When a friend shares a post, the chance that you reshare it depends on what was shared, who shared it, and how long ago. Deep sequence models predict these reshares well, but the way they weigh time is buried in dense attention matrices that nobody can check. Classical influence models are the opposite. They write the effect of time as an explicit decay curve, such as "influence halves every hour," which an analyst can read and verify against data, but they cannot use the content of posts. This note proposes a model that keeps the explicit decay curve while letting a neural network handle content. Each past exposure is first mapped to a few human-readable concepts, for example "came from a mutual follower" or "matches the user's topic." Each concept then contributes through its own small, shape-constrained curve over elapsed time. The prediction is a transparent sum of these contributions plus a term for how many times the user has already been exposed.
The point is not only to build such a model. It is to measure, on real cascade data, exactly how much accuracy is lost by forcing time to pass through a few readable curves instead of free attention. The note also asks whether that loss comes from the concept layer, from the curve shapes, or from assuming exposures add up. That question is open. Video work found a modest cost for constraining attention, while social-diffusion work suggests the true temporal patterns are low-dimensional enough that the cost could be near zero. Either answer is useful: it tells practitioners whether auditable influence models are a sacrifice or a free lunch.
Core Claim
The claim is that on social diffusion streams, almost all of the predictive value of unconstrained temporal attention is recovered by three things together: a concept bottleneck (an intermediate layer whose units are named, human-meaningful quantities, as in Concept Bottleneck Models), one parametric decay kernel per concept, and an explicit non-additive exposure-count term. The count term is the non-obvious ingredient.
Cumulative Effect in Information Diffusion reports three facts about Sina Weibo. Forwarding probability peaks at two exposures and then declines. Forwarding latency is log-normal with a peak at 10 minutes. Users who see a post twice echo the more recent source 85.5% of the time. Any model that treats exposures as purely additive excitations, which is the standard Hawkes-style assumption, increases monotonically with exposure count and therefore cannot reproduce the decline after two exposures. The prediction is that a kernel bottleneck without the count term will show a real accuracy gap to attention, and adding the count term will close most of it. That decomposition is the testable core.
Concretely, for target user $u$ at time $t$ with past exposures $e$ at times $t_e$, the adoption logit is
$$ s_u(t) = \sum_{k=1}^{K} w_k \sum_{e} c_k(e)\, \kappa_k(t - t_e;\theta_k) + h(n_u(t)), $$
where the terms are as follows. $c_k(e)\in[0,1]$ is the activation of concept $k$ for exposure $e$, produced by a neural encoder over the exposure's content and the graph around it. $\kappa_k$ is a decay kernel from a small family: exponential, power-law, and log-normal hump, each with an optional plateau at a learned stabilization horizon. $\theta_k$ holds its two or three parameters. $w_k$ is a scalar weight per concept. $h$ is a monotone-free piecewise function of the exposure count $n_u(t)$ with one value per count bin.
Every quantity that touches time is a curve one can plot and compare against measured latency distributions. Every quantity that touches content is a named concept one can inspect or overwrite.
Relation to Prior Work
Three papers motivate the gap. Each pairs temporal decay with explanation but leaves the model's internals opaque.
Explanations for Temporal Recommendations adds an $\exp(-\alpha t)$ neighbor-explainability regularizer to a recurrent recommender. This slightly improves Netflix RMSE from 0.922 to 0.919, but the authors themselves say the recurrent network's internals remain uninterpretable. We move decay into the forward pass rather than the loss.
Extracting Attentive Social Temporal Excitation for Sequential Recommendation (STEN) models friends' events with Hawkes-inspired mutual and self-excitation attention over dense embeddings. It gives case-level justifications but has no named intermediate layer and no constraint on the shape of its time weighting. It is our strongest accuracy competitor, and our design is in effect STEN with its time weighting forced through readable kernels.
NFTracer criticizes ML impact-dynamics models as incomprehensible and answers with visual analytics. That approach is post hoc rather than a predictive architecture.
Two papers supply the method. Concept Bottleneck Models gives the bottleneck structure and the test-time intervention protocol. On knee osteoarthritis grading, correcting two concepts drops task error from over 0.4 to about 0.3. Concepts in Motion (MoTIF) is the closest architectural relative. It extends bottlenecks to video with per-concept temporal attention over fixed windows and reports a modest accuracy cost relative to fully entangled attention, for example 40.4% on Something-Something V2. Our work differs from MoTIF in three ways. Events here are asynchronous rather than uniformly windowed. Each concept's time weighting is a parametric curve rather than free attention. And we keep a cross-exposure count interaction, which MoTIF's diagonal design forbids.
The decay literature supplies the kernel families. Influence Propagation in Social Networks: A Data Mining Perspective and Herding Effect based Attention supply exponential decay. Temporal scaling in information propagation supplies the power law. StructInf reports that structural influence stabilizes after about 25 hours on Weibo, which motivates the plateau option. Fast Budgeted Influence Maximization over Multi-Action Event Logs supplies a harmonic-mean effective delay that rewards repeated exposure.
Cumulative Effect in Information Diffusion both motivates and constrains the design: it rules out a purely additive kernel model. Graph Concept Bottleneck Models, which shows that letting concepts interact through a learned graph gains 1–2% on static images, suggests a natural extension. We deliberately do not adopt it in the main model, because concept-to-concept message passing would reintroduce the entanglement we are trying to measure.
Research Program
The work has three parts.
