Anchored Ranking for Infinite Latent Features

Recommender and scoring systems usually learn from two kinds of evidence at once. One is which items a person engages with, which is a ranking signal. The other is how much they like them, which is an absolute rating. Models that discover latent features nonparametrically, meaning they let the data decide how many hidden traits exist instead of fixing that number in advance, have so far been fit to only one of these signals at a time. This note proposes a single Bayesian latent feature model in which both signals act on the same unbounded set of binary traits. The pointwise rating term pins scores to an absolute scale. Once the scale is pinned, the sharpness of the ranking term becomes something the model can estimate from data rather than a weight someone tunes by hand.

The motivation is a specific failure that is easy to state. A pure ranking objective cannot tell "item A beats item B by a lot" apart from "all scores are just stretched." Scores can be multiplied by any constant without changing a single ordering, so their scale is unidentifiable. This is the scale drift that makes uncalibrated ranking scores unusable when they must be fused with other signals, as reported for Baidu's production ranker in Pre-trained Language Model based Ranking in Baidu Search. Our second, more interesting claim follows from putting both signals into a model that can grow its feature set. The two signals reveal different hidden traits. Some traits govern what people choose to consume but barely move their ratings. A model fit only to ratings will never create those traits. A model with a fixed number of dimensions has to trade them off against rating traits. A model with an unbounded feature set can simply add them.

Core Claim

We claim two things. First, in a coupled Indian Buffet Process (IBP) model, a joint likelihood over explicit ratings and implicit chosen-versus-unchosen comparisons makes the ranking sharpness parameter identifiable. The IBP is a prior over binary user-by-trait and item-by-trait matrices with an unbounded number of columns. Because the parameter is identifiable, it can be sampled along with everything else, replacing the hand-tuned balancing weight that hybrid losses currently need. Second, the number of features the model discovers grows when the ranking channel is added. The extra features are interpretable as "selection traits," and they account for the hybrid model's top-N gains. If the second claim fails, meaning the hybrid model discovers the same features as the ratings-only model and gains only through reweighting, then the nonparametric ingredient is not doing work, and a fixed-dimension hybrid matrix factorization is the better tool. The experiments are built to show which of these holds.

Relation to Prior Work

Two lines of work meet here, and neither has crossed into the other.

The hybrid-loss line supplies the motivation. Improving Top-N Recommendation with Heterogeneous Loss combines a least-squares reconstruction loss with a pairwise hinge ranking loss in an item-by-item linear model. It shows that the combination beats both pure pointwise and pure pairwise baselines on top-N precision. The model has no latent features at all, its memory grows as the square of the catalog size, and its balancing weight is tuned by hand. CliffRank pairs a mean-squared-error anchor with a thresholded listwise loss in fixed-dimension neural networks. It constrains our design by reporting that ranking terms switched on too early wreck feature learning, which forced a delayed schedule beginning at epoch 120. RRHF uses the same anchor-plus-ranking pattern for language model alignment, adding a supervised term on the best response to a ranking loss. All three treat the mixing weight as a tuning knob. We instead make the ranking sharpness a sampled quantity, which is possible only because the anchor fixes the scale.

The nonparametric line supplies the method. Infinite latent feature models and the Indian buffet process defines the IBP prior. A Coupled Indian Buffet Process Model for Collaborative Filtering (IBP-HFA) places one IBP over users and one over items and fits a Gaussian likelihood to star ratings. It is our direct starting point, and its weakness is that it optimizes only rating error. Infinite Latent SVM and its journal version Bayesian inference with posterior regularization show how to impose margin penalties on IBP posteriors. They also report that learning features jointly with the predictor beats extracting features first, which we test against CliffRank's opposite finding. Max-Margin Nonparametric Latent Feature Models for Link Prediction infers its regularization constant through a hyperprior, which is the precedent for inferring rather than tuning. Discriminative Nonparametric Latent Feature Relational Models with Data Augmentation uses Polya-Gamma augmentation to make logistic losses conditionally Gaussian. We reuse that trick, along with its warning that sampling the binary matrix takes over 95% of training time.

