Recommenders trained on click logs confuse "many people saw this" with "this person wants this." Methods that remove this popularity bias usually subtract a fixed popularity effect from every user. That hurts people who really do like blockbusters. It also misses something that changes over time: a user may start following trends, or stop. The best current method of this kind, CausalEPP, tracks each user's share of popular items inside a sliding time window and adjusts recommendations by that share. But a share is a level, not a cause. A lifelong blockbuster fan and a newly trend-swayed user can have the same share. They differ in one testable way: only the second one starts consuming a group of items after that group's popularity jumps.
We propose to estimate that response directly. For each user, we ask whether a jump in a group of items' platform-wide popularity is followed by that user moving toward the group, beyond what the user's own past tastes predict. We then remove only the part of the recommendation score that comes from this measured, time-varying response. The step that makes this tractable is borrowed from causal discovery, the problem of recovering a cause-and-effect graph from data. The number of possible graphs grows faster than exponentially with the number of variables. Order-based methods tame it by fixing an ordering of the variables first, after which each variable's causes can be found separately. In logs, timestamps supply most of that ordering for free. Only a small population-level graph over item groups within a single time window still needs search. Each user then contributes just a handful of response strengths, shrunk toward the population.
Core Claim
The quantity that should drive personalized, time-aware popularity debiasing is a user's lagged causal response to popularity shocks, not their current popularity share. Three ingredients make this response estimable at recommender scale. The first is the time ordering, which removes most of the graph search. The second is popularity computed with the user's own clicks left out, which blocks the trivial path where a user's clicks create the popularity they appear to follow. The third is a hierarchical model, which pools the population structure while letting each user's response drift. If the claim is right, the learned response should predict who reacts to genuinely new popularity shocks on held-out data where EPP does not. Debiasing with it should also beat CausalEPP on chronological splits, most of all for users whose EPP and measured response disagree.
Research Program
We group items into $K$ clusters, using genre where available and otherwise a clustering of item embeddings in the style of Sequential Recommendation with Causal Behavior Discovery. Causer already showed that item-level graphs are infeasible and that cluster-level ones are useful, but it learned a single static graph with the stiff NOTEARS acyclicity penalty and had no popularity variables. Logs are cut into time windows $t$. For user $u$, let $x_{u,k}(t)$ be the share of the user's interactions that fall in cluster $k$. Let $\Delta p^{-u}_k(t)$ be the change in cluster $k$'s platform popularity, computed with user $u$'s own interactions removed. The structural model is
$$x_{u,k}(t) = \sum_{j} a_{kj}\, x_{u,j}(t-1) + \sum_{j \in \mathrm{pa}(k)} c_{kj}\, x_{u,j}(t) + \big(b_k + \beta_{u,g(k)}(t)\big)\, \Delta p^{-u}_k(t-1) + \varepsilon_{u,k}(t).$$
The terms are as follows. The coefficients $a_{kj}$ are shared lagged taste transitions. The coefficients $c_{kj}$ are within-window edges among clusters, where $\mathrm{pa}(k)$ is cluster $k$'s parents in the within-window graph. The coefficient $b_k$ is the population's response to popularity. The term $\beta_{u,g(k)}(t)$ is user $u$'s personal deviation in response to popularity for a coarse cluster group $g(k)$, and it may drift over time.
Time ordering makes every lagged edge point forward in time, so those edges need no acyclicity constraint. The only real structure search is the $K$-node within-window graph. With $K$ around 30, we can run it with the permutahedron ordering method of DAG Learning on the Permutahedron, which keeps every candidate acyclic by construction. We use it as the supplied method rather than NOTEARS, because the material reports that NOTEARS-style penalties get trapped in local minima and need hard thresholding.
The per-user terms $\beta_{u,g}(t)$ follow a Gaussian random walk across windows and are shrunk toward zero. This captures the "Right Now, Wrong Then" pattern of preference drift studied in Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift, applied here to conformity rather than to taste. Non-stationary discovery methods such as Causal Discovery in Semi-Stationary Time Series and SPACETIME: Causal Discovery from Non-Stationary Time Series motivate letting the mechanism change. However, they target low-dimensional continuous systems. Our design moves the non-stationarity into a few user-level scalars, so the shared graph stays small.
At recommendation time, a base recommender's score for user $u$ and item $i$ in cluster $k$ is split into two parts. The first is the part explained by $\beta_{u,g(k)}(t)\,\Delta p^{-u}_k$. The second is everything else. We remove the first part, scaled by a single tuned coefficient. Popularity that is stable rather than a shock stays in the score. This follows the observation in Popularity Bias is not Always Evil: Disentangling Benign and Harmful Bias for Recommendation that some popularity reflects real quality.
The main identification threat is the platform's own exposure. A recommender that pushes trending items to a user creates a popularity-to-click path that is exposure rather than taste. The public datasets in the file have no impression logs. We therefore treat estimates on real logs as "response including exposure," and we verify identifiability in simulation, where exposure is controlled.
Baselines
The naive reference is the base recommender with no debiasing: matrix factorization and LightGCN trained on the same chronological split. This tells us whether any debiasing helps on future data at all. That matters because How Graph Convolutions Amplify Popularity Bias for Recommendation? shows graph models amplify the bias, while mainstream users may still be served well by it.
The static, uniform causal reference is Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System (MACR). It subtracts a learned direct popularity effect from a fixed four-variable graph and reports HR@20 gains of 153% for MF and 241% for LightGCN on a debiased MovieLens 10M split. Those numbers come from a different evaluation protocol, so we rerun MACR ourselves. The geometric reference is Adap-τ : Adaptively Modulating Embedding Magnitude for Recommendation, which rescales embedding lengths with adaptive temperatures and reports a 9% average gain.
