A Popularity Knob That Still Works When New Items Arrive

Recommender systems learn from logs where a few items get most of the clicks, so they end up recommending what is already popular. A recent, practical fix gives operators a single dial at serving time: divide each item's score by a power of its embedding length, since that length mostly records how often the item was clicked during training. Turning the dial trades raw accuracy for more exposure of long-tail items, with no retraining. The dial has a blind spot, though. Brand-new items have no clicks, so they get their vectors from a content encoder that maps images or text into the recommender's space. The length of such a vector records nothing about popularity. A dial setting chosen on historical warm items therefore does something unpredictable to new items, and new items are exactly where operators most want control.

We propose to make the length of every item vector an explicit, calibrated popularity estimate. For items with clicks, it is learned from those clicks. For new items, it is predicted from content by a head trained to reproduce what warm items' lengths would be. Direction and length are then kept separate. The claim is that one dial setting, tuned once on warm validation data, then yields predictable accuracy and exposure for cold items too. We test this directly: we measure how far the cold-item trade-off lands from the target when the dial is tuned only on warm items.

Core Claim

Test-time popularity debiasing and inductive cold-start recommendation have each been solved separately, but their composition is miscalibrated. Test Time Embedding Normalization for Popularity Bias Mitigation (TTEN) motivates the proposal. It shows item embedding length correlates with training frequency (Pearson up to 0.83) and scores items as

$$\hat r_{ui} = \cos(e_u, e_i)\,\|e_u\|\,\|e_i\|^{1-p},$$

where $e_u, e_i$ are user and item embeddings and $p\in[0,1]$ moves from inner product ($p=0$) to pure cosine ($p=1$). Contrastive Learning for Cold-Start Recommendation (CLCRec) supplies the cold-start machinery. It aligns content projections with collaborative embeddings and, during training, randomly swaps in content projections. Nothing constrains the length of those projections to share the frequency scale of warm items, so $\|e_i\|^{1-p}$ means one thing for warm items and something arbitrary for cold ones.

Our fix is to parameterize every item as $e_i = m_i\, d_i$. Here $d_i$ is a unit direction and $m_i>0$ is a scalar length. For warm items, $m_i$ is learned freely under an inner-product loss, so it absorbs popularity in the way TTEN documents. For cold items, both $d_i$ and $m_i$ come from content, and the length head $g(c_i)$ is trained on warm items whose histories are masked. It regresses onto their learned $\log m_i$. The dial exponent then acts on a quantity with the same meaning, "expected popularity", across the whole catalog. Here "fairness" means item-side exposure fairness across popularity strata, including brand-new items. User demographic fairness is out of scope.

Relation to Prior Work

Three families of work shape the design. The first is geometric popularity debiasing, which motivates the proposal and constrains its form. TTEN and Adap-τ both treat embedding length as the carrier of popularity. Adap-τ automates the temperature rather than exposing a dial, and neither handles zero-interaction items. MACR takes a causal route: it subtracts an item-only popularity branch at inference. That branch is keyed on item identity and so is undefined for cold items. Our length head can be read as a content-driven version of that item-only branch, with a continuous exponent in place of a fixed subtraction. PAAC aligns tail items to co-interacted head items. That requires at least one interaction, and its own reported gains shrink on sparse Gowalla: the popular–unpopular gap is reduced by 34.18% there, versus 87.5% on Yelp2018.

The second family is cold-start methods, which supply the method. We reuse CLCRec's content encoder and its hybrid replacement training. We borrow the masking idea of PAM: hide warm items' histories so the model practices cold prediction, with a self-supervised target for the eventual "hot" state. PAM, however, bins items by fixed popularity thresholds and offers no continuous serving-time control. We keep the masking and drop the bins. MAMO adapts quickly from 5–15 interactions but has no popularity or trade-off mechanism.

The third family is tunable fairness-accuracy trade-offs, which frames the evaluation. GFB and YODO give inference-time control through threshold shifts or weight interpolation. Both are evaluated only on classification benchmarks, and neither asks whether a dial tuned on one population transfers to a population absent at tuning time. That transfer question is our main evaluation criterion. PUFR uses predictive uncertainty as a dial in ranking. It warns that confident-but-biased scores resist correction, which is a reason our length head must be calibrated rather than merely present.

The file also records a tension between the two contrastive traditions. SimGCL fights sparsity by pushing embeddings toward uniformity on the sphere, while CLCRec needs semantic alignment. Splitting direction from length lets us apply uniformity to directions only and leave popularity in $m_i$. We test whether this split helps the alignment side or merely coexists with it.

