Meituan Technical Team's Three Papers Accepted at SIGIR 2024: Ad Auction Integration, Federated Recommendation, and POI Recommendation

01 Deep Automated Mechanism Design for Integrating Ad Auction and Allocation in Feed

Electronic commerce platforms display an ordered list that mixes organic results and ads; the list results from ad auction and ranking, directly affecting ad revenue and overall GMV. Conventional pipelines treat auction and ranking as separate stages, leading to sub‑optimal outcomes because (1) the auction ignores externalities such as position‑dependent CTR, and (2) the ranking uses the winning ad’s payment to decide placement, breaking incentive compatibility (IC).

The paper proposes a deep automated mechanism that jointly optimizes auction and ranking while guaranteeing incentive compatibility and individual rationality (IR) under externalities. Candidate ads are inserted at every possible position in the organic list to generate all candidate allocations. A page‑level model predicts outcomes for each allocation, capturing global externalities. A deep neural network‑based auction then selects the optimal allocation and determines pricing. Offline experiments and online A/B tests show higher ad revenue and GMV than state‑of‑the‑art baselines.

02 ReFer: Retrieval‑Enhanced Vertical Federated Recommendation for Full‑Set User Benefit

Vertical federated learning (VFL) is increasingly used in recommendation systems, but traditional federated schemes ignore non‑overlapping user data, limiting the richness of user interest signals and restricting predictions to a small set of intersecting users. To address this, the authors define a Fully Vertical Federated Recommendation (FullyVFR) paradigm and introduce ReFer, the first retrieval‑enhanced vertical federated recommendation framework.

ReFer employs a two‑stage distributed retrieval architecture together with a distributed attention‑fusion mechanism to mitigate cross‑domain feature missingness and reduce interest bias across user groups. Experiments on public benchmarks and Meituan’s internal datasets demonstrate significant performance gains for the entire user population across multiple recommendation tasks.

03 Disentangled Contrastive Hypergraph Learning for Next POI Recommendation

Next‑point‑of‑interest (POI) recommendation aims to suggest the user’s next location. Existing sequence‑based and graph‑neural approaches often overlook (1) the influence of multiple, evolving decision factors on user preference, leading to coupled and sub‑optimal user representations, and (2) the collaborative relationships among these factors, limiting the ability to capture complementary recommendation signals.

The proposed Disentangled Contrastive Hypergraph Learning (DCHL) framework introduces a multi‑view disentangled hypergraph component that separately models user‑POI interactions from collaborative, transition, and geographic perspectives. View‑specific hypergraph convolution networks learn disentangled POI embeddings, and an adaptive fusion module automatically merges the multi‑view user representations. Cross‑view contrastive learning captures inter‑view correlations, enhancing both user and POI representations. Experiments on three real‑world datasets show that DCHL outperforms a range of strong baselines.

On July 11 (Thursday) from 14:00 to 16:00, the three paper authors will present these works in an online live session. Interested participants can register via the provided link.