A news portal, a video app, or a growing product catalog never ranks a fixed list of items. New articles appear every minute, and the most important ranking decision is often about an item nobody has clicked yet. Two research traditions each handle half of this. Bayesian nonparametric ranking models handle an infinite pool of items and can say how likely the next choice is to be something never seen before. Feature-based rankers can score a brand-new item from its text or category. Neither tradition has been run inside a live stream with a response deadline. We propose a single model that does all three. Each item's prior desirability is set by its features. The total desirability of all items, including those not yet published, stays finite and is computed in closed form. Every click updates the posterior in constant time per item, so the model fits within a stream's latency budget.
In two sentences: a new article gets an informative starting score from its content, then a cheap Bayesian counter refines it with every click. The same machinery tells the system how much attention will flow to articles that do not exist yet, which is the quantity a recommender needs to decide how many slots to give to fresh content. The test is whether this beats the popularity counters that have historically won live news-recommendation contests, under a strictly causal replay where no future information can leak into a prediction.
Core Claim
The claim has three parts. First, a Gamma-process Plackett-Luce model whose base measure depends on item features can be filtered online with per-click cost proportional to the number of items touched, not to the catalog size. Second, the result matches feature-only and identity-only rankers on cold items and on warm items respectively. Third, its forecast of "click mass going to not-yet-published items" is calibrated in a way that a learned "fresh item" classifier is not. If the third part fails, so that a simple classifier forecasts novelty as well, the contribution shrinks to an engineering one, and we say so up front.
Some terms first. The Plackett-Luce model gives each item a positive weight $\lambda_i$ and picks item $i$ with probability $\lambda_i$ divided by the summed weights of the competitors. Bayesian nonparametric models for ranked data. makes this work over infinitely many items. It draws the weights from a Gamma process, a random measure with countably many atoms whose total weight is almost surely finite. It also derives a closed-form novelty mass $w_*$, the weight of all unseen items. That paper motivates our work and supplies the core mechanism. Its limits are also our starting point: its items are featureless atoms, it uses offline Gibbs sampling, and its evaluation is retrospective modeling of weekly bestseller lists. Bayesian nonparametric Plackett-Luce models for the analysis of clustered ranked data extends the same construction to clustered rankings, again without features.
From the feature side, Dyad ranking using Plackett-Luce models based on joint feature representations scores items through a bilinear form $x^\top W a$ or a neural network, which allows zero-shot ranking of unseen candidates. It ranks only finite, static candidate sets and has no notion of unseen mass. We reuse its bilinear scorer as our feature map.
Model
Each item $i$ arrives at publication time with feature vector $x_i$. Its weight has prior
$$\lambda_i \mid x_i \sim \mathrm{Gamma}\big(\alpha_0\, e^{f(x_i)},\ \beta\big),$$
where $f$ is a bilinear or small-network scorer of the item's text, category, and source, $\alpha_0$ is a concentration parameter, and $\beta$ is a rate parameter. New items are treated as arriving as a Poisson process at rate $\rho$, with features drawn from a recent-arrivals distribution. The expected weight of items that will appear in the next window of length $\Delta$ then follows from the Gamma mean and Poisson thinning:
$$\mathbb{E}[w_*^{(\Delta)}] = \frac{\rho\,\Delta\,\alpha_0}{\beta}\,\mathbb{E}_x\big[e^{f(x)}\big].$$
This is the feature-conditioned analogue of the novelty probability in Caron and Teh. It is what the system uses to set the fraction of slots reserved for fresh items.
Updating the posterior relies on two facts. Plackett-Luce has a Thurstonian form, in which each item's "arrival time" is exponential with rate $\lambda_i$ and the earliest arrival is the choice. Given an auxiliary exponential variable, the Gamma prior is then conditionally conjugate. Bayesian inference for Plackett-Luce ranking models exploits this structure with batch Power Expectation Propagation and shows that Gamma priors stop maximum likelihood from overrating rarely observed competitors on NASCAR data.
We replace the batch passes with a single-pass assumed-density filter. Each item keeps a Gamma posterior with shape $a_i$ and rate $b_i$. When a click on item $c$ arrives, $a_c$ increases by one, and the rate of every item that was available receives the expected auxiliary increment. The full catalog's contribution to that increment is maintained as one running sum, so it is never recomputed. A per-tick discount on $(a_i, b_i)$ lets desirability drift.
Feature-seeding also partly addresses the "instant hit" failure that Caron and Teh flag for their smooth Pitt-Walker dynamics. An article whose features predict popularity starts high, with no history needed. The scorer $f$ is trained by stochastic gradient steps on the prior predictive log-likelihood of clicks on items in their first minutes. These steps are run asynchronously, off the serving path.
Research Program
The work has three parts. The first is derivation and analysis. We will derive the filter's update rules. We will bound the per-event cost. We will also study a monotonicity question raised by Generalizing while preserving monotonicity in comparison-based preference learning models. That paper shows that feature-based Bradley-Terry models can lower an item's score after it wins a comparison, and it restores monotonicity with diffusion priors on a static item graph. Our design separates a per-item posterior from a shared feature prior. We conjecture that a click can never lower the clicked item's posterior mean relative to items that were displayed alongside it. The reason is that the feature map changes only the prior, never the per-item evidence. We will try to prove this and will test it empirically. Their result constrains our design: it rules out putting all item evidence into shared feature weights.
