Recommender systems favor already-popular items, and most fixes push every user toward the long tail by the same amount. That uniform push hurts people who really do want mainstream hits. The session-level methods that personalize the push describe a user with a single number, such as the fraction of niche items in the session or the average popularity of clicked items. We propose describing each user's recent history by the shape of its popularity distribution instead: its spread, its asymmetry, and how popularity changes from one click to the next. These are the same small sets of statistical moments that speech systems use to compress a variable-length recording into a fixed-size vector. From this summary we predict how popular the user's next item will be, and we use that prediction to set a user-specific correction on any existing recommender's scores.
The bet is concrete. Take two users whose recent clicks have the same average popularity. One steadily browses mid-popularity items. The other explores obscure items with a single viral hit mixed in. These users should receive different amounts of long-tail promotion. A mean or a tail-ratio cannot tell them apart; variance and skewness can. Because moments can be updated in constant time as each interaction arrives, the summary can also stand in for the time-series forecasting that current dynamic debiasing methods run at inference, which their authors report is costly.
Core Claim
Our claim is about both evidence and method. On the evidence side, the higher-order moments of a user's recent popularity trajectory carry information about the popularity of their next interaction beyond what the mean carries. On the method side, a debiasing correction conditioned on that summary can match the utility of learned or forecasting-based conformity models on accuracy and on long-tail exposure, at close to zero inference cost. Both halves can fail independently, and the experiments are designed so that each failure would be visible.
Each source direction plays a distinct role. Statistical temporal pooling supplies the mechanism. TAP-SLDA pooled the first five moments over frozen speech features. It showed that mean pooling alone leaves keyword classes overlapping, while higher moments separate them, with a 9.9% to 11.3% relative accuracy gain on Google Speech Commands for roughly 0.12% extra parameters. The x-vector age estimator in End-to-end Deep Neural Network Age Estimation similarly needed standard deviation alongside the mean. Popularity bias supplies the problem and the regime. The same file also carries a warning that constrains our design: higher sample moments are noisy on short sequences. Many e-commerce sessions average about four items (Exploring Popularity Bias in Session-based Recommendation), and a kurtosis estimated from four numbers means essentially nothing. We therefore treat sequence length as a first-class variable rather than hoping it does not matter.
Relation to Prior Work
Two papers motivate the gap by personalizing debiasing with a single summary number. Long-Tail Session-based Recommendation from Calibration (CSBR) trains the backbone so that the popularity mix of its recommendations matches the session's scalar tail-ratio. Its authors note that this logic conflicts with user behavior on "pure head" sessions. Wang's study in Exploring Popularity Bias in Session-based Recommendation switches between a nearest-neighbor model and GRU4REC based on the session's mean item propensity. We keep their idea that the user's own history should set the target. We replace the scalar with a distributional summary and replace the tail-ratio target with a predicted distribution over popularity bins. A user with a pure-head history then gets a head-concentrated target automatically instead of a contradictory one.
Taming Recommendation Bias with Causal Intervention on Evolving Personal Popularity (CausalEPP) is the closest dynamic method. It tracks a sliding-window ratio of high-popularity interactions, separates item quality from local popularity with a causal adjustment, and forecasts conformity with moving averages. Its authors flag inference overhead as a limitation. We borrow its insistence that popularity is local in time. Specifically, we measure each item's popularity in a trailing window at the moment of interaction, not from global counts. We do not forecast; the streaming moment summary is the whole temporal model.
Several papers debias through the learned representation itself. DIRAC uses separate interest and conformity encoders with a gate between them. DCRec down-weights conformity-driven interactions in a contrastive loss over transition and co-interaction graphs. PAAC aligns tail and head item representations. These methods change the backbone. Ours is a correction layer placed on top of an unchanged backbone, so it can be stacked on any of them.
Other work constrains the design. Adap-τ shows that item embedding norms grow with interaction frequency. This is why we pool moments of scalar popularity values and not of raw item embeddings: averaging raw embeddings would let popular items dominate the pooled vector. HAM found that on sparse sequential data, self-attention collapses toward uniform averaging and that mean and max pooling suffice. That finding is encouraging, but HAM never measured popularity effects and used only first-order pooling.
Research Program
For each user $u$ at time $t$, we form the sequence of log local-popularities $x_1,\dots,x_n$ of their last $n$ interactions, capped at a window $W$. We also form the first differences $\Delta x_k = x_k - x_{k-1}$. The differences partly restore the temporal ordering that plain moment pooling throws away.
The summary $m_u$ contains the mean, standard deviation, standardized skewness and kurtosis of $x$, the mean and standard deviation of $\Delta x$, and $\log n$. Each moment of order $r\ge 2$ is shrunk toward a population prior, with a weight $n/(n+\kappa_r)$ where $\kappa_r$ grows with $r$. The weight reflects the fact that sample variance of higher moments grows with order, so short histories should fall back to the prior rather than contribute noise. All quantities are running sums, updated in constant time per interaction, as in the streaming discriminant setup of TAP-SLDA.
A small predictor maps $m_u$ to $q(b \mid m_u)$, a distribution over $B$ popularity bins for the user's next item. The predictor is trained by cross-entropy on held-out next interactions. At ranking time, the backbone's score $s(u,i)$ for item $i$ is adjusted as
$$s'(u,i) = s(u,i) + \alpha\left[\log q(b_i \mid m_u) - \log \pi(b_i)\right],$$
where $b_i$ is item $i$'s current popularity bin, $\pi$ is the popularity-bin distribution of the backbone's own top-$K$ recommendations averaged over validation users (what the backbone already favors), and $\alpha\ge 0$ is a single strength tuned on validation. When the user's predicted popularity profile matches what the backbone already produces, the correction is zero. It pushes toward the tail only for users whose history predicts tail interest, and toward the head only for users whose history predicts head interest.
