Recommender systems learn from two very different streams. Cheap actions such as page views, clicks and cart additions arrive in dense bursts. Purchases, the actions a platform actually cares about, arrive rarely. Both streams are also logged with timestamps that are often far coarser than the decisions they record. A user may view ten items and buy one in three minutes, and the log may stamp all eleven events with the same day. Sequence models then put those events in an arbitrary order, usually whatever order the logging pipeline or a sort produced. They read that order as meaningful, and they misjudge when the user's taste actually changed.
The idea of this note is to treat purchases as a slow, undersampled readout of a user's changing preferences, and the dense auxiliary actions as a faster sensor of the same underlying state. The model keeps a continuous-time preference state that moves between observation windows. Inside each coarse window it reads all events as an unordered collection rather than an invented sequence. The claim is testable on real data. Some multi-behavior logs carry fine timestamps, so we can coarsen them on purpose and check whether a model trained on coarse stamps recovers the accuracy and the timing of preference shifts visible at full resolution.
Core Claim
Our central hypothesis has two parts. First, current sequential recommenders lose accuracy as timestamps coarsen, and much of that loss is caused by arbitrary ordering of tied events rather than by lost information. We expect this to show up as large run-to-run variance when the within-window order is randomly shuffled, and possibly as inflated scores when ties happen to be sorted in a leaky way, for example cart before purchase.
Second, a model with two specific properties should be nearly flat across coarsening levels. One property is that it is permutation-invariant within a window and keeps co-occurrence counts. The other is that it lets dense auxiliary behaviors drive the timing of its drifting preference state. On fine-resolution data it should match strong time-aware sequence models, but it should keep its accuracy as resolution drops.
The contribution is therefore a resolution-controlled benchmark plus a model whose advantage is predicted to appear only where resolution mismatch bites. If the advantage does not appear there, the idea is wrong.
Why This Design, Grounded in Prior Work
The temporal-causality literature motivates the problem. Rate-Agnostic (Causal) Structure Learning shows that when a process is sampled more slowly than it acts, standard lag-based methods produce spurious "ghost" edges and miss real ones. Causal Discovery for Irregularly Time Series with Consistency Guarantees argues that binning and interpolation create fake correlations, and it lists same-step effects as outside its scope.
That same literature also constrains the design. iCITRIS: Causal Representation Learning for Instantaneous Temporal Effects proves that the direction of same-time effects cannot be recovered from passive observation without interventions or strong parametric assumptions. Recommender logs are passive. We therefore deliberately do not try to orient edges between events in the same window, which is the "obvious" intersection idea. Instead we model those events as exchangeable given a shared window-level state. This is the honest consequence of the iCITRIS result rather than a workaround.
Structural Hawkes Processes for Learning Causal Structure from Discrete-Time Event Sequences supplies a positive ingredient. It shows that co-occurrence statistics inside coarse bins still carry identifying information, so our within-window encoder keeps counts by behavior type and does not average them away.
On the recommendation side, the closest work splits the problem in incompatible ways:
- Learning Graph ODE for Continuous-Time Sequential Recommendation (GDERec) models preference drift in continuous time between interactions. It uses a single behavior stream, and its evolution step vanishes when events share a timestamp.
- Modeling Contemporaneous Basket Sequences with Twin Networks for Next-Item Recommendation (CBS) runs parallel click and purchase streams over baskets. It indexes time by basket number and ignores elapsed time and drift.
- Personalized Behavior-Aware Transformer for Multi-Behavior Sequential Recommendation (PBAT) uses auxiliary behaviors and temporal decay. It flattens everything into one ordered sequence, which is exactly the setting where tie order matters.
- Causal Disentangled Recommendation Against User Preference Shifts (CDR) shows that accounting for drift across time environments pays off, with Recall@10 gains reported to grow with shift intensity. It defines those environments by hand-chosen bins over a single behavior.
Our proposal differs from all four in the same respect. It uses the dense behavior stream to decide when the preference state moved, at a resolution finer than the purchase stream allows. It also refuses to invent order inside a window.
