PEETS: Preference-aware Embeddings with Explicit Tree Structure for Auxiliary Product Bundling

Maolei Huang, Jing Tian, Dexuan Duan, Yan Chen, Huawei Cao, Shuhan SongarXiv 2026 · arXiv

TL;DR

PEETS is a novel training-time framework that improves auxiliary product recommendations (like flight insurance) by using a hierarchical tree structure to regularize embeddings. By leveraging Hierarchical and Negative Contrastive Losses, it achieves significant gains in both offline AUC and online revenue (GMV) without adding any computational overhead during real-time serving.

What Problem Does This Paper Address?

When you book a flight, you're often prompted to add "auxiliary products" like flight protection, flexible cancellation, or airport transfers. For recommendation systems, this presents a unique challenge called granularity ambiguity.

Within a single product category (e.g., "Flight Insurance"), there are multiple Stock Keeping Units (SKUs) that share similar semantics but differ in price and coverage. Standard recommendation models often struggle with two things:

  1. Semantic Overlap: SKUs in the same category are highly related, but they aren't interchangeable.
  2. Weak Negative Signals: If a user rejects a basic insurance plan, it doesn't necessarily mean they aren't interested in insurance at all—they might just want the premium version.

Existing models often fail to distinguish between "disinterest in the category" and "disinterest in a specific offer," leading to suboptimal bundling.

Methods

The authors introduce PEETS (Preference-aware Embeddings with Explicit Tree Structure). It is a training-time regularization framework that acts on shared SKU representations. The core innovation lies in two auxiliary losses:

  • Hierarchical Contrastive Loss (HCL): This uses a pre-defined catalog tree (based on category and price quantiles). It compares SKU pairs by conditioning their similarity on their deepest common ancestor. This "re-centers" the embeddings, forcing the model to learn what makes a specific SKU unique relative to its siblings in the hierarchy.
  • Negative Contrastive Loss (NCL): This leverages explicit rejection events (e.g., a user removing an item from their cart). When a rejection occurs, PEETS shifts the embedding comparison based on the statistics of the "rejected bucket," helping the model distinguish between passive non-clicks and active rejections.

Crucially, these components are "training-only." The auxiliary branches are discarded after training, leaving the predictive serving path (the MLP heads) exactly the same. This allows for better embeddings without increasing inference latency.

Datasets

The study uses large-scale industrial data from Fliggy (Alibaba Group):

  • Scale: 4.2 million users and 960 SKUs across 6 product types.
  • Training Set: 11 million records collected over four months in 2025.
  • Rejection Events: 0.48 million explicitly logged rejection actions.
  • Baselines: Compared against industry standards like Wide & Deep, DIN, ESMM, and the state-of-the-art TWIN V2.

Key Insights and Ideas

  • The "Free Lunch" of Training-Time Regularization: By moving the complexity of the tree structure into the loss function rather than the model architecture, the authors achieve better performance with zero additional cost at inference time (only 0.1ms difference in p95 latency).
  • Explicit Rejection as a Strong Signal: Most models treat "not clicked" as a generic negative. PEETS recognizes that an explicit rejection is a much noisier but more informative signal that requires specific "shifting" in the embedding space to be useful.
  • Hierarchy as Inductive Bias: Instead of letting the model figure out product relationships from scratch, PEETS uses the existing business catalog (the tree) to provide a structural inductive bias, ensuring that price-sensitive and category-sensitive features are properly disentangled.

Key Results

  • Offline Performance: PEETS achieved 78.20% AUC, outperforming the strong TWIN V2 baseline by 0.90 percentage points and the base model by 1.15 points.
  • Online Impact: In a 14-day A/B test, PEETS improved Gross Merchandise Value (GMV) per exposed user-day by 8.20% compared to the production policy and 0.84% over TWIN V2.
  • Efficiency: Training time increased by only ~8% (about 5.5 minutes), while server-side p95 latency remained stable at 42.1 ms.

Limitations and Future Work

  • Cancellation Rates: While revenue and clicks went up, the "cancellation interval's upper limit" exceeded the pre-specified non-inferiority margin. This suggests that while the model is better at selling add-ons, it might occasionally push products that users later regret, leading to higher cancellation rates.
  • Static Hierarchy: The tree structure is currently based on fixed catalog attributes (price quantiles and types). The authors suggest that future work could involve learning the hierarchy dynamically or incorporating more complex graph structures.
  • Joint Optimization: The current scope is limited to ranking add-ons for a selected itinerary. A more complex "joint optimization" of the flight and the add-on together remains an open challenge.

Summary updated

Abstract

Auxiliary products such as flight protection and flexible-change services are recommended alongside primary travel purchases. This setting creates a granularity ambiguity: SKUs within one product type share semantics but differ in price and coverage, while rejecting one offer does not imply disinterest in the entire type. We present \textbf{PEETS} (\textbf{P}reference-aware \textbf{E}mbeddings with \textbf{E}xplicit \textbf{T}ree \textbf{S}tructure), a training-time framework for auxiliary product bundling. PEETS regularizes an existing click--conversion predictor through two conditional comparisons. Hierarchical Contrastive Loss uses the pair's common ancestor as a reference; Negative Contrastive Loss conditions structural sampling and comparison shifts on logged rejection events. Both act on shared SKU representations during training. This separates catalog relatedness from individual preference supervision and leaves the predictive serving path intact. PEETS achieves 78.20\% AUC, exceeding TWIN V2 by 0.90 percentage points and the matched base model by 1.15 points. In a 14-day online evaluation, gross auxiliary GMV per exposed user-day improves by 8.20\% over Policies, 3.06\% over Base, and 0.84\% over TWIN V2, with all three user-bootstrap 95\% intervals above zero. Bundling rate is 0.51 points below Policies. The cancellation interval's upper limit exceeds the prespecified margin. PEETS adds no auxiliary-loss computation to serving and records 42.1~ms p95 latency, compared with 42.0~ms for Base.
BibTeX
@misc{huang2026peets,
  author = {Maolei Huang and Jing Tian and Dexuan Duan and Yan Chen and Huawei Cao and Shuhan Song},
  title = {{PEETS: Preference-aware Embeddings with Explicit Tree Structure for Auxiliary Product Bundling}},
  year = {2026},
  eprint = {2611.20726},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.20726},
}
Figure