Recommenders learn from everything a user has ever clicked, so a user whose tastes have changed keeps receiving suggestions anchored to who they used to be. The usual fixes are blunt. Systems either down-weight old data uniformly by age, which also erodes interests that have not changed, or they build specialized sequential architectures. We propose a narrower tool. For each user, ask which specific past interactions, if deleted, would make the model better at predicting what that user does now. Then remove exactly those interactions with a cheap approximate unlearning step. The question "what would the model predict if this training point were gone" is answered by influence functions, a classical approximation that estimates a retrained model from gradients and a curvature term without retraining. Recommendation-unlearning work already made these fast and graph-aware for privacy deletion. We aim them at a different target: deleting data that has gone stale.
In two sentences for a colleague: score every historical interaction by its estimated effect on the user's most recent behavior, and unlearn the ones that hurt. The resulting per-user, per-interaction forgetting should keep stable interests intact where time decay cannot. It should also loosen the "filter bubble" that stale interests create, because diversification methods measure distance from a history that now includes obsolete anchors. A user who has drifted is partially cold again, since their recent evidence is only a handful of interactions. So the cold-start regime becomes the hardest stress test, not a separate problem.
Core Claim
We hypothesize three things. First, in chronologically split recommendation data, a small fraction of each drifted user's history carries negative influence on their recent interactions. Second, those interactions can be found with scalable influence estimates. Third, unlearning only them improves next-period accuracy over both uniform time decay and fine-tuning on recent data, specifically for users whose drift is localized to some interests while others stay fixed. A secondary claim concerns the filter bubble. Removing stale anchors should raise fine-grained entity coverage of future recommendations, with no explicit diversity objective and without the usual accuracy cost, because part of what looks like over-specialization is really inertia from obsolete data.
Relation to Prior Work
The method comes from the unlearning literature. Recommendation Unlearning via Influence Function (IFRU) estimates parameter changes from removing interactions using Hessian-vector products. It also tracks "spillover" through graph aggregation layers, and it reports over 250x acceleration versus retraining. RRL: Recommendation Reverse Learning deletes interactions with a reversed ranking loss plus a Fisher-information regularizer, reporting about 3.5 seconds versus over 12,000 seconds for retraining on Gowalla. Both take the set of points to delete as given, because a user or regulator names them. We keep their deletion machinery unchanged and contribute the missing piece: an influence-based rule that decides which interactions to delete, using the user's own recent behavior as the target.
The motivation comes from drift work. Causal Disentangled Recommendation Against User Preference Shifts (CDR) models chronological "environments" with a temporal variational autoencoder. That lets a localized shift, such as a life event, leave unrelated stable preferences untouched, and CDR reports that uniform decay fails for this reason. We accept that diagnosis but test a lighter remedy that changes the data rather than the architecture, so it applies to any trained backbone. Unexpected interest recommender system with graph neural network (UIRS-GNN) separates short-term from long-term interest with an attention-based recurrent unit and scores unexpectedness as distance from clusters of past interests. Stale interactions contaminate exactly those clusters.
The filter-bubble link is constrained by Knowledge Graph Context-Enhanced Diversified Recommendation (KG-Diverse). It pushes user embeddings away from the historical mean and measures entity coverage. As the research file notes, that mean is computed over all history, obsolete interests included. We do not propose a new diversifier. Instead we test whether cleaning the history changes what such diversifiers, and plain accuracy-trained models, surface.
The scalability component borrows an idea from influence maximization. Influence Maximization: Near-Optimal Time Complexity Meets Practical Efficiency and Stop-and-Stare avoid forward simulation by sampling "reverse reachable" sets. These are the nodes that can reach a randomly chosen target, found by walking backward. Maximizing the Diversity of Exposure in a Social Network generalizes the same trick to exposure diversity on static graphs. We use the idea only as a screening step, described below. We do not import their approximation guarantees, which rely on coverage and submodularity structure our problem does not obviously have.
Research Program
The backbones are matrix factorization and LightGCN, trained on all interactions before a chronological cut. For each user $u$ we hold out a probe set $P_u$ of their most recent pre-cut interactions, with sizes swept from 1 to 10. Evaluation is on a later test window the model never sees. For a training interaction $z$, the influence on the probe is
$$ \mathcal{I}(z,u) = -\nabla_\theta L_{P_u}(\hat\theta)^\top H_{\hat\theta}^{-1} \nabla_\theta \ell(z,\hat\theta), $$
where $\hat\theta$ are the trained parameters, $L_{P_u}$ is the ranking loss on $u$'s probe, $\ell(z,\cdot)$ is the training loss on $z$, and $H$ is the Hessian of the total training loss. Up to sign conventions, $\mathcal{I}(z,u)$ is the first-order change in probe loss from up-weighting $z$. Negative removal effect, meaning deletion lowers probe loss, flags $z$ as stale for $u$. We use IFRU's graph-aware formulation so that the spillover of $z$ through LightGCN layers is counted.
Only interactions whose removal is estimated to lower the probe loss by more than a threshold are unlearned. The threshold is tuned on a validation cut and capped at a fraction of history, which prevents over-deletion. Unlearning runs with IFRU's one-step update and, separately, with RRL's reverse loss. Running both separates the value of the selection rule from the choice of deletion method.
