Style-Quizzes for Content-Based Fashion Recommendation in Extreme Cold Start Scenarios
Keywords: Recommender Systems, Cold Start Problem, Fashion
TL;DR: The article discusses a novel method for avoiding the cold start problem by presenting the user with a style quiz during the onboarding phase.
Abstract: This article presents Style-Quiz, a novel method for circumventing the user-based cold start problem in the context of content-based recommender systems. We construct a content-based recommender system for a given environment and generate a quiz built upon its underlying embeddings. During the course of the quiz, the embedding space of the recommender system is segmented via unsupervised hierarchical clustering. The user is presented with a series of images representative of each cluster and tasked with choosing one of them. The chosen cluster is then segmented in the same way as its parent cluster. This process is repeated until the user has honed in on a point in the embedding space that adequately represents that user's tastes.
As a user interested in renting or purchasing fashion items is likely to be interested in several different kinds of fashion articles, we also introduce Style-Vectors. A representation of our items, built on deep-learning encoders and triplet loss, that is indicative of their underlying style, not just physical attributes.
Our results indicate that Style-Quiz significantly improves early personalized recommendation as compared to recommending globally popular items.
To improve reproducibility, we publish both the code and dataset used for the project.
Submission Number: 26
Loading