Item Content Helps the Long Tail but Hurts Popular Items in Sequential Recommendation under a Lexical Encoder: Frequency-Gated Content Injection

Item content is widely used to help sequential recommenders rank long-tail items, yet its effect on popular items is rarely isolated. On three Amazon Reviews 2023 categories, uniformly injecting item titles and categories into a SASRec backbone through a fixed lexical content encoder raises long-tail Recall@10 by 10.6–17.8 points but lowers head Recall@10 by 1.7–4.5 points. This head cost is specific to the lexical encoders we test: with a frozen semantic encoder at 768 dimensions, it disappears in all three categories. Frequency-conditioned fusion weights have been explored in recent multimodal and long-tail recommenders; we study a minimal single-scalar form, Frequency-Gated Content Injection (FGCI), where a per-item scalar α_i scales the content embedding and is learned jointly with the backbone from the item's frozen, standardized log interaction count. FGCI improves head Recall@10 over the no-content baseline in all three categories, while long-tail Recall@10 stays 8.4–15.8 points above it. A natural explanation is that FGCI simply injects more content; we rule this out. A constant scale fixed a priori at the head-item level misses FGCI by 5.1–6.0 points (unpaired p < 0.01 in all three categories). Nor can the gate be fitted after the backbone has been trained. An ablation on Health and Household, one of the three categories, shows the same for training-time gating alone: it drives head Recall@10 23.8 points below the no-content baseline. Together, these controls show that the gain is not reproduced by any unconditional, post-hoc, or single-sided variant we test. What, then, does the gate learn? In all three categories the learned α_i is largest for head items and smallest for long-tail items, inverting the intuition that the items with the least interaction history need the most content. That inversion does not explain the recovery either: head items keep 85–88% of their content weight under the gate, yet a constant scale at that value does not recover head performance at all. This rules out the most natural static-weight reading of the gate and indicates that the recovery requires conditioning to be per item and learned jointly, rather than residing in the specific gate values. Why this conditioning helps remains an open question that our controls have narrowed but not resolved.