Baselines
The naive reference point is each backbone with no correction ($\alpha=0$). We use GRU4REC and HAM for session and sequential data and LightGCN for the chronological collaborative-filtering data. All are trained with identical splits and hyperparameter searches.
The first strong competitor is CSBR on the same GRU4REC backbone. It is the most direct session-personalized calibration method. CSBR reports TailCoverage@20 improved by 14.01% and Tail@20 by 4.67% over its best baseline on YOOCHOOSE, and an 18.83% rise in recommended tail items on Last.fm. The file gives no absolute Recall@20 for CSBR, so we will reproduce it. Beating CSBR means equal or better Recall@20 together with a lower popularity-calibration divergence, meaning the KL between the popularity-bin mix of the recommended lists and that of users' true next items. It also means a clear win on pure-head sessions, which CSBR's authors identify as its failure case.
The second is CausalEPP on LightGCN over Amazon-Music, Ciao and Douban-Movie. It is the strongest dynamic per-user method in the file. On Amazon-Music it reports Recall@20 up 20.4% and NDCG@20 up 14.7% over the best baselines. It also cut LightGCN's popular-item recommendation rate from 75.8% to 55.5%, against a ground-truth rate of 43.3%. The file gives no absolute Recall@20 value for CausalEPP. Beating it means matching its Recall@20 and NDCG@20 within statistical noise while bringing the popular-item rate at least as close to ground truth, at a small fraction of its per-request latency.
The third is Wang's mean-propensity dynamic ensemble. It reports HR@20 of 0.556 on Diginetica and 0.586 on Retailrocket (Exploring Popularity Bias in Session-based Recommendation). It is the cleanest example of first-moment personalization.
We also run MACR as the uniform counterfactual-subtraction reference. MACR reports average HR@20 gains of 153% on matrix factorization and 241% on LightGCN, but under uniform test distributions rather than the chronological next-item evaluation used here.
The ablations isolate the key ingredient. We compare the full summary against truncations to the mean only and to mean plus spread. We also compare against a variant without difference moments, a variant without shrinkage, and a variant using global instead of local popularity. All share the same predictor, bins, $\alpha$ search and compute.
Why not just…?
Why not just use the mean popularity or tail-ratio, as CSBR and Wang do? This is the hypothesis under test, not a settled point. On sessions of about four items, it may well win, because higher moments there are mostly noise. The mean-only ablation, run on identical data and stratified by history length, will settle it. If higher-order moments add nothing even on long histories (30MU sessions average 17.11 items; the chronological user streams are longer), the evidential claim fails and we will report that.
Why not just run a small recurrent network over the raw popularity sequence? This is the most serious threat. A learned encoder can in principle recover any moment, and possibly more. We include a GRU over the same inputs, feeding the same predictor, as an upper-bound competitor. If it clearly beats moments on long histories, the method contribution shrinks to a cost argument. That argument still holds for streaming inference, since moments update in constant time with no hidden state to run, but it is weaker. We treat "moments within noise of the GRU at lower latency" as the success condition and say so in advance.
Why not just apply a uniform correction such as MACR, inverse propensity weighting, or unit-norm embeddings with Adap-τ? CausalEPP and DIRAC both report that uniform penalties hurt users who genuinely want head items. MACR's own paper notes that inverse propensity weighting is high-variance and hard to tune. Our correction reduces to a uniform one when $m_u$ is uninformative. The comparison against MACR, and against $\alpha$ tuned with $q$ replaced by its population average, measures directly how much of any gain comes from personalization rather than from overall debiasing strength.
Proposed Experiments
The first experiment is a pure diagnostic and decides whether to continue. On 30MU, Last.fm, Amazon-Music, Ciao, Douban-Movie, MovieLens-1M, Diginetica and Retailrocket, we measure the held-out log-likelihood of next-item popularity bin under predictors fed moment summaries of increasing order, stratified by history length. We expect higher moments to help only once $n$ exceeds roughly ten, because estimation noise in standardized skewness and kurtosis falls off slowly with sample size, and to give no gain on the short e-commerce sessions.
A sharper test uses users matched on mean popularity within a narrow tolerance. It asks whether their spread and skew still predict different next-item popularity distributions. If they do not, the core claim is false regardless of downstream numbers.
The second experiment is end to end. We measure Recall@20, NDCG@20 and HR@20; per-stratum recall under Wang's target-based popularity stratification; the Gini coefficient of recommendations; the popularity-calibration KL; recall on pure-head sessions; and per-request latency against CausalEPP.
The decisive outcome has two parts. The full summary must beat the mean-only ablation on long-history datasets in both calibration KL and tail-stratum recall without losing overall Recall@20. It must also match the GRU competitor and CausalEPP within confidence intervals. A gain that appears only on short-session data, or that vanishes once $\alpha$ is tuned for the uniform variant, would indicate that the benefit comes from debiasing strength rather than distributional shape.
References
- End-to-end Deep Neural Network Age Estimation
- Exploring Popularity Bias in Session-based Recommendation
- Long-tail session-based recommendation from calibration
- Taming Recommendation Bias with Causal Intervention on Evolving Personal Popularity
- Disentangling Interest and Conformity for Eliminating Popularity Bias in Session-Based Recommendation
- Debiased Contrastive Learning for Sequential Recommendation
- Popularity-Aware Alignment and Contrast for Mitigating Popularity Bias
- Adap-τ : Adaptively Modulating Embedding Magnitude for Recommendation
- HAM: Hybrid Associations Model with Pooling for Sequential Recommendation
- Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System