Automated Graph Representation Learning for KDD Cup 2020 AutoGraph: Technical Solution and Advertising Applications

ACM SIGKDD (International Conference on Knowledge Discovery and Data Mining) is the top conference in the data mining field. The KDD Cup 2020 featured five tracks, among which the AutoGraph track focuses on automated graph representation learning.

Figure 1: KDD 2020 conference.

The Meituan search advertising algorithm team, together with a member from the University of Chinese Academy of Sciences, formed team Aister and won the championship in the Debiasing, AutoGraph, and Multimodalities Recall tracks.

This article introduces the technical solution for the AutoGraph track and the application of graph representation learning in Meituan’s advertising system.

Background

KDD Cup 2020 set up four tracks (Debiasing, Multimodalities Recall, AutoGraph, and others). Graph Neural Networks (GNN) have become widely used in advertising, social networks, knowledge graphs, and even life sciences. The advertising system contains rich structured relationships (User‑Ad, Query‑Ad, Ad‑Ad, Query‑Query). The team applied graph representation learning to improve ad performance and won the AutoGraph track.

Problem Analysis

Five offline graph datasets were provided. They vary in directionality, density, and presence of node features. Most datasets have a strict time budget of about 100 seconds, making exhaustive neural architecture search infeasible.

Table 1: Overview of the five offline graph datasets.

Dataset 5 showed training instability, likely due to over‑fitting on a relatively easy graph. Ranking stability across datasets is more important than raw accuracy, as small accuracy differences can cause large rank changes.

Figure 3: Model training instability on Dataset 5.

Figure 4: Accuracy and rank differences across teams.

Challenges

Graph data diversity – models must work on directed/undirected, dense/sparse, feature‑rich/feature‑less graphs.

Very short time budget – each dataset allows only ~100 s for model training and inference.

Robustness – final submissions are evaluated on unseen datasets, so ranking stability is crucial.

Technical Solution

We designed an automated graph learning framework (Figure 5). The pipeline includes graph preprocessing, feature construction, multiple GNN models, fast hyper‑parameter search, and a multi‑level robust model‑fusion strategy.

Figure 5: Automated graph learning framework.

Data Pre‑processing & Feature Construction

Directed edges are reversed to make the graph undirected, preserving edge weights. Hand‑crafted node features (degree, 1‑hop and 2‑hop neighbor statistics) are bucketed and embedded to reduce over‑fitting.

Graph Neural Network Models

We employed four GNN architectures to handle different graph types:

GCN – spectral method, simple and effective on most datasets.

TAGConv – polynomial‑based spectral method, fast on sparse graphs.

GraphSAGE – sampling‑based spatial method, efficient on dense graphs.

GAT – attention‑based spatial method, excels on feature‑rich graphs.

Formulas for each model are shown in the original paper (omitted here for brevity).

GCN equation.

TAGConv equation.

GraphSAGE equation.

GAT equation.

Fast Hyper‑parameter Search

We split the search into offline (large search space to define bounds) and online (small sub‑space for rapid tuning). Offline we determined ranges for learning rate, number of layers, hidden size, etc., for each graph type. Online we performed a few‑epoch training (e.g., 16 epochs) to pick the best learning rate (0.003) and other hyper‑parameters.

Figure 6: Fast hyper‑parameter search pipeline.

Figure 7: Validation accuracy vs. learning rate after few epochs.

Multi‑level Robust Model Fusion

Because final ranking is based on average rank over unseen datasets, we adopted a hierarchical fusion strategy:

Data‑level split → multiple groups of models trained with early stopping.

Within each group, n‑fold bagging of each GNN model.

Density‑adaptive weighted fusion: model weight = f(density, accuracy) where density = #edges / #nodes. For dense graphs, weight depends mainly on accuracy; for sparse graphs, density influences the weight to improve rank stability.

Final group‑level averaging.

Figure 8: Multi‑level robust fusion architecture.

Evaluation Results

Table 2 shows test accuracy of each GNN on the five offline datasets. GCN performed well across all datasets, TAGConv excelled on sparse graphs, GraphSAGE on dense graphs, and GAT on feature‑rich graphs. Model fusion consistently achieved the best and most stable results.

Table 2: Test accuracy of individual GNNs and fused model.

Table 3 reports the final ranking on the hidden five datasets. Our solution achieved first place overall, confirming the effectiveness of the automated framework.

Table 3: Final ranking of top‑5 teams.

Advertising Business Applications

The advertising system contains heterogeneous nodes (User, Query, Ad, Location) and rich edge types (Query‑Query session, Query‑Query similarity, Query‑Ad click, Ad‑Ad co‑click). We constructed a heterogeneous graph (Figure 9) and applied the learned node embeddings to two core modules:

Trigger module – vector‑based ad recall using ANN, improving recall rate and long‑tail traffic.

Click‑through‑rate (CTR) estimation – augmenting the DSIN model with GNN‑based interest expansion, leading to higher prediction accuracy.

Figure 9: Heterogeneous graph construction.

Figure 10: Trigger network based on graph representation learning.

Figure 11: Personalized CTR prediction network with GNN attention.

Conclusion & Outlook

KDD Cup bridges academia and industry; past winning solutions (e.g., FFM, XGBoost) have become widely adopted. This year’s focus on automated graph representation learning and recommendation systems highlights the synergy between graph learning and AutoML.

Our solution addressed three main challenges—graph diversity, ultra‑short time budget, and robustness—through a modular automated framework, fast hyper‑parameter search, and multi‑level model fusion. The approach not only won the competition but also brought tangible improvements to Meituan’s search advertising trigger and CTR estimation modules.

Future work will further optimize graph models, explore AutoML‑driven end‑to‑end pipelines, and extend graph‑based techniques to other advertising scenarios such as user intent modeling.