None of these papers places pointwise and pairwise terms on a single IBP posterior. None asks whether the ranking term changes which features are discovered.

Model

Each user $i$ has binary traits $z_i$ and each item $j$ has binary traits $x_j$, both drawn from coupled IBPs. Continuous loadings turn these traits into vectors $u_i, v_j$, and the score is $s_{ij} = u_i^\top v_j$, as in IBP-HFA. The observed ratings $r_{ij}$ and the sampled triples $(i,j,k)$ have separate sources. A triple pairs an item $j$ the user rated with an item $k$ they did not. The two channels therefore come from different observations, so the same data is not counted twice. The joint likelihood is

$$ p(r, c \mid s) = \prod_{(i,j)\in R} \mathcal{N}(r_{ij};\, s_{ij}, \sigma^2) \prod_{(i,j,k)\in C} \text{sigmoid}\big(\kappa\,(s_{ij} - s_{ik}) + b\big), $$

where $R$ is the set of rated pairs, $C$ the set of sampled comparisons, $\sigma^2$ the rating noise, $\kappa$ the ranking sharpness, and $b$ an offset. Without the Gaussian term, $\kappa$ and the scale of $s$ trade off exactly. With it, the scale of $s$ is set by the ratings, so $\kappa$ is identified and receives a Gamma prior.

Polya-Gamma augmentation, which introduces one auxiliary variable per comparison, makes both channels conditionally Gaussian in the loadings. Loadings, $\sigma^2$, $\kappa$ and $b$ then all have closed-form Gibbs updates. Each binary entry $z_{ik}$ is updated using user $i$'s ratings plus a freshly resampled set of a few negatives per positive. The cost is therefore linear in the user's interactions rather than quadratic in the catalog. For larger datasets we also derive a truncated mean-field variant, following IBP-HFA. We do not claim to solve the binary-sampling bottleneck. We report its cost honestly.

Research Program

The work has three parts. First, we derive and implement the augmented Gibbs sampler and the variational variant, and verify that $\kappa$'s posterior concentrates when ratings are present and fails to concentrate when they are removed. That contrast is the identifiability claim made visible.

Second, we build a synthetic generator with three kinds of planted traits. Rating traits shift rating values only. Selection traits shift the probability of interaction only. Shared traits do both. Some traits in each group are rare, owned by only a few items. This generator lets us ask the mechanistic question directly: which planted traits does each model recover? We measure recovery by matching discovered columns to planted columns, and we separately track how often rare traits are recovered.

Third, we run the model on real recommendation data and measure, for each discovered feature, how much it contributes to rating likelihood versus comparison likelihood. This splits features into rating-dominant and selection-dominant groups. We also test whether the top-N gains disappear when the selection-dominant features are removed after training. We additionally test CliffRank's scheduling finding. If the rating anchor gives the IBP enough grounding, joint training from the first sweep should be stable. If it does not, we will see the feature count blow up or collapse early, and a delayed schedule should fix it.

Baselines

The naive reference point is item popularity for top-N ranking and a user-plus-item bias model for rating error. Any method that cannot clearly beat these on sparse users is not discovering anything.

The strongest existing nonparametric method is IBP-HFA from A Coupled Indian Buffet Process Model for Collaborative Filtering. On MovieLens it reports RMSE 0.8015 on the full test set and 0.8318 on users with at most 20 ratings, against 0.9366 for SVD in that sparse regime. We rerun it on our own splits with the same truncation level and sweep budget as our model. Our model beats it if it matches IBP-HFA's rating error within noise while clearly improving Precision@5 and MAP. Matching on rating error is required, because losing calibration would defeat the purpose of the anchor.

