When a recommender meets a brand-new item, it judges the item by its first handful of clicks. The system effectively asks how many users touched the item, and which ones. Those early clicks were not earned in a vacuum. People have a fixed amount of attention, so adding options to a page mostly redistributes clicks rather than creating new ones, as Improving Website Hyperlink Structure Using Server Logs showed from server logs. A new game released the same week as a blockbuster in its genre, or a new film released next to a franchise sequel, collects few early clicks because it was crowded out, not because it is bad. A cold-start model that reads raw early clicks as a quality signal confuses "lost to a strong rival" with "unappealing." The same item released in a quiet week would be ranked very differently.
We propose to treat each early interaction of a cold item as a choice made among the items that were competing for that user's attention at that moment. Under this view, the item's latent position is inferred from which competitors it beat, not only from how often it was picked. Two time-varying quantities enter the estimate. The first is the user's drifting taste, which determines which rivals are relevant to that user. The second is the changing pool of concurrently popular items, which sets how hard it was to win a click. The claim is narrow and testable: cold-item estimates that ignore this local, time-varying competition are biased in proportion to the competitive pressure at launch, and correcting for it improves cold-item ranking beyond what a global time adjustment can achieve.
Core Claim
Few-shot cold-start recommenders treat a cold item's support set as exchangeable evidence. Here the support set means the first $K$ interactions, and exchangeable means the model ignores when and against what those interactions happened. Cold-start Sequential Recommendation via Meta Learner (Mecos) matches a session against $K\in\{1,3,5\}$ such interactions. MAMO and Preference-Adaptive Meta-Learning for Cold-Start Recommendation (PAML) personalize the starting parameters a model fine-tunes from. All three score candidates independently, and none conditions on the competitive context in which the support interactions were generated.
We claim that the support set should instead be explained by a choice model over a time-indexed, user-specific competitor set:
$$ P(u \text{ picks } i \text{ at } t) = \frac{\exp s(h_u(t), e_i)}{\sum_{j \in C_u(t)} \exp s(h_u(t), e_j)}, $$
Here $h_u(t)$ is user $u$'s sequential state at time $t$, which carries their drifting preferences, and $e_i$ is the item embedding. The score $s$ is the base recommender's compatibility function. The set $C_u(t)$ contains the items that were available and plausibly salient to $u$ at $t$: items already released, weighted toward those currently popular in $u$'s neighborhood of the embedding space. Fitting $e_i$ for a cold item under this likelihood credits a click won against strong local rivals more than one won in an empty field. A meta-learned initializer turns the $K$ support choices into an embedding in one forward pass.
The specific novelty is the local competitor set. A global, per-time-period normalization would only rescale everyone's clicks equally. The hierarchical-competition finding of App Download Forecasting: An Evolutionary Hierarchical Competition Approach suggests otherwise: attention is reallocated within categories, so a blockbuster mainly displaces items near it. If that holds at the level of individual users, then local competition is what matters.
Relation to Prior Work
The motivating evidence comes from the attention-competition literature. Feature Driven and Point Process Approaches for Popularity Prediction shows that cascade models which ignore interference overestimate final popularity. Public Opinion Field Effect and Hawkes Process Join Hands for Information Popularity Prediction (POFHP) adds an inter-topic "distraction" term to a self-exciting process and reports that this lowers popularity-prediction error. Both operate on aggregate popularity counts, and POFHP assumes every topic has an observable early history. We move the competition term into personalized, few-shot item representation, where the early history is exactly what is scarce.
The closest method is Debiasing the Cloze Task in Sequential Recommendation with Bidirectional Transformers (ITPS-BERT4Rec). It already models a user's single choice among exposed items and lets exposure probability vary over the sequence. It differs from our proposal in three ways. Its propensities depend on sequence position and the user's own history, not on which rival items were trending. It is evaluated on warm items. It does not estimate representations for items with $K$ interactions.
We reuse its semi-synthetic evaluation idea: a MovieLens-100K-derived world with known exposure and relevance. We also run it as a baseline.
