Two Forgetting Clocks: Time-Driven versus Interference-Driven Decay in Parallel Sequential Recommenders

Recommenders that read a user's history in order have to decide how quickly to forget. Almost every current answer uses one clock: an interaction loses weight as time passes, either measured in steps or in wall-clock time. That rule breaks for a common kind of user. Someone who buys running gear every spring, or returns to a favourite artist after months away, has not lost that preference. It has gone quiet. A single recency clock fades it anyway, so when the user comes back the model has to relearn the interest from scratch.

This proposal separates two reasons for forgetting inside a fast, parallelizable recurrent recommender. The first is time-driven decay, where an interest fades simply because real time has passed. The second is interference-driven decay, where an interest fades because the user keeps acting on things that compete with it. The model learns, for each interest channel, how much of each clock to apply. The claim is that interest channels which hold periodic or dormant preferences will learn mostly interference decay and almost no time decay, and that this matters most on the events where users return to a long-dormant interest. The model also keeps the logarithmic-time parallel training and constant-time streaming updates of linear recurrent recommenders.

Core Claim

Linear recurrent units are recurrences whose state update is linear in the previous state, so a whole sequence can be computed with a parallel prefix scan in $O(\log L)$ steps for a sequence of length $L$. They have removed the recurrence bottleneck from sequential recommendation. Linear Recurrent Units for Sequential Recommendation (LRURec) reports reaching peak performance in about 23k steps versus over 50k for SASRec, and 7.3× higher incremental-inference throughput on XLong. Its decay, however, is a fixed per-dimension base applied once per step. It therefore ignores real time intervals and applies equally to every interest. Behavior-Dependent Linear Recurrent Units for Efficient Sequential Recommendation (RecBLR) makes the decay depend on the current item and reports gains of 7.46% HR@10 and 7.28% NDCG@10 on MovieLens-1M over its strongest baseline. Its gate is unstructured, though, and still counts steps rather than time.

Continuous-time methods take physical intervals seriously but embed them in quadratic attention. Self-attention with Functional Time Representation Learning does this with kernel-derived time features, and Learning Self-Modulating Attention in Continuous Time Space with Applications to Sequential Recommendation does it with point-process intensities. None of these methods separates why something should be forgotten.

We propose a channelled linear recurrence. Each state channel $c$ has its own update:

$$h_{t,c} = a_{t,c}\, h_{t-1,c} + p_{t,c}\, B x_t, \qquad a_{t,c} = \exp\!\big(-\lambda_c\,\Delta t_t \;-\; \mu_c\,(1-p_{t,c})\big).$$

Here $x_t$ is the embedding of the $t$-th interaction and $B$ is an input projection. The term $p_{t,c}\in[0,1]$ is a soft routing weight saying how much event $t$ belongs to interest channel $c$; it is computed from $x_t$ alone, so it adds no sequential dependency. The quantity $\Delta t_t$ is the real time elapsed since the previous event. The parameter $\lambda_c \ge 0$ is the channel's time-decay rate, and $\mu_c \ge 0$ is its interference-decay rate.

A channel with small $\lambda_c$ and positive $\mu_c$ forgets only when the user spends events elsewhere. Across a long quiet gap it barely decays. A channel with large $\lambda_c$ behaves like classic recency. Because $a_{t,c}$ depends only on the current input and $\Delta t_t$, the recurrence is still diagonal and input-dependent, so it can be computed with an associative scan and has an $O(1)$ streaming update.

Trainable exponential decays are known to destabilize training. Disentangle to Decay: Linear Attention with Trainable Decay Factor (D2D) addresses this by splitting the decay into a fixed base and a bounded trainable factor. It reports that this lowers PG19 extrapolation perplexity from 40.52 to 29.76. We reuse that split for $\lambda_c$ and $\mu_c$: each rate is a fixed multi-scale base times a bounded learned multiplier.

What is new is not "add $\Delta t$ to an LRU," which is a natural but small step. The new part is the structural hypothesis that forgetting should be split into a time component and an interference component, per interest. We test it directly on the events where the two hypotheses disagree, which are returns to dormant interests.

Research Program

The work has three parts.

Build the model. We implement the channelled recurrence on top of the RecBLR and LRURec codebases, keeping their scan kernels, embedding sizes and prediction heads. That way, any difference comes from the decay parameterization rather than from engineering.

Build return-event evaluation slices. These are carved out of standard benchmarks. A test event is a return if its item's category, genre or artist has been absent from the user's history for more than a threshold. The threshold is defined both in elapsed time and in number of intervening events, so that the two clocks can be separated. The remaining events form a continuation slice. The data come from Amazon subsets (with item categories), MovieLens-1M (genres), LastFM (artists, with periodic listening), and Tmall (timestamped clicks).

Build a controlled simulator. In the simulator, synthetic users have a known set of latent interests, known dormancy periods, and known drift. We generate two kinds of forgetting as ground truth: truly time-driven fading and truly interference-driven fading. This lets us check that the learned $(\lambda_c, \mu_c)$ recover the generating mechanism, which offline logs cannot confirm. The file notes the same problem for drift evaluation: passive logs rarely record the user's real motivation. AcCoRD: Evaluating User-Agent Collaboration Under Realistic User Preference Dynamics motivates this part. It shows that "triggered preferences", which surface late in an interaction, drive agent satisfaction scores down to 0.09, whereas explicitly underspecified preferences stay at 0.7–0.9. That is the conversational version of a dormant interest resurfacing.

