Multi-Objective Optimization with MMoE for Taobao "Lying Flat" Channel

Taobao's "Lying Flat" channel is a scenario‑based recommendation platform for home‑goods, where users can enter via three methods and interact with content cards that link to product detail pages, aiming to increase GMV through multi‑step user actions.

The business goal requires optimizing not only first‑click (CTR) but also second‑click and downstream actions such as add‑to‑cart and purchase, revealing a gap in the existing ranking model that only considers first‑click.

Existing multi‑objective approaches include sample‑weighting, model stacking, shared‑bottom networks, and MMoE. Sample‑weighting lacks true multi‑task modeling; stacking increases model size; shared‑bottom works well only when tasks are highly correlated.

MMoE improves upon shared‑bottom by introducing multiple experts shared across tasks and a gating network that learns task‑specific expert weights, allowing effective learning even when task correlations are low.

We adopted MMoE for the "Lying Flat" channel and applied four layers of optimization: (1) adding a linear model output, (2) feeding linear features into task towers, (3) incorporating linear features into expert and gate networks, (4) combining (1) with (3), and (5) combining (1) with (2). The best performance came from the fourth variant.

Feature processing was enhanced by introducing local activation units inspired by Deep Interest Network to capture user‑content relevance, and linear features were integrated to correct model generalization.

Loss weighting experiments showed that a 1:1 weight between first‑click and second‑click losses yielded the best offline AUC improvements (CTR +0.24%, CVR +6.01%).

We extended the model to a three‑objective setup (first‑click, second‑click, and post‑click purchase actions) and performed model slimming, reducing depth from five to three layers while maintaining performance.

Online A/B testing over seven days demonstrated significant gains in second‑click rate, with a slight trade‑off in first‑click rate, confirming the effectiveness of MMoE‑based multi‑objective optimization for this scenario.

Future work includes incorporating GMV‑related features and exploring advanced loss‑weighting methods to further maximize revenue.

References: [1] Covington et al., 2016; [2] arXiv:1706.05098; [3] Caruana, 1993; [4] Luming Dong blog; [5] Ma et al., 2018; [6] Zhou et al., 2018; [7] Kendall et al., 2018.