When a recommender adds a new item, it must decide how much exposure to give it before it has any behavior data. As the first impressions arrive, it must also decide how quickly to revise that guess. That second decision depends on something most cold-start systems never model: how bursty the item's engagement is. Engagement counts on real platforms are overdispersed, meaning their variance far exceeds their mean. Burstiness is the within-item version of this: one share can produce a spike of clicks that says little about the item's long-run appeal. A model that treats counts as Poisson (variance equal to mean) reads that spike as strong evidence and over-promotes the item. The same model reads an early drought as evidence the item is bad and buries it.
We propose to learn, from an item's content alone, two quantities besides its predicted engagement rate. The first is how informative the content prior is. The second is how bursty the item's counts will be, which sets how much each early observation should count. Both enter a simple Gamma-rate posterior. That posterior drives Thompson sampling, the exploration rule that shows each item with probability proportional to its chance of being best, during the item's warm-up period. The claim is testable: if burstiness is predictable from content, content-conditioned dispersion should give better-calibrated warm-up forecasts and fewer false promotions than either a Poisson update or a single global dispersion constant. If it is not predictable, the idea fails early and cheaply.
Core Claim
Overdispersion matters little for ranking cold items at time zero, because ranking depends on the predicted mean and not on the variance. It matters a great deal for the transition from cold to warm, because the variance-to-mean ratio sets the effective sample size of early counts. This follows from the law of total variance: a count with rate heterogeneity has variance equal to its mean plus the variance of its rate. Under a Negative Binomial model with mean $\mu$ and dispersion $\phi$, the variance-to-mean ratio is $1 + \mu/\phi$. A count therefore carries only $1/(1+\mu/\phi)$ as much information about the rate as a Poisson count of the same size.
Our model predicts, from content features $x_i$ of item $i$, a prior mean rate $m(x_i)$, a prior strength $\kappa(x_i)$, and a burst dispersion $\phi(x_i)$. After exposure windows $t$ with exposures $e_t$ and counts $c_t$, the approximate posterior over the item's rate $\lambda_i$ is
$$\lambda_i \mid \text{data} \approx \text{Gamma}\Big(\kappa m + \sum_t \rho_t c_t,\; \kappa + \sum_t \rho_t e_t\Big), \qquad \rho_t = \big(1 + e_t \hat m / \phi(x_i)\big)^{-1},$$
where $\hat m$ is the current posterior mean. Each window is tempered by its inverse variance-to-mean ratio. When $\phi \to \infty$, the update reduces to the standard Poisson-Gamma update. The new ingredient is that $\phi$ is predicted from content, so a meme video and a technical paper warm up at different speeds before either has been seen.
Relation to Prior Work
Two lines of work sit on either side of this proposal, and neither covers the warm-up transition. The first handles counts but ignores cold start. Hierarchical Negative Binomial Factorization replaces Poisson with Negative Binomial in collaborative factorization and outperforms Poisson factorization on Last.fm play counts. Its dispersion is tied to collaborative latents, so it cannot say anything about an item with no interactions. Content-based recommendations with Poisson factorization (CTPF) recommends new arXiv and Mendeley articles through text topics. It forces equi-dispersion on both words and clicks, so its posterior would react to early counts in exactly the over-confident way we target. Deep Double Poisson Networks supply the mechanism we reuse. They predict input-dependent dispersion, and their learned down-weighting of surprising counts is the regression analogue of our tempering factor $\rho_t$. They also report that a Poisson network's out-of-distribution detection collapses to 0.330 AUROC against 0.887 for Double Poisson ensembles. That result motivates, but does not establish, our hypothesis that mismodeled noise corrupts exploration.
The second line handles cold start but ignores count noise. DropoutNet trains content networks by randomly hiding collaborative inputs, so the model learns to work without history. We borrow that trick directly: during training we truncate item histories to random prefixes, including empty ones. DropoutNet differs from our proposal in two ways. It regresses warm scores with squared error, and it produces no notion of how fast to update. Epinet for Content Cold Start adds an epistemic head, a small network that expresses uncertainty about the model's own prediction, and uses it for Thompson sampling over new Facebook Reels videos. It models per-impression binary likes, which is the strongest reason to doubt our proposal. Correlated bursts across impressions of the same item still violate its independence assumption, and nothing in Epinet adjusts for that. Coupled Compound Poisson Factorization shows empirically that residual variance in sparse rating data scales with observation probability. This supports the premise that variance is structured and potentially predictable, not constant.
Research Program
The first stage is a cheap go/no-go test. On items with long histories, we fit per-item Negative Binomial dispersion from windowed counts, conditioned on exposure. We then regress log-dispersion on content features: text embeddings for arXiv and CiteULike, and job-posting profiles for XING. If held-out explained variance is negligible, content cannot tell us how bursty an item is. In that case the proposal reduces to "use a global dispersion," and we would report that as the finding.
