Fashion Apparel Product Recommendation Engine With Automated Tagging and Consumer Preference Data
The Challenge
A large fashion retailer in the UK wanted to build an online apparel fitting application as the manual effort in tagging dresses was fraught with inconsistency, errors and not efficient for online shoppers. The current process resulted in incorrect sizing and a high risk of product returns. They also wanted to leverage multiple sources of data with implicit signals of shopper preferences to optimize product recommendations.
Hypothesis
- Vision based algorithms can extract all useful tags for women’s fashion.
- Ensemble of recommender systems using different buyer signals can be built without introducing scale bias.
- Selection bias in items viewed and bought can be algorithmically addressed.
Execution
- Multiple base recommenders built using traditional approaches as well as novel item and customer models.
- Novel approaches to scaling base recommender scores to remove implicit scale bias.
- Items were modeled with respect to their sales patterns.
- Customers were modeled based on the types of products and their eclectic test.
- Local rather than global weights derived for created ensemble of recommenders for each user.
Outcomes
- 60% reduction in cost of tagging dresses and other woman wear.
- 30% increase click through rate.
- Tags had confidence metrics associated with them to enable human intervention where needed.
- 30% increase in click through rate on recommendation.
Project Highlights
12%
increase sales revenue