Baselines

Naive references. The first is SASRec from Self-Attentive Sequential Recommendation, with its standard positional encodings. The second is the same model given only the last $K$ events, with $K$ tuned. Beating these confirms that the proposed model is competent. It does not show that the hypothesis is right.

Strongest parallel competitor: RecBLR. It shares our efficiency profile and already has input-dependent forgetting. Its reported result is a 7.46% HR@10 and 7.28% NDCG@10 improvement over the strongest baseline on MovieLens-1M. The file gives only relative improvements, not absolute scores, so we will rerun it ourselves. We also give RecBLR the strongest fair upgrade: $\Delta t$, encoded with Mercer time features, fed into its gates. We call this RecBLR+Δt. It is the model a skeptic would say already subsumes ours.

Strongest time-aware attention competitors. These are MOJITO from Attention Mixtures for Time-Aware Sequential Recommendation and MEANTIME from MEANTIME: Mixture of Attention Mechanisms with Multi-temporal Embeddings for Sequential Recommendation. MOJITO reports HR 82.09% on MovieLens-1M and HR@10 83.26% / NDCG@10 59.94% on Amazon Book. MEANTIME reports an average improvement of 9.07% Recall@5 and 10.43% NDCG@5. These numbers come from sampled-candidate protocols that are not comparable across papers, so they only tell us the expected range. We will rerun both models under a single protocol. The self-modulating attention layer, which reports HR@10 of 0.51347 versus 0.50433 for SASRec on Tmall, is included on Tmall.

Ablations that isolate the key ingredient. We run the model with $\mu_c$ fixed to zero (time-only decay), with $\lambda_c$ fixed to zero (interference-only decay), and with routing $p_{t,c}$ replaced by a uniform $1/C$ (which removes the channel structure). We also run it without the D2D-style fixed base, to check the stability claim.

Fairness rules. Every model gets the same splits, the same full-catalog ranking protocol (we avoid sampled negatives, which inflate numbers like those above), the same embedding dimension, and the same number of hyperparameter trials. We use equal wall-clock GPU budgets for training, and three or more seeds.

What counts as beating the baselines. The proposed model must be statistically ahead of RecBLR+Δt on NDCG@10 on the return slice, with a margin larger than the seed variance. It must be within noise of it on the continuation slice. It must also stay within 1.5× RecBLR's training throughput.

Why not just…?

Why not just feed $\Delta t$ into RecBLR's gates and let them learn whatever forgetting they need? This is the real threat. An unstructured, input-dependent gate can in principle express interference-style forgetting. What it lacks is any pressure to do so. The training signal comes mostly from continuation events, and on those, recency decay is nearly optimal. RecBLR+Δt is our headline baseline for exactly this reason. If it matches us on the return slice, the structural hypothesis adds nothing, and we will report that as the finding. The simulator settles the question more cleanly: we will check whether RecBLR+Δt's gate values over a dormant gap stay near 1, as interference-driven ground truth requires.

Why not just retrieve past items from the candidate's category at scoring time? Category-conditioned retrieval over the full history handles returns directly, with no decay modelling at all. Its cost is a lookup over history per candidate, and it discards the order in which interests were abandoned. We include it as a strong return-slice baseline: SASRec plus category-matched target attention over the full history. If retrieval matches us on returns at similar latency, the recurrence contributes only efficiency. That is still useful to know, and we would narrow the claim to "retrieval-quality returns at $O(1)$ streaming cost."

Why not just truncate history or discount old training examples? Truncation removes exactly the dormant interests at issue. TLSAN: Time-aware Long- and Short-term Attention Network for Next-item Recommendation finds that dropping its long-term stream hurts AUC; with both streams it reaches 0.9230 on Amazon Electronics. Discounting the training loss, in the style of Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift, improves LC win rate from 7.29 to 8.93 under preference flips. It uses one global discount, though, so it cannot hold on to one interest while forgetting another. We include a globally loss-discounted LRURec to measure this directly, and we expect it to help continuation events and hurt returns.

Proposed Experiments

Main comparison. We report full-ranking HR@10 and NDCG@10 on Amazon Beauty, Sports and Games, on MovieLens-1M, LastFM and Tmall. Each result is split into return and continuation slices, at several dormancy thresholds.

Separating the two clocks. The decisive check sorts return events into two groups: long time gaps with few intervening events, and short time gaps with many intervening events. If forgetting is interference-driven, the first group should be recovered well and the second poorly. Pure time decay predicts the opposite. We compare how accuracy moves across these groups for the proposed model, the time-only ablation and RecBLR+Δt. We also inspect the learned $(\lambda_c,\mu_c)$. On LastFM we expect interference-dominant channels, where listening is periodic. On Tmall we expect time-dominant channels, where intent changes session to session. If every channel learns $\mu_c \approx 0$ on real data while the simulator recovers the ground truth, the hypothesis is refuted for these domains.

Efficiency. We measure training throughput and per-event streaming latency against LRURec and SASRec on XLong, with sequences of up to 800 interactions. The model should keep the scan advantage that LRURec reports.

Stability. We track gradient norms and divergence rates with and without the fixed-base split, since D2D identifies trainable decay as the main source of failure.