Toss TASTE Improves Recommendation Performance by 35.1% via Multitask Learning and Demonstrates Transfer Effects to Other Services

• Toss TASTE extracted user representations through multitask learning to improve recommendation performance by approximately 35.1%
• Applying TASTE embeddings to models for other services increased mini-app recommendation CTR by up to 209%
• Introducing Auto Targeting improved push campaign CTCVR from 0.25% to 2.70%
• Built TASTE Bench to establish a process for quickly verifying the effects of embedding improvements

Technical Architecture and Learning Method

• Utilized MMoE, DCN-V2, and PLE structures to jointly learn app events and consumption data
• Compressed recent app event sequences using a Causal Transformer to complement the limitations of summary information
• Separated app event and consumption prediction models to apply an optimization structure tailored to specific objectives

Operations and Transfer Effects

• Introduced the Auto Targeting feature in IM (Intelligence Manager) to reduce targeting burden
• Achieved a 40% increase in conversions despite a 35% reduction in send volume when sending LMS for financial product categories
• Added TASTE embeddings to shopping and advertising models, confirming improvements in AUROC and NE metrics