Clicks as Noisy Sensors: A Kalman-Filtered Preference Trajectory for Sparse, Drifting Users

Most recommenders learn one fixed vector per user, or read a user's history as an ordered list of positions. Both choices lose information that matters when data are scarce. Users change what they want over time, and the time between their actions is uneven. A single user's purchases are also far too few to track that change. We propose to treat a user's taste as a smooth curve in real calendar time. Every interaction, whether a view, a click, a cart addition or a purchase, is a noisy reading of where that curve is at the moment it happens. Cheap, frequent behaviors are loud but unreliable readings, and purchases are rare but precise. A temporal smoothness prior, meaning a prior belief that taste at nearby times is similar, is what lets a click on Tuesday tell us about purchase intent on Thursday.

Stated in two sentences: model each user's preference as a sum of a slow-moving and a fast-moving random process in continuous time, and fuse all behavior types into it with a learned noise level per behavior. Because the processes are Markovian, exact Bayesian updating is a cheap Kalman filter. This yields preference estimates that automatically fall back to long-term taste after long absences and that weight new evidence by how much recent evidence already exists.

Core Claim

In the sparse regime, the useful part of a temporal prior is not smoothing for its own sake. Its value lies in two properties that current sequential recommenders do not have together.

The first is that the prior supplies a principled weighting of evidence. How much a new interaction moves the estimate depends on its behavior type, the elapsed time, and the density of surrounding observations. The Kalman gain computes exactly this.

The second is that the prior separates timescales. After a long gap, the estimate of the fast component decays toward zero while the slow component persists, so the prediction reverts to long-term taste instead of either freezing on stale session intent or forgetting the user.

We claim this probabilistic multi-behavior fusion in real time will beat deterministic continuous-time models specifically on two groups: users with few target actions, and predictions made after long inactivity gaps. We do not expect a gain on dense, regularly active users, and we will report that honestly.

Relation to Prior Work

Three papers motivate the design by showing that temporal structure helps.

FMLP-Rec replaces self-attention with learnable frequency filters over sequence positions. Its learned filters turn out to be low-pass, which is direct evidence that smooth preference signals dominate noisy logs. However, it works on integer positions, is deterministic, and ignores behavior types.

GP-VAE puts a Gaussian-process prior (a distribution over smooth functions) on a variational autoencoder's latent space to interpolate across gaps in time series. Two of its documented limits constrain us. It relies on an amortized encoder, which introduces an "amortization gap" between the encoder's output and the true posterior. It also needs banded precision matrices for tractability.

Our design avoids both limits. We use Ornstein–Uhlenbeck processes, a Matérn-1/2 Gaussian process whose Markov structure makes exact filtering linear in sequence length, with no encoder approximation.

Three continuous-time recommenders are the closest existing work, and each one supplies part of the method.

SS4Rec sets a state-space model's discretization step from the real elapsed time and is robust to randomly removing 10% of interactions. Its state update is deterministic and heuristic, and it has a single behavior type.

GDERec evolves user and item states between interactions with a graph neural ODE (a learned differential equation over the interaction graph). It reports average gains of 6.5% Recall@5 over SASRec and LightGCN. It is also deterministic, with no uncertainty and a single edge type.

Self-Modulating Attention silences past items whose point-process intensity has decayed. Its authors report that its timestamp regularizer is fragile on sparse Amazon Electronics, which warns us that time-likelihood terms can hurt under sparsity. We therefore make our marginal-likelihood term optional and ablate it.

Our model is close to SS4Rec in form, since both run a linear continuous-time state update. The differences are the ones our claim rests on: a posterior variance that sets the update gain, explicit slow and fast components, and per-behavior observation noise.

On the multi-behavior side, PBAT is the closest competitor. It represents items and behaviors as Gaussians and uses Wasserstein-distance attention with temporal decay. Its uncertainty sits on static embeddings, and time enters as a decay term inside attention, not as a stochastic process over the user's state.

CML and S-MBRec learn behavior weights through contrastive objectives on static graphs, discarding timestamps. Our per-behavior noise variance plays the role of their learned behavior weights, but it is interpretable and interacts with time: a click one hour ago can outweigh a purchase six months ago.

CDSRNP also treats preferences as a stochastic process. Its process runs across users and domains for transfer, not across a user's own timeline.

FlowRec and PreferDiff speak of "trajectories", but their trajectories live in the generator's internal integration time, not calendar time. They are orthogonal to us and could consume our filtered state as their conditioning prior.

Research Program

Each user $u$ has a latent preference $z_u(t) \in \mathbb{R}^d$, where $t$ is calendar time. We write it as $z_u(t) = s_u(t) + f_u(t)$, a slow and a fast component, each an independent Ornstein–Uhlenbeck process per dimension with learned lengthscales $\ell_s \gg \ell_f$ and variances. The slow process has a learned per-user mean initialized from a population prior, so that new users start at the population average. An interaction with item $i$ of behavior type $b$ at time $t_k$ is modelled as

$$ e_i = z_u(t_k) + \epsilon_k, \qquad \epsilon_k \sim \mathcal{N}(0, \sigma_b^2 I), $$

where $e_i$ is the item's learned embedding, possibly produced by any item encoder, and $\sigma_b$ is a learned noise scale for behavior $b$. Because everything is linear-Gaussian and Markov, the posterior over $z_u$ at any query time comes from a Kalman filter run over the user's events. The cost is $O(Td)$ per user, where $T$ is the number of events. The filter is differentiable end to end.

