Forgetting by Shape: Interest-Level Decay for Drifting User Preferences

Recommenders and preference-tuned language models usually handle changing tastes with one blunt tool. They down-weight old interactions by their age, either through exponential decay or a sliding window. That tool cannot tell an interest the user has abandoned from one they simply have not touched lately. A user who stopped watching cooking videos because they lost interest and a user who stopped because the feed stopped showing them look the same under time decay, and both get forgotten. We propose to forget by interest rather than by time.

The method splits a user's history into its separate interest clusters using topological tools that find which groups of embedded items are connected at many scales at once. It then compares the recent window against the long-term history to see which clusters have gone quiet, and checks the user's own negative feedback to decide whether each quiet cluster was rejected or only unvisited. Only rejected clusters decay. The working rule is "absence of evidence is not evidence of absence."

Core Claim

Our main hypothesis is that a large share of real preference drift is component-level: whole interest clusters appear, die, or go dormant. Dormant clusters are mishandled by global recency weighting. A second hypothesis is that a cross-barcode, a topological comparison that isolates the connected pieces of one point cloud that a second cloud does not cover, finds these clusters without fixing their number in advance. Combined with explicit skip and short-view signals, it should yield a history-weighting rule that beats time decay on users whose interests change shape, and that matches time decay on users whose tastes are stable.

A third hypothesis, which we test before building the method, is that drift in real exposure-controlled logs is not mostly "reweighting among stable topics." Persistent Homology for Distribution Drift Detection in LLM Embedding Streams found that this is the regime where topology loses to a plain covariance test, with AUC 0.641 against 0.740. If real drift is mostly of that kind, the proposal's premise fails, and we will report that.

Relation to Prior Work

The closest work is Persistent Homology for Distribution Drift Detection in LLM Embedding Streams. It shows that a Wasserstein distance between connected-component persistence diagrams detects synthetic, mean-preserving reorganizations of text-embedding streams better than MMD-RBF (Maximum Mean Discrepancy with an RBF kernel), with AUC 0.858 against 0.782. That paper motivates our work and also constrains it. It produces one alarm per stream window on synthetic shifts. It does not model individual users, it does not say which part of the distribution changed, and it does not feed the alarm back into a model. We differ in three ways: we work per user on real logs, we localize drift to specific interest clusters, and we use that localization to set training weights.

It also supplies practical settings that we adopt. Windows of 50–100 points worked best, and projecting embeddings to 20–50 dimensions with PCA (principal component analysis) reduced noise.

Our localization tool comes from Manifold Topology Divergence. Its Cross-Barcode(P, Q) collapses distances inside the reference cloud Q, so that the surviving features belong to P and are not represented in Q. On CIFAR-10 mode dropping it tracked disturbance with Kendall-tau 0.89, against 0.36 for Geometry Score. Geometry Score compares hole statistics without regard to position, which is why we do not use it here: a user drifting into a new region with the same shape would be invisible to it. We also run The Shape of Data as a spectral, coordinate-free detector. Its heat-kernel distance is cheap but gives one global number, with no per-cluster localization.

On the preference side, Right Now, Wrong Then introduces NS-DPO. It discounts preference pairs by $\gamma^{T-t_i-1}$ and keeps accuracy above 60–70% under preference flips, where static DPO falls to about 50–55%. It concedes that the right $\gamma$ depends on an unknown budget of change. Its decay is global, and our method is a per-cluster replacement for it.

Beyond Markovian Drifts already argues for memory that suppresses interests using negative actions, through separate click and skip memories in Walk2Pers. Its memory is a pointwise walk in embedding space, with no notion of which interest cluster a step belongs to. We borrow its insight that skips carry information that decay ignores.

KuaiRand supplies the key experimental resource. Because some videos were shown at random, user responses to them estimate preference without the bias of what the deployed policy chose to show. This is what lets us tell "unvisited" from "rejected" honestly. That bias is exactly why The Topological Face of Recommendation could only audit graphs crawled from a single platform. That work used discrete link graphs and static new-versus-returning snapshots, not drift over continuous embeddings.

Research Program

For each user at time $T$, we embed their engaged items and form two point clouds. The first is the recent window $R$, made of the last 50–100 positive interactions. The second is the history $H$, made of everything before $R$. Both are projected to about 30 dimensions with PCA.

We compute Cross-Barcode($H$, $R$) on connected components. Long bars mark clusters of $H$ that the recent window does not cover, which we call quiet clusters. We also compute Cross-Barcode($R$, $H$), whose long bars mark newly opened interests. We read each long bar back to the history items it contains, so each quiet cluster $C$ becomes an explicit set of past items.

We then classify each quiet cluster $C$ with a rejection score $\rho_C$. This score is the fraction of recent exposures falling near $C$ that drew negative feedback: skips, short views, or dislikes, all of which are among the twelve KuaiRand signals. When there were no recent exposures near $C$, $\rho_C$ is undefined, and the cluster is treated as dormant. Each history item $i$ in cluster $C$ then receives a training weight

$$w_i = \gamma^{\,\rho_C (T - t_i)},$$

where $t_i$ is the item's timestamp and $\gamma<1$ is a single decay base. Rejected clusters ($\rho_C \to 1$) therefore decay at the full rate. Dormant clusters ($\rho_C$ undefined, which we set to 0) and still-active clusters keep weight near 1. Items in newly opened clusters receive a fixed upweight.

