Protecting Seasonal Demand When Debiasing Recommenders

Recommender systems learn from click logs in which a few items collect most of the interactions. Much of the field's work on "popularity bias" tries to stop models from recommending those items just because they are frequent. Recent work adds a refinement: some popularity reflects real quality and should be kept, while some reflects herd behavior or over-exposure and should be removed. Every existing method draws that line without asking when the clicks happened relative to the calendar. A Christmas album that surges each December looks the same to these methods as a track pushed by a viral loop. Current debiasers therefore either suppress it every December or treat each year's surge as fresh, transient herd behavior. We propose to draw the line with a seasonal forecast. The part of an item's popularity that a seasonal time-series model predicts from previous cycles is treated as legitimate demand and kept at recommendation time. Only the part the calendar cannot explain, the residual shock plus the long-run exposure-driven level, is handled by standard causal debiasing.

Stated for a colleague in two sentences: split each item's popularity into a long-run level, a recurring seasonal shape shared with similar items, and unexplained shocks, then debias only the level and the shocks. This should raise accuracy on in-season niche items and at seasonal turning points without letting the head of the catalogue back in.

Core Claim

The closest existing work already makes popularity time-dependent, but none of it models periodicity, and none uses a forecast to decide what to remove. TIDE splits interactions into a static item quality term, a conformity term, and user–item matching. The conformity term is a sum of exponentially decaying contributions from past clicks, and TIDE blocks it at inference. This motivates our proposal: it shows that keeping a benign popularity signal improves accuracy. It also constrains it, because under monotonic decay a recurring seasonal surge can only be classified as new conformity. CausalEPP tracks each user's evolving taste for popular items and each item's local popularity. It removes confounding with backdoor adjustment and uses rolling moving averages to project both forward. It is the strongest dynamic competitor. Its forecaster lags at inflection points, and its authors name forecasting cost as a limitation. DPA learns a popularity-free score, then re-injects the observed popularity of the current day at test time. This supplies our inference mechanism of "strip, then add back the right amount," but it adds back empirical counts rather than a forecast, and its Alibaba log covers one month.

Our claim differs in which component is protected. Popularity that is predictable from past seasons is protected, and everything else is debiased. The forecaster is a means, not the contribution. We expect the result to hinge on the decomposition and on pooling seasonal shapes across items, not on choosing SARIMA over a simpler seasonal predictor. We designed the experiments so that this expectation can fail.

Model

Let $n_{i,t}$ be the interaction count of item $i$ in week $t$, let $s$ be the season length (52 weeks for annual cycles), and let $g(i)$ be a group of items (a product category, or a cluster of co-consumed items). We fit

$$\log(1 + n_{i,t}) = \ell_{i,t} + \sigma_{g(i)}\big(t \bmod s\big) + e_{i,t},$$

where $\ell_{i,t}$ is a slowly varying item level, $\sigma_g$ is a seasonal profile shared by all items in group $g$, and $e_{i,t}$ is the residual shock. The seasonal profile is shared for a practical reason. Fitting a seasonal ARIMA model to each item is too expensive at catalogue scale, and it is meaningless for long-tail items with a handful of clicks. Both problems appear among the open issues of the forecasting direction. Pooling lets a niche winter item borrow the winter shape of its category.

The seasonal ARIMA model, whose seasonal differencing $(1-B^s)Y_t$ compares each week to the same week one cycle earlier, is fit per group on aggregated counts, using only data before the prediction time. The same holds for its simpler substitutes, discussed below. This mirrors how the tuberculosis disruption study fit ARIMA to pre-pandemic years and read deviations from the projection as the effect of an external shock. Here the "shock" is exposure and herd behavior rather than a pandemic.

The recommender then follows TIDE's structure with one change. The level $\ell_{i}$ is decomposed into static quality and conformity exactly as in TIDE. The residuals $e_{i,t}$ feed the decaying conformity pathway in place of raw clicks. At inference we block the conformity pathway, and we add the forecast seasonal index $\sigma_{g(i)}$ for the week in which the recommendation is served. That index is a multiplicative boost on the score, clipped so it cannot exceed the forecast's own uncertainty band.

Persistent exposure-driven popularity lives in the level and is still removed. Only the recurring shape is protected. This answers the objection that a seasonal forecast would simply preserve the head of the catalogue: a seasonal index averages to zero over a cycle, so it cannot favor always-popular items.

Research Program

The work has three parts. First, a measurement study asks whether benchmark logs contain seasonality worth protecting. For each dataset we measure the share of groups with significant autocorrelation at lag $s$ and the share of interactions that fall in groups with a strong seasonal profile. We report this before building anything. If Ciao or Douban-Movie show no annual structure, the method should collapse to TIDE there, and we will say so.

Second, we build the decomposed recommender on MF and LightGCN backbones. We implement the seasonal index with four interchangeable forecasters: a seasonal ARIMA per group, a seasonal-naive "same week last cycle" estimate, a moving average (to reproduce CausalEPP's choice inside our decomposition), and zero (which recovers TIDE).

