Technical Summary: Representation Curriculum for Robust Ranking and Allocation

1. Problem Statement

Modern digital marketplace ranking systems operate as dynamic exposure-allocation mechanisms. These systems rely heavily on endogenous, exposure-confounded signals (e.g., popularity estimates, historical CTR/CVR, and ID-based representations) because such signals are highly predictive under approximately stationary demand and explain significant variance in logged outcomes.

However, this predictive power creates a learning pathology:

• Shortcut Learning: Early access to exposure-dependent belief signals during training steers optimization dynamics toward over-reliance on these signals.
• Gradient Starvation: In pairwise or listwise learning-to-rank (LTR), strong predictors (historical beliefs) can create large score gaps early in training, causing the loss function to saturate. This suppresses gradients for weaker but essential cues, specifically exposure-independent merit signals (e.g., content-based competitiveness, semantic intent affinity).
• Consequences: The learned policy tends to entrench incumbents, degrade cold-start generalization, and reduce robustness under distribution shifts. This leads to self-reinforcing exposure loops that harm discovery, fairness, and long-term platform health.

Existing mitigation strategies (counterfactual learning, supervision shaping, shrinkage, architectural constraints, and stochastic masking) are often semantics-agnostic or fail to control how learning capacity is allocated across feature groups over time.

2. Methodology: Representation Curriculum (RC)

The authors propose Representation Curriculum (RC), a semantics-aware training-time intervention that temporally stages feature utilization to prevent shortcut reliance on historical signals.

Core Mechanism

RC partitions features into two semantic groups:

1. Content-based merit signals ($x_C$): Available regardless of prior exposure (e.g., item attributes, query–item affinity).
2. Exposure-dependent historical belief signals ($x_H$): Derived from past interactions (e.g., stabilized CTR, historical transactions).

RC implements a two-stage training curriculum:

• Stage 1 (Content-Only): The model is trained exclusively on content-based signals ($x_C$), with historical signals masked ($x_H = 0$). This forces the learner to develop a robust "content pathway" capable of explaining label variation without relying on historical shortcuts.
• Stage 2 (Full + Anchoring): Historical signals ($x_H$) are introduced. However, the optimization is anchored to the Stage 1 solution to preserve the learned content representations. This is achieved via two complementary anchoring strategies:
  1. Prediction Anchoring (Masked Consistency): Enforces that the Stage 2 model's predictions on masked inputs (where $x_H=0$) remain consistent with the Stage 1 model.
  2. Parameter Anchoring (Content-Path Protection): Explicitly regularizes the content pathway parameters ($\theta_C$) to remain close to their Stage 1 values ($\theta_C^{(1)}$).

The Stage 2 objective function combines the primary task loss, standard regularization, and the anchoring terms:
$$\mathcal{L}_{total} = \mathbb{E}[\mathcal{L}(f_\theta(x_C, x_H); y)] + \lambda \Omega(\theta) + \mu_1 \mathcal{L}_{pred}(\theta) + \mu_2 \mathcal{L}_{par}(\theta)$$

Model Agnosticism

RC is applicable across different hypothesis classes:

• Neural Networks: Implemented in two-tower architectures where the item-content tower is anchored.
• Gradient Boosted Decision Trees (GBDTs): Implemented by fitting an initial set of trees ($M$) using only content features, then continuing to boost the remaining trees ($T-M$) using all features, treating the Stage 1 output as a fixed base margin.

3. Theoretical Guarantees

The authors formalize RC in a Gaussian linear ridge regression setting to derive closed-form solutions and verifiable sufficient conditions.

• Pareto Tradeoff: The analysis demonstrates that RC allows for a quantified tradeoff between performance on the logged source distribution ($P$) and a strategic target distribution ($Q$) that emphasizes cold-start or under-exposed segments.
• Risk Bounds:
  • Source Performance: RC maintains bounded regret relative to a standard full-model (trained on all features) on the logged distribution $P$.
  • Target Robustness: RC significantly reduces population risk on the target distribution $Q$ (specifically in "frozen-start" scenarios where $x_H$ is absent or unreliable).
• Conditions for Improvement: The paper provides sufficient conditions under which RC strictly reduces risk on the target distribution while remaining competitive on the source. These conditions depend on the conditioning of the Schur complement of the covariance matrix and the margin between the full model and the content-only model.

4. Experimental Results

The efficacy of RC is validated through public benchmarks and a large-scale online deployment:

• MSLR-WEB (Learning-to-Rank): RC successfully reduced the feature importance of highly predictive historical belief signals. This shift improved the ranking quality for cold items with minimal degradation in overall performance.
• MovieLens (Recommendation): In a neural two-tower setting, RC shaped the model toward stronger reliance on content features. This resulted in significantly better "frozen-start" performance (simulating scenarios where historical signals are unavailable).
• eBay Sponsored Search (Online A/B Test): A randomized experiment in a live e-commerce system showed that a policy trained via RC:
  • Increased exposure and sale velocity for new listings (cold-start items).
  • Maintained neutral aggregate Key Performance Indicators (KPIs), indicating that the shift toward content-based merit did not harm overall marketplace efficiency.

5. Significance and Claims

The paper claims that Representation Curriculum offers a distinct advantage over existing methods by explicitly controlling the temporal trajectory of learning rather than just the final objective or data distribution.

• Mechanism Design: RC acts as a mechanism design tool that aligns the learning process with welfare-aligned goals (e.g., supporting new suppliers) without sacrificing the predictive power of historical data.
• Robustness: By preventing the "shortcut" of over-relying on endogenous signals, RC produces policies that are more robust to distribution shifts and better at generalizing to strategic subpopulations (cold-start).
• Practicality: The method is model-agnostic and integrates seamlessly into existing training pipelines (neural and tree-based) with minimal architectural changes, requiring only a staged training schedule and anchoring constraints.

The authors conclude that RC effectively shifts reliance from historical belief signals toward content-based merit signals, yielding consistent gains on cold populations with a controlled trade-off in head performance.