Debiasing Competition Solution: Multi‑hop i2i Graph Modeling for Advertising Recommendation

The ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) hosts an annual competition that attracts both academia and industry. In KDD Cup 2020, the Debiasing track focused on eliminating selection and popularity bias in next‑item prediction for e‑commerce advertising.

Data analysis revealed two major biases: (1) Selection Bias – the exposure data used for training is a subset selected by the system, and (2) Popularity Bias – popular items receive disproportionate clicks, leading to a “Matthew effect”. The dataset contains over 1 M clicks, 100 k items, and 30 k users, with rich multimodal item features (text and image vectors).

To address these challenges, the team built a heterogeneous item‑to‑item (i2i) graph with two edge types: click‑co‑occurrence edges (weighted by time interval, user activity, and popularity penalties) and multimodal similarity edges (cosine similarity of text/image vectors). Multi‑hop random walks on this graph generated unbiased candidate samples, expanding the training set and mitigating selection bias.

The modeling pipeline consists of three stages:

Construction of the i2i graph and multi‑hop walk to produce candidate samples.

Creation of i2i training pairs, automatic high‑order feature engineering, and training with LightGBM using a popularity‑weighted loss:

L = -\sum_{i}\left[\alpha\cdot y_i\log(p_i) + (1-y_i)\log(1-p_i)\right]

, where \(\alpha\) is inversely proportional to item popularity.

Aggregation of scores from multiple source items via max‑pooling and post‑processing that boosts low‑popularity items, improving NDCG@50_half.

The solution achieved 1st place out of 1 895 teams, with NDCG@50_half 6.0% higher than the runner‑up. The same techniques were applied to Meituan’s search advertising system, where bias mitigation led to significant business improvements across matching, ranking, and creative‑selection stages.

In summary, converting traditional user‑to‑item (u2i) CTR modeling to user‑to‑item‑to‑item (u2i2i) i2i modeling, enriching the candidate set via multi‑hop graph walks, and incorporating popularity penalties in both graph construction and loss functions effectively resolve both selection and popularity bias in large‑scale recommendation systems.