Third, we run a semi-synthetic study in which ground truth is known. Real logs cannot tell us which part of a December surge is genuine demand. We therefore extend the closed-feedback-loop simulator of Zhu et al., which retrains models on their own recommendations over iterations and showed MF drifting to a Gini of about 0.8, with injected seasonal utility for a subset of items. We then check whether the protected seasonal index amplifies the Matthew effect across iterations.

Baselines

The naive reference points are the undebiased MF and LightGCN backbones and a MostPop recommender. They show how much popularity alone explains, and at seasonal peaks MostPop may be surprisingly strong.

Among static debiasers we run MACR and Adap-τ. MACR estimates and subtracts the direct effect of item popularity with counterfactual inference. Adap-τ normalizes embedding length and adapts the softmax temperature per user. They represent the view that popularity is time-invariant.

The strongest competitor is CausalEPP. It is already dynamic and forecasting-guided, and it ranked first over DICE, MACR and TIDE. It reports Recall@20 up 20.4% and NDCG@20 up 14.7% on Amazon-Music relative to the strongest baseline. It also reduced LightGCN's share of popular items in recommendations from 75.8% to 55.5%, against a ground-truth share of 43.3%. No absolute Recall@20 or NDCG@20 values were found on Lacuna for CausalEPP, so we will reproduce it rather than quote numbers.

TIDE is the direct parent. It reports significant Precision, Recall and NDCG gains on Amazon-CDs, Ciao and Douban, but no absolute values were found on Lacuna. DPA, adapted to weekly counts, tests whether re-injecting observed rather than forecast popularity suffices. It reports Recall@1 +117.30% and Recall@5 +43.13% over DeepMatch with full softmax on the Alibaba log. We also include PPAC, which counterfactually balances popularity within a user's neighborhood of similar users against global popularity, as a static personal-popularity reference.

Every method uses the same chronological splits, backbone, embedding size and a fixed budget of hyperparameter trials tuned on a validation window. Where a forecast is used, that window lies before the test window. The ablation that isolates our ingredient keeps our model fixed and swaps only the seasonal index among zero, moving average, seasonal naive and seasonal ARIMA, each pooled and per item.

Beating CausalEPP means higher Recall@20 and NDCG@20 on test windows at seasonal transitions, significant across seeds. It also requires a popular-item share at least as close to ground truth, and no loss on off-season windows.

Why not just…?

Why not just add calendar features, letting TIDE's quality term vary by month? This is the most serious alternative, because it learns seasonality end to end without an external forecaster. Its weakness is that month-by-item parameters are estimated from very few interactions for tail items. The embedding-growth effect documented by Adap-τ suggests those parameters will grow fastest for items that dominate peak months. We will run this variant, with and without category pooling. If the pooled calendar variant matches ours, the lesson is "pool seasonal shape and keep it out of the debiased pathway," and the forecaster is dispensable. We would report it that way.

Why not just use the seasonal-naive estimate instead of seasonal ARIMA? It may well work equally well. Weekly counts on review datasets are noisy, and the file offers no evidence that ARIMA beats a same-week-last-year estimate on such data. The fits in the file, such as Liao's exchange-rate model, are in-sample, on dense series, and flagged for overfitting. The ablation settles this question. We state in advance that the contribution survives if seasonal naive wins.

Why not just lengthen CausalEPP's moving-average window? A window long enough to be stable averages across the season and erases the peak. A short window lags the onset and overshoots the end. Plotting forecast error by week relative to the seasonal peak, for the moving-average variant inside our model, shows directly whether that lag costs accuracy.

Why not just recompute a static debiaser within each month? Monthly slices starve tail items of training signal, and they only react to the current month rather than anticipating the next one. Its closest runnable form is DPA with weekly counts, which is in the baseline set.

Proposed Experiments

We use datasets with multi-year timestamps: Amazon-Music and Amazon-CDs as the primary tests, where holiday cycles are plausible, and Ciao and Douban-Movie as likely weakly seasonal controls. The Alibaba log is too short for annual cycles and is used only for weekly seasonality ($s=7$ days). We make chronological splits with several test windows per dataset, labeled by the measurement study as rising, peak, falling or flat relative to the item group's seasonal profile. We report Recall@20 and NDCG@20 overall and per phase. We also report Recall@20 on in-season long-tail items, the popular-item share against ground truth (following CausalEPP), and the Gini coefficient of exposure.

The first decisive check is phase-specific. Our method should gain most at rising and falling windows and should be indistinguishable from TIDE on flat windows and on non-seasonal datasets. Uniform gains everywhere would suggest we are measuring tuning, not seasonality. The second check is the semi-synthetic loop. With injected seasonal utility, the protected index should recover in-season items while keeping Gini near FPC-DScale-like levels across iterations. If Gini climbs toward the MF trajectory, protecting forecast demand feeds the Matthew effect, and the clipping design fails. The third check is the forecaster ablation, which determines whether any time-series machinery beyond seasonal naive earns its cost.