MGOE: Macro Task Graph Experts¶

Core mechanism evidence. Paper and local evaluation are separate.

Paper information¶

Field	Value
Paper	KDD 2026
Affiliations	Alibaba
First public release	2025-06-12 (arXiv v1)
Authors' code	Author code.
Adapter / method	`mgoe`
Local code	`src/auto_research/reproductions/mgoe/`

Original paper¶

Background and contribution¶

MMoE shares at example level but does not express stable inter-task structure. Build the Macro Task Merging Graph, propagate via graph bottom and experts, and predict with task towers.

flowchart LR
  T[Task correlations] --> G[Macro graph] --> E[Graph experts] --> Y[Task gates and towers]

Key figure from the paper¶

Original Figure 2: MGOE. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ H^{l+1}=\sigma(\tilde A H^lW_l),\qquad \hat y_t=\operatorname{Tower}_t\!\left(\sum_e g_{t,e}H_e\right). \]

Paper results¶

Public and Alibaba datasets outperform Shared-Bottom, MMoE, PLE, and MacGNN. Online versus MMoE: PCTR +2.16%, UCTR +1.63%, CVR +5.88%, GMV +16.46%, and stay time +4.12%.

Local reproduction¶

Build three behavior/genre proxy tasks on MovieLens. Against separate MMoE gates: NDCG@10 0.07514→0.08042 (+7.03%), Hit@10 0.14048→0.15952, with increased head share. Metrics.

auto-research reproduce --paper mgoe --dataset-dir data --seed 42

Reproduction limits¶

Actual graph estimation, propagation, and task gates, not billion-scale embeddings, distributed graph construction, or private online features.