterogeneous customer information, thereby mapping various data types into a unified numerical space. Subsequently, a clustering technique is employed to group customers according to a similarity measure, facilitating the efficient extraction of complex yet interpretable features based on their shared characteristics. Experimental evaluations conducted on composite datasets, such as product purchase records, demonstrate the effectiveness of the proposed method in uncovering new personas and underscore its potential for informing data-driven marketing strategies. The findings further indicate that an integrated data analysis framework not only enhances the depth of customer understanding but also enables the derivation of concrete, actionable insights, ultimately contributing to improved business performance.