Research Program
The first component is a benchmark. We start from multi-behavior logs and, after verifying granularity, use those whose released timestamps are fine-grained, with Taobao the primary candidate. We then build versions coarsened to minute, hour, and day resolution. Each version comes with three tie-breaking protocols:
- the original file order,
- random shuffling with several seeds,
- a deliberately adversarial sort that places behavior types in their funnel order.
Datasets whose stamps are already coarse, such as Tmall and IJCAI-Contest if inspection confirms day-level stamps, serve as natural, uncontrolled test beds.
For drift, we compute reference shift points from the full-resolution data. These are change points in each user's category distribution of views, and they are used only for evaluation. The question is whether a model trained on coarse data can localize them.
The second component is the model. Let window $k$ contain the multiset $S_k$ of (item, behavior-type) events, and let $\Delta_k$ be the elapsed time since window $k-1$. The preference state updates as
$$ z_k = A(\Delta_k)\, z_{k-1} + g\big(S_k,\, z_{k-1}\big), $$
where:
- $A(\Delta)$ is a learned elapsed-time decay. We use a closed-form exponential per dimension, so no ODE solver is needed.
- $g$ is a permutation-invariant set encoder that pools event embeddings with behavior-specific counts and is conditioned on the previous state.
Purchase scores in the next window are read out from $z_k$. The auxiliary behaviors enter both as inputs to $g$ and as a dense self-supervised target: the model also predicts the next window's view and cart distribution from $z_k$. Because views are plentiful, this target supplies most of the gradient that tells the state when to move. The sparse purchase loss mainly shapes what the state means for conversion. To keep a sudden shift in one product area from contaminating stable interests, we optionally reuse CDR's sparse latent-to-category mapping.
The third component is analysis. We will measure three things: how each model's accuracy and seed variance change across the coarsening ladder, how well each model's internal state change aligns with reference shift points, and how gains split between sparse users and active users.
Baselines
The naive reference point is a popularity-plus-recency recommender: the user's most recent category's top items, or global popularity for cold users. It shows how much any sequence modeling buys at each resolution.
The strongest direct competitor is a time-aware self-attention recommender trained on a single chronological stream of all behaviors. It is a SASRec-style model with behavior-type and continuous time embeddings, and tied events receive identical time encodings. This is the reviewer's default answer. Done carefully, it may absorb most of the coarse-timestamp problem without any new machinery. PBAT is the published version of this design. The file reports relative gains on Taobao of 6.37% in HR@5 and 10.3% in NDCG@5, and on IJCAI-Contest of 8.20% HR@5 and 12.9% NDCG@5, over its baselines. No absolute Taobao HR@5 is recorded on Lacuna, so we will rerun PBAT and the time-aware transformer ourselves rather than compare against a quoted figure.
CBS represents the parallel-streams view. The file reports Recall@10 of 3.41% for CBS-CFN on Alibaba mobile shopping, against 1.94% to 2.39% for a single-sequence baseline. We will reproduce that setting as a sanity check before using CBS on our coarsened data.
GDERec represents continuous-time drift on a single behavior. Its reported advantage is an average 6.5% Recall@5 and 8.8% Recall@10 over SASRec and LightGCN on Amazon and MovieLens. CDR represents environment-based drift handling, with a reported 28.1% Recall@10 relative gain on Amazon Book. Are Graph Augmentations Necessary?: Simple Graph Contrastive Learning for Recommendation (SimGCL) represents the sparsity-only fix of embedding-noise contrastive regularization, applied to a static graph. Its reported Recall@20 gain on Amazon-Book is over 25% against LightGCN.
Fairness rules:
- All models see identical splits.
- Splits are chronological by global cutoff, following CDR's protocol, so that test windows truly follow training windows.
- All models get identical behavior inputs where they can accept them, the same embedding dimension, and the same hyperparameter search budget in GPU-hours.
- Results are averaged over five seeds and five tie-shuffles.
Ablations isolate the key ingredients. One replaces the set encoder with an order-dependent recurrent encoder over the arbitrary within-window order. Another removes the auxiliary-behavior prediction target, so that drift is driven by purchases alone. A third fixes $A$ to the identity, which removes elapsed-time drift.
Beating the time-aware transformer means one of two things, and either must hold without losing more than a small margin at full resolution:
- a statistically significant NDCG@10 advantage at hour and day resolution, or
- a markedly smaller degradation slope across the ladder, together with lower tie-shuffle variance.