These weights multiply the per-item loss of a sequential recommender during training and, when we fine-tune, during refresh. The recommenders are SASRec, a causal self-attention model, and BSARec, which adds a frequency-domain filter. The weights also multiply the per-pair loss in a DPO variant. Per-user cost is a few Ripser calls on clouds of roughly a hundred points each, which is in the tens-of-milliseconds range reported for single windows.

The program has three stages. The first is a measurement study on KuaiRand, asking what drift looks like in real users. We label drift events using only randomly exposed items: a drift event in an embedding region is a significant change in a user's response rate to random exposures there. Each event is then typed as cluster death, cluster birth, dormancy-then-revival, or reweighting within stable clusters. The second stage compares detectors on those labels. The third trains recommenders with each weighting scheme.

Baselines

The naive reference point is uniform weighting: standard SASRec and BSARec trained on full histories. On MovieLens-1M, the standardized replication A Systematic Replicability and Comparative Study of BSARec and SASRec reports BSARec at NDCG@10 of 0.07704, 5.7% above SASRec. On Foursquare-NYC it reports BSARec at NDCG@5 of 0.2668 against 0.2334 for SASRec. We reproduce these numbers in the same EasyRec-style setup before adding any weighting. No reported result was found on Lacuna for sequential models on KuaiRand, so we establish those baselines ourselves.

The strongest existing competitor for history weighting is global exponential decay in the NS-DPO style, together with sliding windows. For recommenders, this means per-item loss weights $\gamma^{T-t_i}$. For preference optimization, it means NS-DPO itself, which reports above 60–70% accuracy under flips and a length-controlled win rate of 8.93 against 7.29 for DPO on UltraFeedback-LM Right Now, Wrong Then. It is the strongest competitor because it is what our method must beat on its own ground: it is cheap and tuning-light, and it already captures most recency effects. For fairness, $\gamma$ and window length receive the same grid-search budget as our single $\gamma$ and upweight constant, on the same validation users, over the same seeds and the same backbone. Beating it means a statistically significant NDCG gain on the full population, and a larger gain on the users whose history contains dormant-then-revived clusters.

For the detection study, the classical baselines are MMD-RBF and covariance shift, as run in Persistent Homology for Distribution Drift Detection in LLM Embedding Streams, plus that paper's own $H_0$ Wasserstein detector and the spectral distance of The Shape of Data. All run on identical per-user windows and PCA settings.

Two ablations isolate the proposal's ingredients. The first keeps the feedback-gated rule but finds clusters with k-means, with $k$ tuned, or with a density clusterer instead of cross-barcodes. This tests whether topology adds anything beyond "use clusters." The second keeps cross-barcode clusters but drops the feedback gate, so that every quiet cluster decays. This tests whether telling rejected clusters from unvisited ones is what matters.

Why not just…?

Why not just use exponential decay with a well-tuned $\gamma$? Decay is our main baseline and may win on stable datasets like MovieLens-1M; we expect parity there and would call that a success, not a failure. Its known weakness is structural. One $\gamma$ applies to every interest, and NS-DPO's own analysis ties the right $\gamma$ to an unknown variation budget. The decisive evidence is the dormancy-revival subset of KuaiRand users. Decay must under-predict revived interests there, and our rule should not. If decay matches us on that subset, the component-level hypothesis is wrong.

Why not just run MMD or a covariance test per user? These tests are simpler, and they already win on reweighting drift (AUC 0.740 for covariance against 0.641 for topology in the embedding-stream study). Their limitation for our purpose is that they return a yes/no alarm with no localization, and our weighting needs to know which items to forget. The measurement study reports how much real drift falls in the reweighting regime. If it is the majority, we will recommend covariance alarms with a clustering step, rather than topology.

Why not just cluster the history with k-means and decay per cluster? This is the most serious threat, since connected-component persistence is close to single-linkage clustering. The first ablation answers it directly. Our expected edge comes from two properties: cross-barcodes need no fixed $k$, which varies widely across users, and they measure what the recent window fails to cover rather than clustering history alone. If tuned k-means matches us, the honest conclusion is that feedback-gated cluster-level forgetting is the contribution and topology is replaceable. We would report it that way.

Why not just let a large language model read the history? AcCoRD finds frontier agents achieve perfect outcomes in fewer than 30% of dynamic-preference episodes, and scores as low as 0.09 on triggered preferences. Separately, Beyond Markovian Drifts beats 14B-parameter models by 0.22 PerSEval with a 170M-parameter model. Both suggest that long, drifting histories are not solved by context alone, so we do not prioritize this baseline.

Proposed Experiments and Decisive Checks

The measurement study uses KuaiRand-27K for scale and KuaiRand-1K for full-history analysis. Item embeddings come from a SASRec model trained on non-random exposures; as a check, we repeat the study with category-based features. We report the distribution of drift types, and detector AUC per drift type against labels derived from random exposures.

The recommendation study trains SASRec and BSARec under the six weighting schemes on KuaiRand, MovieLens-1M and Foursquare-NYC. We report NDCG@5/10 and Hit Rate, measured both on the standard next-item split and, for KuaiRand, on held-out random exposures, which removes the bias of what the old policy chose to show.

A secondary test applies cluster-gated pair weights inside NS-DPO on TV-HH (time-varying helpful-harmless). There the helpful and harmless preference regions form two natural clusters, and drift changes one of them, so gating should beat global discounting on reward accuracy.

The study has three decisive checks. First, our rule must beat tuned global decay on the dormancy-revival subset while matching it on MovieLens-1M, where little drift is expected. Second, the no-feedback ablation must lose on that same subset, which would show that the rejected-versus-unvisited distinction is doing the work. Third, the k-means ablation decides whether topology itself earns its place.