The second part is a strictly causal streaming harness built on the NewsREEL Replay setup. CLEF 2017 NewsREEL Overview: Offline and Online Evaluation of Stream-based News Recommender Systems and CLEF 2017 NewsREEL Overview: A Stream-Based Recommender Task for Evaluation and Education replay timestamped plista messages from February 2016, preserving arrival order and response deadlines. Following Time Travel is Cheating: Going Live with DeepFund for Real-Time Fund Investment Benchmarking, which shows that systems that look good in back-tests lose money once evaluated after their training cutoff, we enforce three rules. All normalization statistics are computed from the past only. Pretrained text encoders must have cutoffs before the replay window. Any response that misses the deadline counts as a miss.
The third part ports the same harness to chronological streams from the Micro-video, Kwai, and Amazon Clothing datasets. These datasets are currently evaluated with static 8:1:1 splits by Temporally and Distributionally Robust Optimization for Cold-Start Recommendation and Contrastive Learning for Cold-Start Recommendation. Porting them tests whether those papers' static-split rankings survive causal replay.
Baselines
The naive reference point is a recency-weighted popularity counter that falls back to most-recent items in the same category for zero-click articles. Its NewsREEL Live counterpart is BL2Beat, reported at 1.17% online click-through rate (CTR) in CLEF 2017 NewsREEL Overview: Offline and Online Evaluation of Stream-based News Recommender Systems. The strongest live result in that paper with meaningful volume is ORLY_KS at 2.09% CTR over 42,786 impressions. Riadi_NV_01 reached 2.70%, but over only 443 impressions. That paper reports that popularity-style methods beat complex content models online, which makes this the baseline to fear.
No reported result found on Lacuna for NewsREEL Replay accuracy, so we will reimplement popularity and hybrid fallbacks in our harness. We will tune their decay windows on the first replay week, with the same tuning budget we give our own hyperparameters.
The strongest feature-based competitor is CLCRec, which aligns content features with collaborative embeddings contrastively. Its reported results are relative cold-start Recall@10 gains of 103.95% on Tiktok and 96.24% on Amazon over prior baselines, with no absolute values in the file. We also include TDRO, which corrects group-robust training toward emerging feature trends. No reported number was found in the file for TDRO's absolute Recall@20 on Micro-video. Both models will be run causally under the same deadline. They will be warm-started on the pre-window history and fine-tuned online at a fixed update cadence on the same hardware.
Three ablations isolate each ingredient. The first removes features, giving a streaming Caron-Teh model with a constant base measure. The second removes the infinite measure, giving a finite feature-conditioned Plackett-Luce model in the style of bilinear dyad ranking with no $w_*$. The third removes the per-item posterior, so items are scored by $f(x)$ alone.
Beating the field means two things. On the plista replay, the full model must exceed the tuned popularity hybrid on replay clicks overall and on items in their first 30 minutes, with confidence intervals from day-level blocks, while meeting the deadline at the 99th percentile. Its novelty forecast must also have lower calibration error than every alternative forecaster.
Why not just…?
Why not just use a trending counter with a content fallback? This is the real threat. NewsREEL's live results favor it, and it costs almost nothing. Its weakness is structural. A counter gives every zero-click article the same score, so it can only order fresh items by an ad-hoc fallback rule. It also cannot say how much traffic fresh items deserve. The first-30-minute slice of the replay will settle this. If the counter matches us there, the feature prior adds nothing, and we will report that.
Why not just fine-tune a two-tower contrastive model online and add a "fresh item" classifier? CLCRec-style training gives strong cold-start retrieval offline. A separate classifier predicting "the next click goes to an item younger than $\Delta$" could imitate $w_*$. We test this head-on by comparing that classifier's calibration against our closed-form forecast across arrival-rate shocks, such as breaking-news hours. Our forecast moves with the observed $\rho$ and the feature mix of arrivals without retraining. A classifier must relearn to track these shifts. A further practical constraint is that CLCRec reports degradation beyond about 256 negatives, which limits how far its normalizer can stand in for an open catalog.
Why not just run an online Bayesian linear bandit or sliding-window maximum likelihood? Contextual bandits appeared among NewsREEL entries. A linear bandit is essentially our third ablation plus exploration. Sliding-window maximum likelihood is known to overrate rarely seen items, as the NASCAR analysis in Guiver and Snelson shows. Including both as baselines tells us whether the per-item Gamma posterior is doing the work.
Decisive Checks
The plista replay is the headline experiment. The metrics are replay click recall at top-$k$ overall and for items under 30 minutes old, the 99th-percentile response latency against the platform deadline, and the expected calibration error of the novelty forecast in hourly windows.
The Micro-video, Kwai, and Amazon causal replays test whether the static-split ordering of CLCRec and TDRO survives once leakage is removed and deadlines are enforced. A reversal there would be a finding on its own.
A targeted stress test injects synthetic "instant hits" into the replay: articles with high-popularity features and no history. This measures how many clicks each method needs before ranking such an article in its top five.
If NewsREEL Live's successor platform accepts submissions, we will run a live comparison against the popularity hybrid. If it does not, we will state plainly that all results come from replay. The project fails cleanly if the feature-free ablation matches the full model on fresh items, or if the classifier forecasts novelty as well as $w_*$.
References
- Bayesian nonparametric models for ranked data.
- Bayesian nonparametric Plackett-Luce models for the analysis of clustered ranked data
- Dyad ranking using Plackett-Luce models based on joint feature representations
- Bayesian inference for Plackett-Luce ranking models
- Generalizing while preserving monotonicity in comparison-based preference learning models
- CLEF 2017 NewsREEL Overview: Offline and Online Evaluation of Stream-based News Recommender Systems
- CLEF 2017 NewsREEL Overview: A Stream-Based Recommender Task for Evaluation and Education
- Time Travel is Cheating: Going Live with DeepFund for Real-Time Fund Investment Benchmarking
- Temporally and Distributionally Robust Optimization for Cold-Start Recommendation
- Contrastive Learning for Cold-Start Recommendation