Scalability is the engineering challenge. One inverse-Hessian-vector product per user is affordable, since it is shared across all of that user's candidates. Scoring every interaction against every probe is not. For LightGCN, a probe prediction depends only on edges within $K$ hops of the probe user and item. We therefore sample backward random walks from each probe pair, weighted by the model's normalized aggregation coefficients. This mirrors how reverse reachable sets walk backward from a target. The walks produce a small candidate set of training edges, and exact influence is computed only on those. Whether this screening loses important edges is an empirical question, checked against exhaustive scoring on MovieLens-1M.
The analysis stratifies users by drift type. We label a user's drift as localized when some categories or knowledge-graph entities disappear from recent activity while others persist, and as global otherwise. The core claim predicts an advantage mainly in the localized stratum. We also stratify by probe size, because with one or two recent interactions the influence estimate is noisy. That is the "re-cold user" regime, where the method is most useful if it works and most likely to fail.
Baselines
The naive reference is the backbone trained on full history with no forgetting. Every method must beat it on next-window Recall@10 and NDCG@10 for the result to matter.
The simplest serious competitors are an exponential time-decay weighting of training interactions and a sliding window that retrains on recent data only. Each gets its decay rate or window length tuned on the same validation cut and with the same tuning budget as our threshold. Fine-tuning the trained backbone on each user's probe data, with the same number of gradient steps our unlearning uses, is the strongest cheap competitor.
The strongest drift-specific method is CDR. It will be run on its own chronological environment splits of Yelp, Amazon-Book and Amazon-Electronics, using its released configuration, with our method applied to the same splits. CDR reports relative Recall@10 improvements over its strongest baselines of 19.5% on Yelp, 28.1% on Amazon-Book and 26.9% on Amazon-Electronics. These are relative figures only; no absolute Recall@10 for CDR was found on Lacuna. We will therefore reproduce CDR's absolute numbers ourselves. "Beating it" means matching or exceeding reproduced CDR Recall@10 on the localized-drift stratum while applying to an unmodified backbone. Matching it at a fraction of the training cost would be a meaningful but weaker result.
For the diversity question, the comparators are maximal-marginal-relevance reranking over categories and KG-Diverse. KG-Diverse reports Entity Coverage@20 of 87.2 on Last.FM, up from about 68 for matrix factorization. We also combine our cleaned history with KG-Diverse's repulsion term to test whether cleaning complements it.
Three ablations isolate the key ingredient, each pruning the same number of interactions per user. The first prunes at random. The second prunes the oldest interactions, which separates influence from age. The third prunes by gradient similarity alone without the inverse Hessian. A final check retrains exactly without the flagged interactions, which bounds the error of the approximate unlearning.
Why not just…?
Why not just decay old interactions by age? Decay assumes staleness is a function of time. CDR reports that uniform down-weighting damages stable long-term interests when shifts are localized. Our design tests exactly this: if tuned decay matches us in the localized-drift stratum, the per-interaction rule adds nothing and we will report that.
Why not just fine-tune on the user's recent interactions? This is the real threat. Fine-tuning also moves the model toward current behavior at similar cost. The difference is that fine-tuning adds pressure toward recent items while leaving stale evidence in place, and with tiny probes it can overfit. Deletion removes the conflicting evidence. The stratified comparison against fine-tuning at probe sizes 1 to 10 settles this. If fine-tuning matches us everywhere, the proposal's accuracy claim fails and only its interpretability, meaning which interactions went stale, remains.
Why not just use CDR? CDR requires training a specific temporal generative model. Our method is a post-hoc operation on whatever model is deployed. If CDR is clearly more accurate, our method is still useful only as a cheap retrofit, and the comparison will quantify that gap honestly.
Why not just rerank for diversity? Reranking adds variety on top of a model still anchored to obsolete interests, and KG-Diverse shows category-level reranking is too coarse. We measure whether cleaned histories raise entity coverage at fixed Recall. If reranking achieves the same coverage-accuracy curve, the filter-bubble claim is refuted.
Proposed Experiments and Decisive Checks
The first decisive check uses planted drift. On MovieLens-1M and Amazon-Book we splice real histories so that a known block of a user's older interactions comes from a different user in a disjoint genre. These are "stale" by construction, while the rest of the history and the recent probe stay consistent. We then measure precision and recall of flagging the planted block, compared against age-based and gradient-only flagging. If influence cannot find planted staleness, the method fails at its premise.
The second check runs on natural data. We report next-window Recall@10 and NDCG@10 on Yelp, Amazon-Book and Amazon-Electronics against all baselines above, split by drift type and probe size. The prediction is a gain concentrated in localized drift and a degradation, relative to fine-tuning, at probe size one.
The third check concerns the filter bubble. On Last.FM and Amazon-Book we plot Entity Coverage@20 and UIRS-GNN-style unexpectedness against Recall@10 for cleaned versus uncleaned histories, with and without KG-Diverse. The claim holds only if cleaning shifts the curve outward rather than trading along it.
The fourth check is scalability. On Gowalla we report the wall-clock time of scoring plus unlearning, placed next to RRL's reported 3.5 seconds for deletion and the over-12,000-second retraining reference. We also report candidate-screening recall versus exhaustive influence scoring on MovieLens-1M. A screen that misses most negative-influence edges would reduce the method to a small-graph tool.
References
- Recommendation Unlearning via Influence Function
- RRL: Recommendation Reverse Learning
- Causal Disentangled Recommendation against User Preference Shifts
- Unexpected interest recommender system with graph neural network
- Knowledge Graph Context-Enhanced Diversified Recommendation
- Influence maximization: near-optimal time complexity meets practical efficiency
- Stop-and-Stare: Optimal Sampling Algorithms for Viral Marketing in Billion-scale Networks
- Maximizing the Diversity of Exposure in a Social Network