Browse Prior Art Database

Method and System for Providing Latent Visual Preference Learning (LVPL) for Deep Personalized Image Search in Online Shopping

IP.com Disclosure Number: IPCOM000242659D
Publication Date: 2015-Aug-03
Document File: 2 page(s) / 20K

Publishing Venue

The IP.com Prior Art Database

Related People

JH Hsiao: INVENTOR

Abstract

A method and system is disclosed for providing latent visual preference learning (LVPL) for deep personalized image search in online shopping. The method and system delivers a more accurate preference-level image search satisfying user's personal visual taste and greatly boosts the search quality.

This text was extracted from a Microsoft Word document.
This is the abbreviated version, containing approximately 52% of the total text.

Method and System for Providing Latent Visual Preference Learning (LVPL) for Deep Personalized Image Search in Online Shopping

Abstract

A method and system is disclosed for providing latent visual preference learning (LVPL) for deep personalized image search in online shopping. The method and system delivers a more accurate preference-level image search satisfying user's personal visual taste and greatly boosts the search quality.

Description

Image search is steadily becoming more important and it enables the possibility of many use cases with great business potential. Take e-commerce as an example, visual search technique can enhance the product recommendation since traditional collaborative filtering methods consider only the co-occurrence relationship among products yet ignore the content relevancy (i.e., visual similarity of product appearance, which is extremely important in some product categories, such as fashion-related products) for product suggestion.  Despite its importance, there still remain limitations and weakness for current image search engine.  Current researches focus on finding good image feature to describe the image and use universal distance function for calculating the similarity among images.  However, the determination of an appropriate distance metric plays a very key role in building an effective visual search system.

Disclosed is a method and system for providing latent visual preference learning (LVPL) for deep personalized image search in online shopping.  The method and system implements a deep learning technique on a large image dataset of Electronic Commerce (EC) products with corresponding categories as labels to enhance the image feature discriminability.  LVPL then directly models users' latent visual preferences from click through data.  A set of past viewed or ignored products is used and analyzed to infer user's personal preference regarding different products.  A personalized distance metric that better describes user’s visual preference is eventually learned from the past browsing history and the resulted feature space separates visually preferred and non-preferred product by a large margin.  The method and system thus achieves a more accurate preference-level image search.  

In accordance with the method and system, we us...