The strongest competitor is Taming Recommendation Bias with Causal Intervention on Evolving Personal Popularity (CausalEPP). It is the only method that is already personalized and time-varying. It reports Recall@20 improved by 20.4% and NDCG@20 by 14.7% over the best baseline on Amazon-Music, with an Average Rank of 1.00 on Ciao and Douban-Movie. No absolute Recall or NDCG values were found on Lacuna, so we reproduce it.
For session data we add Disentangling Interest and Conformity for Eliminating Popularity Bias in Session-Based Recommendation (DIRAC), which uses two encoders for interest and conformity. No reported numeric result was found on Lacuna for it.
All methods share several controls. They use the same base backbones and the same chronological splits. Each gets an equal hyperparameter search budget of 50 trials, tuned on a validation window that precedes the test window. Our method's per-user parameters are fit only on data before the test window.
Two ablations isolate the key ingredient. The first, "EPP-for-β," replaces the learned response with CausalEPP's popularity share inside our score correction. The second uses the response computed with the user's own clicks included. Beating CausalEPP means a statistically significant gain in Recall@20 and NDCG@20, measured over five seeds with paired tests across users. It also means the gain must concentrate in users whose EPP and measured response disagree. A gain spread uniformly across users would point to tuning rather than mechanism.
Why not just…?
Why not just use CausalEPP's sliding-window popularity share? This is the real threat. If EPP and the measured response rank users almost identically, our method adds machinery for nothing. The EPP-for-β ablation and the shock-prediction test below settle it directly. We report the rank correlation between the two quantities. If that correlation exceeds about 0.9 and the ablation matches the full method, we will report that EPP suffices.
Why not just add popularity features to a sequential transformer such as SASRec and let it learn conformity? A transformer that is fed $\Delta p$ can learn to predict trend-following. But prediction does not give the decomposition needed to remove the trend-driven part. Sequential Recommendation with Causal Behavior Discovery also found that attention latches onto co-occurrences that are not causal. We run SASRec with popularity-change features, plus a counterfactual variant that zeroes those features at inference. If that variant matches us, the structural model is not needed, and we would say so.
Why not just run per-user Granger regressions, a regression that asks whether past popularity changes improve prediction of the user's next behavior, with no causal discovery at all? This is largely what our lagged component is. The honest claim is that time order removes most of the combinatorial problem. Two additions are what remain. The first is the within-window cluster graph. Without it, two clusters that are both popular and consumed together in the same window would be misread as popularity response. The second is hierarchical pooling, without which most users have too few windows to estimate anything. An ablation that drops the within-window edges and an ablation without pooling measure each addition. If both ablations are harmless, the contribution reduces to leave-one-out lagged response, which is still a measurable improvement over EPP if that holds.
Why not just normalize embeddings or calibrate the tail ratio? Evolution of Popularity Bias: Empirical Study and Debiasing shows that static corrections do not stop bias from growing in closed feedback loops, where the model's recommendations become its own future training data. The calibration authors of Long-Tail Session-based Recommendation from Calibration concede that calibration conflicts with intent in sessions made only of popular items. Adap-τ stays in our comparison so this is tested rather than assumed.
Decisive Experiments
The first test is a semi-synthetic closed loop on MovieLens 1M, following the iterative retraining setup of Evolution of Popularity Bias: Empirical Study and Debiasing. There, plain MF reaches a Gini coefficient of about 0.8, while FPC-DScale holds it near 0.2. Simulated users are given known mixes of drifting taste and drifting conformity, and exposure is controlled. We measure how well $\beta_u(t)$ recovers the true conformity, using area under the curve against ground truth and including users whose conformity switches mid-stream. We also track Gini and total clicks over retraining rounds against FPC-DScale. This test checks identifiability, which real logs cannot provide.
The second test uses natural shocks in real logs on Amazon-Music, Ciao, and Douban-Movie, which are CausalEPP's chronological datasets. We define a shock as a cluster whose leave-one-out popularity jumps into the top percentile of week-over-week change. We then predict which users move toward that cluster in the following window, comparing pre-shock $\beta_u$ against pre-shock EPP. This out-of-sample causal-direction check is the core falsification test. If EPP predicts shock response as well as $\beta_u$, the central claim fails.
The third test is end-to-end recommendation on the same chronological splits. We report Recall@20 and NDCG@20, tail coverage, and Gini, broken down by EPP-versus-response disagreement quartile. We also run session-level replications on Last.fm and YOOCHOOSE against DIRAC and the calibration method. Success means three things: the shock test favors $\beta_u$, accuracy gains are largest in the disagreement quartiles, and the method holds Gini down in the closed loop at no cost in clicks.
References
- Sequential Recommendation with Causal Behavior Discovery
- DAG Learning on the Permutahedron
- Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift
- Causal Discovery in Semi-Stationary Time Series
- SPACETIME: Causal Discovery from Non-Stationary Time Series
- Popularity Bias Is Not Always Evil: Disentangling Benign and Harmful Bias for Recommendation
- How graph convolutions amplify popularity bias for recommendation?
- Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System
- Adap-τ : Adaptively Modulating Embedding Magnitude for Recommendation
- Taming Recommendation Bias with Causal Intervention on Evolving Personal Popularity
- Disentangling Interest and Conformity for Eliminating Popularity Bias in Session-Based Recommendation
- Evolution of Popularity Bias: Empirical Study and Debiasing
- Long-tail session-based recommendation from calibration