---
title: "Making Ads Count: Using MMoE and Auxiliary Tasks to Better Connect Buyers & Sellers"
description: "Etsy’s Ads Search ranking model needed a stronger purchase-intent signal because clicks are noisy and purchases are sparse later in the buyer journey. It retained its shared feature-representation and interaction layers, then replaced the shared feed-forward network with Multigate Mixture of Experts (MMoE), using task-specific gates and towers for click-through rate (CTR) and post-click conversion rate (PCCVR). Add-to-cart was added as an auxiliary task, while favorites were rejected after producing negative transfer. Offline, the resulting model increased Purchase and Click PR AUCs by average 3.5% and 1%, respectively, and Purchase ROC AUC by 0.5%. Online, it improved purchases and purchase calibration, supplied better inputs to auto-bidding, kept model size flat through pruning, and reduced inference cost."
---

# Making Ads Count: Using MMoE and Auxiliary Tasks to Better Connect Buyers & Sellers

[Etsy](https://yomu.fyi/company/etsy) · Amanda Steigman · Mar 20, 2026

**Type:** Explainer

## Summary

Etsy’s Ads Search ranking model needed a stronger purchase-intent signal because clicks are noisy and purchases are sparse later in the buyer journey. It retained its shared feature-representation and interaction layers, then replaced the shared feed-forward network with Multigate Mixture of Experts (MMoE), using task-specific gates and towers for click-through rate (CTR) and post-click conversion rate (PCCVR). Add-to-cart was added as an auxiliary task, while favorites were rejected after producing negative transfer. Offline, the resulting model increased Purchase and Click PR AUCs by average 3.5% and 1%, respectively, and Purchase ROC AUC by 0.5%. Online, it improved purchases and purchase calibration, supplied better inputs to auto-bidding, kept model size flat through pruning, and reduced inference cost.

## Context

Clicks and purchases were the primary signals for Etsy’s Ads Search ranking model, but clicks could be noisy and purchase data became increasingly sparse later in the buyer journey. The team sought to improve meaningful buyer engagement and purchase-intent prediction without causing one task’s performance to degrade when another improved.

## Approach / What changed

The model introduced a Multigate Mixture of Experts layer between shared interaction layers and task-specific towers. Task-specific gates combine expert outputs for CTR and PCCVR, while add-to-cart serves as an auxiliary task. Favorites were evaluated but excluded from the production model after causing negative transfer.

## Takeaways

- MMoE lets CTR and PCCVR use different weighted subsets of experts, allowing task-specific patterns while retaining shared representations.
- Add-to-cart improved purchase metrics, while favorites were noisy and caused negative transfer; only add-to-cart was ramped up in production.
- The deployed model improved purchase calibration for auto-bidding, kept overall model size flat through pruning, and made serving inference less costly.

**Tags:** [Architecture](https://yomu.fyi/topic/architecture), [Experimentation](https://yomu.fyi/topic/experimentation), [Machine Learning](https://yomu.fyi/topic/machine-learning)

- Source: [Etsy](https://www.etsy.com/codeascraft/making-ads-count-using-mmoe-and-auxiliary-tasks-to-better-connect-buyers--sellers)
- Source URL: https://www.etsy.com/codeascraft/making-ads-count-using-mmoe-and-auxiliary-tasks-to-better-connect-buyers--sellers
- Ingested by Yomu: 2026-09-09T20:02:46.788Z

[Read original post](https://www.etsy.com/codeascraft/making-ads-count-using-mmoe-and-auxiliary-tasks-to-better-connect-buyers--sellers)