Research Program

We build the model on a LightGCN-style backbone with CLCRec's content encoder. It has an explicit direction/length split and a length head $g$. Training minimizes three terms: a sampled-softmax recommendation loss on $m_i d_i$, CLCRec's user–item and representation–embedding contrastive losses applied to directions, and a masked regression loss $\|\log g(c_i) - \operatorname{sg}(\log m_i)\|^2$ on warm items with hidden histories. Here $\operatorname{sg}$ stops gradients, so popularity estimates are not pulled toward content. At serving time the TTEN exponent is applied to $m_i$ for every item.

Beyond ranking metrics, we analyze two quantities. The first is how well $g$ ranks cold items' eventual popularity. We compare $g(c_i)$ with the length the item acquires after it warms up, using chronological splits. The second is whether cold-item exposure varies monotonically with $p$, as warm-item exposure does.

Baselines

The naive reference is CLCRec as published with plain inner-product scoring, which is $p=0$ with no dial. It shows how far cold items are over- or under-exposed by default. No absolute Recall numbers for CLCRec were found on Lacuna. The file reports only relative cold-start Recall@10 gains of 103.95% on TikTok and 96.24% on Amazon over its baselines, so we rerun it from the same splits.

The strongest competitor is TTEN applied directly to CLCRec's output embeddings. This is the obvious composition. It uses the same dial and differs from our method only in having no calibrated cold-item length. TTEN reports 4.26%–31.5% Recall@20 gains over its best baselines on warm benchmarks. No reported result was found on Lacuna for TTEN on cold-start items.

The second post-hoc competitor subtracts $\lambda \log(1+n_i)$ from each score, where $n_i$ is the item's observed interaction count.

The strongest weight-space alternative is YODO-style interpolation between a CLCRec model trained with inner product and one trained with cosine scoring.

PAM is included on MovieLens-25M in its streaming protocol, where it reports +32.73% Recall@5 and +57.81% NDCG@5 over its best baseline. It serves as a cold-start-and-popularity reference, not a dial.

Two ablations isolate the key ingredient. In the first, every cold item gets the median warm length. In the second, $g$ is trained without masking, i.e. on unmasked warm features.

All methods share the backbone, embedding size, data splits, and an equal hyperparameter search budget per method. Every dial method is swept over the same number of settings. Beating the field means two things. Our frontier's hypervolume (HV, the area a trade-off curve dominates, as used by GFB) on the cold split must exceed TTEN-on-CLCRec across seeds, with no loss on the warm split. And our warm-to-cold dial-transfer error, defined below, must be clearly smaller.

Why not just…?

Why not just apply TTEN to CLCRec's embeddings? This is the main threat. If CLCRec's content projections happen to inherit sensible lengths, the plain composition may already transfer well. Nothing in the file shows it fails. Our transfer-error experiment compares it directly with our method. If its transfer error matches ours, the calibration head is unnecessary and we would report that.

Why not just penalize log interaction count? It is monotone and trivially covers cold items. But it gives every new item the same maximal boost regardless of content, so it treats "no data yet" as "niche". We expect the cold-item precision it achieves at a matched exposure level to be worse. The matched-exposure comparison in the experiments settles this.

Why not just set cold-item lengths to the median warm length? If content carries little signal about future popularity, this constant does as well as a learned head and is simpler. It is included as an ablation. Our correlation analysis between $g(c_i)$ and each item's eventual length says in advance which regime a dataset is in. We expect short-video datasets with visual features to favor the learned head more than MovieLens genres do.

Why not use YODO weight interpolation? It gives a dial without any geometric assumption. But both endpoint models must handle cold items, and interpolating weights does not fix a mismatch in length scale between content and collaborative embeddings. Its hypervolume and transfer error are measured on the same footing as ours.

Proposed Experiments

Cold-start experiments use datasets with content features: TikTok and Kwai (multimodal), Amazon (text), and MovieLens-1M (metadata). A held-out fraction of items gets zero training interactions. MovieLens-25M uses chronological splits, both to compare against PAM and to observe eventual item lengths. Warm-only sanity checks on Yelp2018 and Gowalla with the TTEN codebase confirm that our split does not hurt TTEN's warm behavior.

Metrics are Recall@20 and NDCG@20, reported separately on warm, tail, and cold items. We also report the share of top-20 slots going to cold and tail items, and HV of the accuracy–exposure frontier.

The decisive check is dial transfer. An operator picks the $p$ (or $\lambda$, or YODO $\alpha$) that hits a target tail-exposure share on warm validation data. We then measure realized cold-item exposure and cold Recall@20 on test. Transfer error is the gap between that realized cold-item trade-off and the one obtained by an oracle dial tuned on the cold items themselves. The proposal succeeds if its transfer error is consistently lower than TTEN-on-CLCRec across targets and datasets, with cold exposure monotone in $p$. It fails if the median-length ablation or plain TTEN matches it.

A secondary check is whether confining uniformity to directions changes CLCRec's reported sensitivity to more than 256 negatives. This would tell us whether the direction/length split eases the tension between uniformity and alignment.