Representation Curriculum: Stagewise Training for Robust Ranking and Allocation

Researchers propose Representation Curriculum (RC), a machine learning training method that improves ranking systems in digital marketplaces by strategically controlling when different data signals are introduced during model training. The approach reduces over-reliance on exposure-dependent historical signals and strengthens content-based merit evaluation, yielding better performance on cold-start scenarios and improved robustness across distribution shifts.

This research addresses a fundamental problem in modern recommendation and ranking systems: the tendency of machine learning models to exploit shortcuts in training data at the expense of genuine signal quality. Digital marketplaces rely heavily on historical engagement metrics like click-through rates and popularity scores because these signals are highly predictive under normal conditions. However, this predictive strength creates a training trap where models learn to depend almost exclusively on exposure-dependent signals, essentially memorizing past allocation patterns rather than learning true content merit.

The Representation Curriculum method intervenes during training by temporally sequencing feature introduction. By initially training on content-based features while gradually introducing historical signals, RC anchors the model's learned representations in genuine merit before allowing it to incorporate historical bias. This curriculum approach parallels human learning, where foundational concepts must be established before building complex knowledge.

The implications extend across e-commerce search, recommendation systems, and marketplace ranking broadly. Current systems often entrench successful sellers while disadvantaging new entrants with quality offerings, creating market friction and reduced discovery diversity. By improving cold-start performance and robustness during market shifts, RC enables fairer allocation mechanisms that benefit both consumers seeking better products and newer sellers seeking market access. The theoretical analysis in Gaussian linear settings provides formal guarantees, while practical validation on public benchmarks and large-scale e-commerce experiments demonstrates real-world applicability.

The controlled trade-off against head performance suggests implementations must balance discovery fairness against immediate revenue optimization—a key consideration for platform operators. As ranking systems become increasingly sophisticated, curriculum-based training approaches may become standard practice for building more robust, generalizable allocation mechanisms.

• →Representation Curriculum addresses the shortcut-learning problem where ranking models over-rely on historical engagement signals rather than content merit
• →The method temporally stages feature introduction during training, prioritizing content-based signals before gradually introducing exposure-dependent historical data
• →RC demonstrates measurable improvements on cold-start populations and maintains robustness under distribution shifts with quantified trade-offs in head performance
• →Theoretical analysis provides closed-form solutions and sufficient conditions for population risk reduction in Gaussian linear settings
• →Online experiments in large-scale e-commerce validate the approach, showing practical shifts in model reliance from historical bias toward content-based merit