To score the next target action at query time $t^\ast$, the model propagates the filtered posterior to $t^\ast$ and ranks items by $e_j^\top \mu_u(t^\ast)$, where $\mu_u$ is the posterior mean. It is trained with sampled softmax on held-out next purchases, optionally adding the Gaussian marginal likelihood of all events as an auxiliary term.

The posterior variance does not change a linear ranking. Its role is to set the Kalman gain, and it is also reported for calibration and gap analysis.

We will build the model in two forms. The first is a standalone model. The second is a drop-in "filter head" that replaces the final state of SASRec or SS4Rec, which tests whether the benefit survives on top of a strong sequence encoder.

We will also test one extension aimed at informative missingness. The length of a gap is itself evidence: a user who stops buying may have lost interest, which is exactly the "missing not at random" situation that GP-VAE's assumptions exclude. In the extension, the fast component's lengthscale is made a function of recent event rate.

Baselines

The naive reference points are popularity ranking and a static per-user embedding (matrix factorization on purchases plus auxiliary behaviors). They show how much any sequence or time modeling buys at each sparsity level.

The strongest single-behavior continuous-time competitor is SS4Rec SS4Rec. We will run it with a behavior-type embedding added to each input token, so that it sees exactly the information our model sees. The file reports state-of-the-art HR@10 for it on five datasets but gives no absolute value, so no reported result found on Lacuna for SS4Rec's HR@10 on our target datasets.

GDERec GDERec is the deterministic continuous-time graph alternative. Its reported figure is relative only: an average 6.5% Recall@5 gain over SASRec.

Self-Modulating Attention SMLayer reports HR@10 of 0.51347 on Tmall. This is the one absolute number in the file, and we will reproduce it on the same Tmall protocol before comparing.

For multi-behavior sequences, the strongest competitor is PBAT PBAT. It already combines Gaussian uncertainty, behavior types and temporal decay on Taobao, IJCAI and Yelp. The file reports only relative gains for it, including +6.37% HR@5, and no absolute values we can quote.

We add FMLP-Rec FMLP-Rec as the position-based smoothing baseline.

All methods get the same splits, the same item-embedding dimension, full-catalog ranking, and an equal budget of 30 hyperparameter trials each, selected on validation NDCG@10.

Four ablations isolate the key ingredients:

  1. Fixed decay. Replace the Kalman gain with a fixed exponential time decay using the same learned lengthscales.
  2. Shared noise. Tie all $\sigma_b$ to a single value.
  3. Single timescale. Remove the fast component.
  4. Positions instead of time. Replace timestamps with integer positions.

Beating PBAT and SS4Rec means a statistically significant gain in NDCG@10 over five seeds on the sparsest user tercile and on the longest-gap tercile, without a loss beyond noise on the densest tercile.

Why not just…?

Why not just add an exponential time-decay weight to past interactions? This is the most serious threat. After a gap, our posterior mean does decay the fast component much as a fixed decay would. The difference is that our gain depends on how dense the recent evidence is and on the behavior type, whereas a fixed decay weights the fifth click in an hour the same as a lone click. The fixed-decay ablation, run with identical lengthscales, tests exactly this. If it matches the full model on sparse users, we will conclude that the probabilistic machinery is unnecessary and report that.

Why not just give SS4Rec behavior-type tokens? SS4Rec already evolves a state in real time, and a learned selective update could learn behavior-dependent gains. We run this augmented SS4Rec as our main baseline, and we also put our filter on top of it as a head. If the head adds nothing on top of SS4Rec, the claim reduces to "a simpler, interpretable equivalent". That is still useful, but it is weaker, and we would say so.

Why not just use PBAT, which already has Gaussian uncertainty and behaviors? PBAT's uncertainty belongs to static item and behavior embeddings. It does not grow during a user's inactivity, and its time handling is a decay term inside attention. The gap-stratified evaluation below is where the two should separate. If PBAT is equally strong in long-gap cases, the continuous-time process is not paying for itself.

Proposed Experiments

We will use the multi-behavior datasets that PBAT evaluates on (Taobao, IJCAI and Yelp), Tmall as used by Self-Modulating Attention, and Amazon categories with timestamps for single-behavior sanity checks. The protocol is leave-last-purchase-out with full ranking, reporting HR@10 and NDCG@10. We evaluate both with the true query timestamp and with the query time set to the last observed event, so that the model is not credited for knowing when the user returns.

Results are stratified by the number of target actions per user and by the time since the last interaction. This stratification is where the claim lives.

We run four decisive checks:

  1. Positions instead of timestamps. If swapping real timestamps for integer positions causes no drop, the continuous-time claim fails.
  2. Learned noise scales. If the learned $\sigma_b$ do not order as purchases < carts < clicks < views, the sensor interpretation is wrong.
  3. Calibration after gaps. We check whether the posterior variance predicts ranking error after gaps, using rank correlation between predicted variance and reciprocal rank of the held-out item.
  4. Synthetic drift. In a controlled test, we inject preference switches into held-out Taobao users by relabelling their later items from a different category cluster. This measures how quickly each model recovers relative to how long it remembers pre-switch taste.