A Probabilistic Multi-Topic Model Learned From Aesthetic Quality Communities for Dress Colors Recommendation

We have developed a Probabilistic Multi-Topic Model that segments large-scale online clients, such as those on e-commerce platforms or social media, into distinct aesthetic quality communities based on their shared preferences for dress colors. This method leverages a geometry-based feature selection strategy to effectively capture and model each client’s aesthetic tendencies, characterized by a diverse set of features, including those learned under weak supervision. These features are further refined through a feature selection process. The model represents each client’s color preferences as a probabilistic multi-topic distribution within a hidden feature space, offering a comprehensive understanding of their aesthetic inclinations. To uncover the underlying connections between clients with similar preferences, we construct a graph-based framework and employ a specialized algorithm to identify tightly connected subgroups, or aesthetic quality communities. These communities are defined by shared dress color preferences, such as an affinity for bold, vibrant hues or subtle, neutral tones. To enhance personalization, the model integrates a ranking mechanism that recommends dress colors and styles aligned with the identified communities. By learning from these aesthetic quality communities, the system ensures that recommendations are both relevant and deeply resonant with the client’s unique aesthetic profile. We validated the effectiveness of this approach using a dataset comprising over one million clients. Results demonstrate that our model accurately identifies and classifies distinct aesthetic quality communities, enabling internet-scale personalized dress color recommendations that are aesthetically aligned and client-centric.