Towards intelligent online cross-selling

Abstract

The ubiquity of online cross-selling for fashion demands a large number of qualified outfit compositions. This paper targets practical and intelligent online cross-selling by providing a more accurate fashion compatibility model and a reliable evaluation protocol for evaluating the fashion compatibility model. Firstly, a Hierarchical Outfit Network (HON) is proposed to leverage multi-layer relations among attributes, items, and outfits. The awareness of multiple relations hidden in various outfits enables the HON to outperform all state-of-the-art methods on fill-in-the-blank (FITB) accuracy and compatibility Area Under Curve (AUC) on the Maryland and Type-aware test sets. Meanwhile, a new evaluation protocol is introduced to assess the fashion compatibility model more objectively and accurately, namely, Aesthetic 100 (A100). A100 possess three desirable characteristics: 1) Completeness. All types of standards in the fashion aesthetic system are covered through two tests, namely LAT (Liberalism Aesthetic Test) and AAT (Academicism Aesthetic Test); 2. Reliability. It is an agnostic protocol and consistent with major indicators. It provides an objective and fair assessment for model comparison. 3. Explainability. A100 assesses the model on more fine-grained dimensions, e.g., Color, Material, and Balance demonstrating its superiority in identifying essential characteristics of fashion aesthetics. Experimental results demonstrate the progress of A100 in the aspects of Reliability and Explainability. The evaluation results on A100 also show the generalization ability of HON from both quantitative and qualitative perspectives. Finally, solutions for multiple applications in fashion retailing are proposed to show how HON can be utilized to help online cross-selling.