The cold-start meta-learners supply our few-shot machinery and the evaluation protocol for "cold" items, which follows Mecos's bottom-20%-by-interaction split. They also constrain the design. PAML and Public Opinion Field Effect Fusion in Representation Learning for Trending Topics Diffusion both report that widening a neighborhood beyond one or two hops adds noise. Mecos reports that too many matching iterations crowd the embedding space. For these reasons we build $C_u(t)$ from a small, one-hop neighborhood rather than broad propagation.
Research Program
The system is a wrapper around an existing sequential recommender. We plan to use SASRec and SR-GNN, the base models in the Mecos experiments.
We will first construct, for every logged interaction, a competitor set $C_u(t)$. It contains released items restricted to the user's top-$M$ nearest neighbors under $h_u(t)$, with each neighbor's weight raised by its popularity over a trailing window. Because $h_u(t)$ is the user's current sequential state, a user whose taste has drifted faces a different set of rivals than their long-run profile would suggest. This is how temporal drift enters the method. We are not building a separate drift model.
From these sets we will define a per-interaction competitive pressure: the popularity-weighted similarity mass of rivals at time $t$. This quantity is both a diagnostic and an input feature.
The cold-item encoder takes the $K$ support interactions, each paired with its user state and pooled competitor embeddings, and outputs $e_i$. It is meta-trained on warm items by simulating cold starts: we truncate each warm item to its first $K$ interactions in chronological order and train against the choice likelihood above on held-out later interactions.
The analysis side asks whether the conventional estimate is biased by launch conditions. For each cold item we will compute the error of its few-shot embedding relative to its eventual warm embedding, learned from all of its data. We then regress that error on its average launch pressure. If competition confounds early evidence, baseline errors should correlate with pressure, and our estimator should flatten that relationship.
Baselines
The naive reference point is the base sequential recommender with cold items embedded from their $K$ interactions by ordinary training. We also include an item-popularity ranker, which shows how much any personalization adds.
The strongest cold-start competitor is Mecos Cold-start Sequential Recommendation via Meta Learner layered on SR-GNN. It is the only prior method in our file that does few-shot cold start inside a sequential model without metadata, which matches our setting exactly. Its paper reports average HR@10 improvements over plain SR-GNN of 99% on Steam, 91% on Amazon Electronic, and 70% on Tmall. These are relative gains; no absolute HR@10 value was reported on Lacuna for Mecos, so we will reproduce it ourselves.
The strongest competitor on the exposure side is ITPS-BERT4Rec Debiasing the Cloze Task in Sequential Recommendation with Bidirectional Transformers. It reports NDCG@10 of 0.3844 on MovieLens 1M against 0.3262 for BERT4Rec, though that result is on warm items. We will run it on the cold split with its temporal propensities applied to the support interactions.
We include MAMO MAMO as a non-sequential meta-learning reference. It reports a cold-user, cold-item MAE of 0.8894 on MovieLens 1M against 0.9299 for MeLU. That is a rating-error metric, so it does not transfer to our ranking metrics, and we cite it only as a reference.
Two ablations isolate the key ingredient. The first, the global-normalization ablation, replaces $C_u(t)$ with all items released by $t$ and reweights each support click by the inverse of total platform activity in its time window. This keeps time-awareness but removes locality. The second, the static-competitor ablation, uses a user's long-run profile instead of $h_u(t)$ to choose rivals. This tests whether the drift component matters.
All methods share the same chronological splits, the same $K$, the same base encoder dimension, and the same candidate sets at evaluation. Each gets an equal hyperparameter search budget (the same number of trials on a shared validation fold) and comparable GPU hours.
We count the proposal as winning if it beats Mecos and the global-normalization ablation on cold-item NDCG@10 across seeds at the $K=1$ and $K=3$ settings. It must also substantially reduce the correlation between launch pressure and embedding error. Gains confined to $K=5$, or matched by the global ablation, would count as a loss.
Why not just…?