The strongest existing hybrid-loss method is Improving Top-N Recommendation with Heterogeneous Loss, which reports Precision@5 = 0.318 on Yahoo! Music. The file gives no reported numbers for it on MovieLens 100k or 1M. We rerun it on Yahoo! Music and MovieLens with its balancing weight tuned on validation, giving it the same number of hyperparameter trials our model spends on its two IBP concentration parameters. Beating it means higher Precision@5 on Yahoo! Music under the same protocol, together with calibrated rating predictions, which it cannot produce at all.

We also run a fixed-dimension hybrid matrix factorization with the same two losses. Its dimension and balancing weight are tuned jointly within the same trial budget. This is the baseline most likely to tie us.

Ablations isolate the key ingredients. A ratings-only version of our model is essentially IBP-HFA trained by our sampler. A comparisons-only version should show drifting score scale and an unidentified $\kappa$. A version with $\kappa$ fixed on a grid tests whether inferring $\kappa$ matters. A decoupled version first learns features from ratings and then fits only the ranking head, which tests the joint-versus-decoupled finding of Infinite Latent SVM.

Why not just…?

Why not just use fixed-dimension matrix factorization with a mixed rating-and-ranking loss and tune the dimension? This is the real threat. It may tie us on accuracy, and the file reports no numbers that rule it out. The equal-budget comparison settles the accuracy question. The proposal's distinctive claim does not rest on accuracy alone, though. It rests on the synthetic recovery experiment, where a fixed dimension must split a single budget between rating traits and selection traits, while the IBP can add selection traits without displacing rating traits. If the tuned fixed-dimension model recovers planted selection and rare traits as well as ours, we concede that the nonparametric part is unnecessary.

Why not just train IBP-HFA on ratings and then rerank or recalibrate its scores? Post-hoc reranking can only reorder items using features that already exist. If selection traits never enter a ratings-only posterior, no reranker can recover them. The decoupled ablation measures exactly this gap. Infinite Latent SVM found joint learning superior to decoupled learning for classification, but whether that holds for ranking is what we test.

Why not just use many continuous dimensions with automatic shrinkage instead of binary IBP features? Continuous shrinkage priors can prune unused dimensions. They tend to keep a dense set of small effects rather than a few rare, crisp traits owned by a handful of items. We include a large-dimension shrinkage hybrid model as an extra baseline under the same budget. We compare both top-N accuracy and rare-trait recovery on the synthetic data. If shrinkage matches us on both, the binary structure is not earning its sampling cost.

Proposed Experiments

The datasets are the synthetic planted-trait generator; MovieLens 1M, using both rating values and the fact of rating; and Yahoo! Music, which allows a direct comparison with the reported heterogeneous-loss result. Metrics are RMSE, overall and for sparse users; Precision@5 and MAP; expected calibration error of ratings; posterior concentration of $\kappa$; the number of discovered features; and planted-trait recovery.

MovieLens needs a guard. Our Model Achieves Excellent Performance on MovieLens: What Does it Mean? shows that MovieLens interactions largely reflect the platform's elicitation process. Shuffling interaction order cut HR@10 by 41.6–52.0% for sequential models. This means "selection traits" found on MovieLens could be elicitation artifacts rather than taste. We therefore also fit the model on interactions from a single collection period, and report whether selection-dominant features survive or vanish. We treat Yahoo! Music as the cleaner real-data test.

Three outcomes decide the proposal. First, $\kappa$'s posterior must concentrate with ratings present and spread out without them; otherwise the identifiability argument is wrong. Second, on synthetic data the hybrid model must recover planted selection-only traits that the ratings-only model misses, and must do so better than the tuned fixed-dimension hybrid model; otherwise the nonparametric claim is wrong. Third, on Yahoo! Music, removing selection-dominant features after training must erase most of the Precision@5 gain over IBP-HFA, while rating error stays within noise of IBP-HFA's. Passing all three would show that pairing a rating anchor with a ranking term changes what an unbounded latent feature model discovers, not merely how it weights its losses.