The second stage trains $m$, $\kappa$ and $\phi$ jointly. Each training example takes a warm item's history, samples a prefix length from zero upward, forms the tempered posterior above, and scores the posterior-predictive likelihood of the item's next windows of counts. The posterior predictive of a Gamma rate under Negative Binomial noise is approximated as Negative Binomial by moment matching. This objective asks directly how well the model forecasts the future given what it has seen so far, at every stage of warm-up. The content network is shared across the three heads.
The third stage plugs the posterior into Thompson sampling in a replay simulator. The simulator is built from the full logged trajectories of items. Each simulated new item replays its real exposure-conditional count stream, resampled by block bootstrap so that bursts are kept intact. The scope is item-level engagement rate: the item's appeal averaged over the users it is shown to. Personalization is left to the underlying ranker and combined with the rate as a multiplier. This keeps the contribution isolated and makes it compatible with two-tower systems.
Baselines
The do-nothing reference is greedy allocation on the content-predicted mean with no exploration and no updating beyond a running average. A Poisson-Gamma update with content-predicted mean and prior strength ($\phi = \infty$) is essentially CTPF's implied behavior at the item level. It is the most direct test of whether overdispersion matters. A global-dispersion Negative Binomial update, with one $\phi$ tuned on validation data, is the ablation that isolates our key ingredient: content-conditioned dispersion. A log-transformed Gaussian Thompson sampler, with DropoutNet-style content training and a tuned noise variance, is the practitioner's default.
The strongest existing method is an Epinet-style ensemble head over a Bernoulli per-impression model. We will reimplement it on the same content network with the same number of hyperparameter trials. For reference, Epinet reports roughly a 32.7% increase in likes per impression for videos with 0–100 impressions over greedy on Facebook Reels (Epinet). That is a live A/B result that cannot be reproduced offline. DropoutNet reports a 6% cold-start recall@100 improvement over DeepMusic on CiteULike, which we will reproduce as a sanity check on our content network. For warm-up predictive likelihood or replay regret on arXiv, XING or Last.fm, no reported result was found on Lacuna.
All methods share data splits, the content encoder architecture, and an equal tuning budget. Beating the strongest method means lower cumulative regret over each new item's first 100 exposures, together with better held-out log-likelihood of next-window counts. Both improvements must hold with confidence intervals across bootstrap replays and not only on average.
Why not just…?
Why not just use a single global dispersion constant? This is the most serious threat. If most of the benefit comes from switching Poisson to Negative Binomial at all, content-conditioned dispersion adds complexity for nothing. The global-dispersion ablation settles it directly. We will also report where any gap appears. Our hypothesis predicts gains concentrated among items whose predicted $\phi$ lies in the extreme deciles. It predicts little difference near the median.
Why not just log-transform counts and use Gaussian Thompson sampling? A constant noise variance on log counts cannot represent variance that depends on both the mean and the item type, which CCPF's heteroscedasticity finding contradicts. Gaussians also assign mass to impossible values, which Deep Double Poisson Networks link to poor probabilistic calibration (measured by CRPS, a score for how well a predicted distribution matches outcomes). This baseline is nonetheless tuned and included, because a well-tuned variance might recover most of the regret gap. If it does, we will say so.
Why not just use Epinet on per-impression likes? Per-impression Bernoulli losses sidestep count dispersion only if impressions are independent. Shares and trending create correlated runs that the epistemic head interprets as reduced uncertainty. Our simulator preserves those runs by block bootstrap. We will run an Epinet head alongside our tempered posterior and a hybrid that feeds $\rho_t$-tempered counts to Epinet. That comparison tells us whether dispersion modeling complements or replaces learned epistemic uncertainty.
Why not just fit Negative Binomial factorization on warm data and regress its factors from content? The two-stage pipeline recovers a mean prior but leaves dispersion unpredicted for new items, as noted above for HNBF. It is also trained only on the endpoints of full histories, not on the warm-up states in which it must operate. We include it with an added content-to-dispersion regression, which isolates the value of the joint prefix-truncation objective.
Decisive Checks
The first check is the go/no-go regression of per-item dispersion on content. The second tests the mechanism rather than the headline metric. We measure the false-promotion rate, the fraction of items that receive top-quintile allocation after their first window but finish in the bottom half of true rate. We also measure the premature-burial rate, the mirror-image error. Both should fall relative to the Poisson update, and most sharply for high-burst items. The third check is calibration of forecasts after the first $k$ exposures, for $k$ in {0, 1, 5, 20, 100}. We use probability integral transform histograms, which should be flat if forecasts are calibrated, and CRPS. Gains should shrink as $k$ grows, since the prior washes out with data. If the gains instead persist or appear only at large $k$, the improvement comes from something other than warm-up dispersion. We run these checks on arXiv usage logs, where new submissions arrive continuously; on the XING RecSys 2017 data for job-posting cold start; and on Last.fm play counts, where the overdispersion documented by HNBF is strongest.
References
- Hierarchical Negative Binomial Factorization for Recommender Systems on Implicit Feedback
- Content-based recommendations with Poisson factorization
- Fully Heteroscedastic Count Regression with Deep Double Poisson Networks
- DropoutNet: Addressing Cold Start in Recommender Systems
- Epinet for Content Cold Start
- Coupled Compound Poisson Factorization