Why not just…?
Why not just bin timestamps into fixed intervals or interpolate them? Binning is part of our ladder, so the experiment measures its cost directly rather than assuming it. The prior evidence is that fixed binning creates aggregation bias and interpolation creates spurious correlations, as argued in Causal Discovery for Irregularly Time Series with Consistency Guarantees and Rate-Agnostic (Causal) Structure Learning. Those papers address causal discovery rather than ranking accuracy, so the recommendation-side cost remains an empirical question we answer here.
Why not just use a Transformer with time encodings and give tied events equal timestamps? This is the real threat. Self-attention with tied time encodings is already nearly permutation-invariant within a window, apart from its positional encoding. If we remove positional encodings within ties and that model matches ours across the ladder, our architecture is unnecessary. The contribution would then shrink to the benchmark and the tie-order diagnosis, and we will report that outcome as such. We expect our model to differ mainly in drift timing. Attention has no explicit state that must move when dense view distributions shift, whereas our auxiliary prediction target forces such a state. The shift-localization metric is designed to expose that difference.
Why not just fix sparsity with contrastive regularization such as SimGCL? SimGCL spreads embeddings more uniformly on a static collapsed graph. It has no notion of time, so it cannot tell historical habit from a new interest. CDR reports that static graph baselines including LightGCN, SimGCL's backbone, lose accuracy under chronological shift. We include SimGCL so that any gain we report is not just a sparsity gain. Its contrastive loss can also be added to our model if it helps.
Why not just learn which within-window event caused which, as in instantaneous causal discovery? iCITRIS shows that this direction cannot be identified from passive data without interventions or strong assumptions, and recommender logs offer neither. Treating window events as exchangeable gives up a question that cannot be answered and keeps the co-occurrence information that remains usable. A drift-as-natural-experiment approach in the spirit of Temporally Disentangled Representation Learning remains a possible follow-up. It is not part of the central claim.
Proposed Experiments and Decisive Checks
The first check is the coarsening ladder on fine-timestamped Taobao. We report NDCG@10 and HR@10 for every model at every resolution, with mean and spread across tie-shuffles, and we separately report scores under the adversarial funnel-order sort.
We predict three things. Flattened-sequence baselines will show a downward slope as resolution coarsens, nontrivial shuffle variance, and suspicious gains under funnel sorting, which would expose leakage. Our model will show a flat curve with low shuffle variance by construction.
If flattened-sequence baselines show neither degradation nor variance, the premise that resolution mismatch matters for ranking is false at these scales, and we will say so.
The second check is drift timing. Models are trained at day resolution. We measure the time between each reference shift point, computed at full resolution, and the window where the model's state change peaks. For baselines we use the change in their final user embedding.
The third check is a sparse-user breakdown. We split by purchase count, mirroring the under-four-interactions analysis in Contrastive Meta Learning with Behavior Multiplicity for Recommendation. The question is whether the auxiliary drift target helps most where purchases are rarest.
Finally, we run the CDR chronological protocol on Amazon Book and Electronics. These datasets have a single behavior and coarse stamps, which tests whether the elapsed-time state and set encoding help even without dense auxiliary behaviors. We expect smaller gains there, and finding them would be the clearest sign that the dense-sensor mechanism, not extra capacity, drives the main results.
References
- Rate-Agnostic (Causal) Structure Learning
- Causal Discovery for Irregularly Time Series with Consistency Guarantees
- iCITRIS: Causal Representation Learning for Instantaneous Temporal Effects
- Structural Hawkes Processes for Learning Causal Structure from Discrete-Time Event Sequences
- Learning Graph ODE for Continuous-Time Sequential Recommendation
- Modeling Contemporaneous Basket Sequences with Twin Networks for Next-Item Recommendation
- Personalized Behavior-Aware Transformer for Multi-Behavior Sequential Recommendation
- Causal Disentangled Recommendation against User Preference Shifts
- Are Graph Augmentations Necessary?: Simple Graph Contrastive Learning for Recommendation
- Temporally Disentangled Representation Learning
- Contrastive Meta Learning with Behavior Multiplicity for Recommendation