<p>Long-tail recommenders often over-fit popular items because the training log is highly skewed: a small head catalogue dominates both ranking gradients and the negative samples used by contrastive learning. This paper studies that failure mode under full-ranking evaluation and proposes <b>Adaptive Popularity-Debiased Contrastive Learning (A-PDCL)</b>, a lightweight training-time correction for graph contrastive recommendation. The core idea is simple: in the item-side InfoNCE denominator, popular in-batch negatives are down-weighted and tail negatives are up-weighted using smoothed inverse item degree. Rather than requiring a separate manual debiasing sweep for every dataset, A-PDCL sets the exponent automatically as <InlineEquation ID="IEq1"><EquationSource Format="TEX">$\beta _{\textrm{eff}}=\beta _{\max }\cdot \textrm{Gini}_{\textrm{batch}}$</EquationSource><EquationSource Format="MATHML"><math><mrow><msub><mi>β</mi><mtext>eff</mtext></msub><mo>=</mo><msub><mi>β</mi><mo movablelimits="true">max</mo></msub><mo>·</mo><msub><mtext>Gini</mtext><mtext>batch</mtext></msub></mrow></math></EquationSource></InlineEquation>, so batches with stronger popularity concentration receive stronger correction. We evaluate A-PDCL on Yelp, Amazon Books, and MIND using a unified global-time-split, full-ranking protocol. Multi-seed results show that A-PDCL selects regime-appropriate correction (<InlineEquation ID="IEq2"><EquationSource Format="TEX">$\beta _{\textrm{eff}}\approx 0.75$</EquationSource><EquationSource Format="MATHML"><math><mrow><msub><mi>β</mi><mtext>eff</mtext></msub><mo>≈</mo><mn>0.75</mn></mrow></math></EquationSource></InlineEquation> on sparse Amazon Books and <InlineEquation ID="IEq3"><EquationSource Format="TEX">$\approx 0.59$</EquationSource><EquationSource Format="MATHML"><math><mrow><mo>≈</mo><mn>0.59</mn></mrow></math></EquationSource></InlineEquation> on dense Yelp), recovers catalogue coverage lost by standard contrastive training in the sparse regime, and closely matches tuned fixed-<InlineEquation ID="IEq4"><EquationSource Format="TEX">$\beta $</EquationSource><EquationSource Format="MATHML"><math><mi>β</mi></math></EquationSource></InlineEquation> PDCL without a per-dataset <InlineEquation ID="IEq5"><EquationSource Format="TEX">$\beta $</EquationSource><EquationSource Format="MATHML"><math><mi>β</mi></math></EquationSource></InlineEquation> search. To motivate the design, we also report controlled diagnostics of a text-gated knowledge-distillation variant and identify two dense-domain failure channels: gate-gradient collapse and self-normalized inverse propensity scoring (SNIPS) optimization instability. The resulting method and analysis give practical guidance on when popularity-debiased contrastive learning is useful, when fixed debiasing is preferable, and where true cold-start items remain outside the method’s scope.</p>

Adaptive popularity-debiased contrastive learning for long-tail recommendation: diagnosing failure mechanisms and bridging the accuracy-diversity gap

期刊论文
出版模式：

开放获取
发表日期：

2026年7月9日

查看全文

Yunan Zhang,
Jingjing Fan,
Yanxiao Liu

摘要

Long-tail recommenders often over-fit popular items because the training log is highly skewed: a small head catalogue dominates both ranking gradients and the negative samples used by contrastive learning. This paper studies that failure mode under full-ranking evaluation and proposes Adaptive Popularity-Debiased Contrastive Learning (A-PDCL), a lightweight training-time correction for graph contrastive recommendation. The core idea is simple: in the item-side InfoNCE denominator, popular in-batch negatives are down-weighted and tail negatives are up-weighted using smoothed inverse item degree. Rather than requiring a separate manual debiasing sweep for every dataset, A-PDCL sets the exponent automatically as $\beta _{\textrm{eff}}=\beta _{\max }\cdot \textrm{Gini}_{\textrm{batch}}$ $β_{eff} = β_{max} \cdot {Gini}_{batch}$ , so batches with stronger popularity concentration receive stronger correction. We evaluate A-PDCL on Yelp, Amazon Books, and MIND using a unified global-time-split, full-ranking protocol. Multi-seed results show that A-PDCL selects regime-appropriate correction ($\beta _{\textrm{eff}}\approx 0.75$ $β_{eff} \approx 0.75$ on sparse Amazon Books and $\approx 0.59$ $\approx 0.59$ on dense Yelp), recovers catalogue coverage lost by standard contrastive training in the sparse regime, and closely matches tuned fixed-$\beta $ $β$ PDCL without a per-dataset $\beta $ $β$ search. To motivate the design, we also report controlled diagnostics of a text-gated knowledge-distillation variant and identify two dense-domain failure channels: gate-gradient collapse and self-normalized inverse propensity scoring (SNIPS) optimization instability. The resulting method and analysis give practical guidance on when popularity-debiased contrastive learning is useful, when fixed debiasing is preferable, and where true cold-start items remain outside the method’s scope.