Alibaba Mama Team Papers Accepted at KDD 2022 and Other Top Conferences

Recently, the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2022) announced its paper acceptance results. KDD 2022, an A‑class international conference recommended by the China Computer Federation (CCF), received 1,695 submissions to the Research track, of which 254 were accepted (acceptance rate 14.9%).

The Alibaba Mama technical team had five papers accepted at this top data‑mining conference. In the following sections we present the titles and abstracts of these papers.

ROI‑Constrained Bidding via Curriculum‑Guided Bayesian Reinforcement Learning

Abstract: Real‑Time Bidding (RTB) is a key form of online advertising. Advertisers aim to maximize ad delivery while satisfying return‑on‑investment (ROI) constraints. Existing works often assume a static or smoothly changing market, which fails under highly non‑stationary conditions. We formalize the problem as a Partially Observable Constrained MDP and handle constraints via a latent function without extra hyper‑parameters. We propose a Curriculum‑Guided Bayesian Reinforcement Learning (CBRL) framework that uses a series of approximated problems to guide learning under sparse rewards, enabling the bidding policy to infer market dynamics and adapt the trade‑off between constraints and objectives. Experiments on two large‑scale datasets demonstrate CBRL’s superior performance, stable convergence, and strong out‑of‑distribution generalization.

Adversarial Gradient Driven Exploration for Deep Click‑Through Rate Prediction

Abstract: Exploration‑exploitation methods address data‑loop issues in large‑scale online recommendation systems. Prior work focuses on uncertainty estimation, overlooking the impact of explored samples on subsequent model learning. We design a Pseudo‑Exploration module that simulates the effect of successfully explored samples on the recommender’s future training, implemented by adding adversarial perturbations to model inputs. This leads to the Adversarial Gradient driven Exploration (AGE) strategy. To improve efficiency, we introduce a dynamic gating unit that filters low‑value samples. Extensive experiments on public academic datasets validate AGE’s effectiveness, and it has already been deployed on Alibaba’s display advertising platform with positive online gains.

EXTR: Click‑Through Rate Prediction with Externalities in E‑Commerce Sponsored Search

Abstract: Click‑through rate (CTR) prediction is crucial for e‑commerce advertising. Items surrounding a target ad—both organic results and other ads—exert externality effects on the target’s CTR. Existing CTR models ignore these externalities. We propose the Externality Transformer (EXTR), which treats all possible display positions of the target ad as queries and surrounding items as keys/values, modeling diverse externalities. A Position‑Aware Gating (PAG) module learns latent ad arrangements to address incomplete externalities. Experiments on Alibaba’s real platform show that EXTR effectively captures externality effects and improves platform revenue.

Pretraining Representations of Multi‑modal Multi‑query E‑commerce Search

Abstract: Modeling contextual information across multi‑modal, multi‑query search sessions is essential for e‑commerce. Users may issue text queries, upload photos, or use visual search within a single session. Prior work only models textual queries. We propose a heterogeneous graph neural network that learns representations for multi‑modal, multi‑query (MM) sessions, using a multi‑view contrastive learning framework to pre‑train the graph network. Two views capture intra‑query, inter‑query, and cross‑modal information propagation. Experiments demonstrate that the pre‑trained session representations boost state‑of‑the‑art baselines on downstream tasks such as personalized CTR prediction, query recommendation, and intent classification.

Streaming Graph Neural Networks with Generative Replay

Abstract: Graph neural networks (GNNs) have achieved great success on static graph tasks, but real‑world graphs often evolve as streams. Existing methods either retrain on the whole graph at each time slice (high computational cost) or suffer from catastrophic forgetting in online learning. We introduce a novel framework that employs an auxiliary graph generative model to preserve the full historical distribution, enabling the GNN to learn new patterns while consolidating existing knowledge. Experiments on multiple real‑world networks show that our approach maintains efficiency while delivering superior performance.

Further detailed analyses of these papers will be presented by the authors in upcoming sessions. Stay tuned for more insights.