Why not just normalize early clicks by overall platform traffic in that week? This is the most serious threat. If competition is effectively a global volume effect, a time fixed effect captures it at almost no cost. The hyperlink-log result is compatible with either story: a fixed click budget does not say whether displacement is global or local. The global-normalization ablation exists to settle this. If it matches the full method, we will report that the useful correction is global and drop the local competitor machinery. What would favor locality is a clear difference between in-genre and out-of-genre blockbusters. On Steam and MovieLens we will compare cold items launched next to a same-genre hit with those launched next to an off-genre hit, holding total traffic fixed.
Why not just use ITPS-style temporal propensity weighting? ITPS corrects for exposure changing with a user's position in their sequence. It is the right tool when the confound is "this item only appears late in histories." Its propensities carry no term for which rivals were trending, so two cold items at the same sequence position but in different competitive weeks receive the same weight. Running ITPS on the cold split shows directly whether its correction already absorbs launch-pressure bias. The pressure–error regression is the specific test: if ITPS flattens it, our contribution shrinks to the few-shot encoder.
Why not just mask the softmax to items that were available at time $t$? Restricting training negatives to released items is simple and partly captures zero-sum choice, so we treat it as part of every baseline rather than as our contribution. It treats every available item as an equally strong rival. A new indie game then "competes" with the whole back catalog as much as with that week's hit in its own genre. The difference between availability masking and our popularity-weighted local set is exactly what the global-normalization ablation measures.
Why not just wait for more interactions, or use text features? Waiting makes the problem disappear, but only after the item has already been buried. Mecos's large gains at $K\le 5$ show that the early regime is where ranking decisions matter. Content features bypass the noisy-clicks problem, but they lose collaborative signal; A survey on sequential recommendation describes this as the trade-off between learned item identifiers and content features. The competition correction applies to whatever few-shot evidence exists, so we will also test it on top of a text-initialized embedding to check that the two are complementary.
Proposed Experiments
The first experiment is semi-synthetic, where the bias is known. Following the semi-synthetic MovieLens-100K protocol of Debiasing the Cloze Task in Sequential Recommendation with Bidirectional Transformers, we simulate users with known relevance and drifting preference states. Clicks are generated by a choice process over displayed slates. We then inject "blockbuster" items at controlled times, controlled genre overlap, and controlled strength.
Ground-truth cold-item quality is known in this world, so we can measure estimation bias directly. We do this both as a function of competitor strength and as a function of genre overlap, comparing the full method, both ablations, Mecos, and ITPS. This experiment decides whether local competition can be recovered at all when it truly exists. It also shows whether the global ablation fails when competition is genuinely local and succeeds when it is genuinely global, which validates that the ablation discriminates between the two.
The second experiment uses real data with natural shocks: Steam, Tmall, and Amazon Electronics under the Mecos cold split, plus MovieLens 1M and 20M using rating timestamps. Game launches and movie release clusters provide natural variation in launch pressure. We report HR@10 and NDCG@10 on cold items with $K\in\{1,3,5\}$, evaluated against time-respecting candidate pools.
The decisive check is stratified. We split cold items by launch-pressure quintile, and separately by whether the dominant rival shared their category. The proposal predicts three things. Baselines should degrade most in the high-pressure, same-category stratum. Our method's advantage should concentrate there. The static-competitor ablation should lose ground mainly for users whose sequential state moved substantially between their long-run profile and the launch window.
If none of these patterns appears, the honest conclusion is that cold-start evidence is not meaningfully confounded by competition in these logs. That negative result would still matter, because it would tell the cold-start literature that its exchangeable support sets are a safe simplification.
References
- Improving Website Hyperlink Structure Using Server Logs
- Cold-start Sequential Recommendation via Meta Learner
- MAMO: Memory-Augmented Meta-Optimization for Cold-start Recommendation
- Preference-Adaptive Meta-Learning for Cold-Start Recommendation
- App Download Forecasting: An Evolutionary Hierarchical Competition Approach
- Feature Driven and Point Process Approaches for Popularity Prediction
- Public Opinion Field Effect and Hawkes Process Join Hands for Information Popularity Prediction
- Debiasing the Cloze Task in Sequential Recommendation with Bidirectional Transformers
- Public Opinion Field Effect Fusion in Representation Learning for Trending Topics Diffusion
- A survey on sequential recommendation