The first part builds the model. The concept encoder produces roughly ten to twenty concepts per exposure. Some are structural and computed directly from the graph: mutual follow, whether the source closes a triad with the target, and the source's follower-count bucket. Others are content-based and predicted by a small text encoder supervised on weak labels: topical match to the target's history, the presence of a URL or media, and stance agreement where stance labels exist. Kernels are selected per concept by a likelihood-based choice among the three families. We train jointly, with sequential training as in Koh et al. as a check.
The second part is a gap decomposition. Starting from an unconstrained attention model over exposures, we remove freedom one step at a time and record held-out accuracy after each step. First we replace dense embeddings with concepts. Then we replace free attention with per-concept nonparametric time bins. Then we replace the bins with parametric kernels. Finally we remove the count term. Each step's accuracy drop is one named component of the price of interpretability.
The third part tests faithfulness and usefulness. It checks whether the learned curves match independently measured temporal statistics, whether they recover known kernels on synthetic cascades, and whether overwriting a concept changes predictions in the expected direction. These checks are described with the experiments below.
Baselines
The naive reference is a recency-and-count logistic regression: adoption predicted from time since the most recent exposure and the number of exposures, with no content. It sets the floor that any learned temporal structure must clear.
The strongest interpretable non-neural competitor is a generalized additive model, meaning a model that adds up one learned curve per feature, fit over hand-built features. The features are exposure count, binned elapsed times, structural motifs as in StructInf, and source attributes. Its only reported result is StructInf's 1.73% F1 improvement for Weibo retweet prediction from structural patterns; no absolute F1 is reported on Lacuna. We fit it on the same splits and features available to our structural concepts, with the same hyperparameter search budget.
The strongest accuracy competitor is STEN. On Delicious it reports Recall@10 3.87% and MRR@10 5.74% higher than its strongest baseline, given as relative gains; no reported absolute result was found on Lacuna for STEN on Delicious, Yelp, or Ciao. We run it ourselves on Delicious and Weibo with the same exposure histories, the same embedding size as our concept encoder backbone, and an equal tuning budget in GPU-hours.
We also include a plain transformer over timestamped exposures with learned time encodings, which serves as the unconstrained ceiling. The TempEx-Fluid-style regularized recurrent model is included on Netflix, where its RMSE of 0.919 is the reference to match. Beating STEN is not required. The proposal succeeds if the full kernel bottleneck stays within one point of STEN's Recall@10 and the transformer's AUC while clearly beating the additive model.
The key ablation removes the count term $h$ while keeping everything else. If our core claim is right, this ablation loses substantially more accuracy than switching from nonparametric bins to parametric kernels.
Why not just…?
Why not just explain a black-box transformer with attention weights or SHAP? Attribution can say which past exposure mattered, but it cannot guarantee that moving that exposure an hour earlier changes the prediction the way the explanation implies. Post-hoc influence estimates in deep networks are known to be unstable, as shown in Influence Functions in Deep Learning Are Fragile. We test this directly. We shift exposure timestamps in held-out data and compare each method's predicted change in its explanation against the actual change in model output. In our model the two agree by construction. For attention and SHAP the agreement must be measured, and we report it.
Why not just fit a multivariate Hawkes process? Hawkes processes have readable kernels but cannot use post content. They are also additive, so they cannot produce the two-exposure peak. Our model can be read as a Hawkes process whose excitation is gated by learned concepts and corrected by a count term. Including a plain Hawkes fit with the same kernel families as a baseline isolates exactly what the concept encoder buys. If the Hawkes fit matches our model, the content concepts are unnecessary, and we would say so.
Why not just fit an additive model over hand-crafted temporal and structural features? This is the most serious threat. If the Weibo signal is mostly structural and temporal, a well-engineered additive model with an exposure-count feature may match us. The gap decomposition settles it. If our structural-concepts-only variant matches the full model, the content encoder adds nothing and the honest conclusion is that hand-built additive models suffice. The proposal is designed so that this outcome is a reported finding rather than a failure.
Proposed Experiments
The primary benchmark is the Sina Weibo WISE 2012 cascade data used by Bao et al., for next-hour reshare prediction. We use time-based train and test splits and report AUC and F1. The StructInf Weibo stream serves as a second retweet task. Delicious, Yelp, and Ciao provide sequential-recommendation comparisons to STEN, scored by Recall@10 and MRR@10. The Twitter multi-action event log from Yu et al. tests repeated exposure. Twitter15 and Twitter16 supply content-heavy cascades where stance and rumor concepts matter.
Four checks are decisive.
The first is the gap decomposition. We report accuracy at each step of removing freedom, with confidence intervals over five seeds. The claim is falsified if parametric kernels lose more than two AUC points to nonparametric bins, or if the count term contributes nothing.
The second is external validity of the learned curves. The kernel learned on Weibo should place its latency peak near the 10-minute log-normal mode reported by Bao et al. Its plateau should fall near the 25-hour stabilization horizon StructInf reports. Neither statistic is used in training.
The third is synthetic recovery. We generate cascades with known concept-specific kernels and a known peak at two exposures. We measure kernel-parameter error for our model, and we compare how well curves extracted from the transformer's attention, and from STEN's, recover the same kernels.
The fourth is intervention. On Twitter15 and Twitter16, we replace predicted content concepts with human or metadata labels on a held-out subset, following the Koh et al. protocol. We report how prediction error falls as the number of corrected concepts grows, with STEN given the closest analogue: replacing an exposure's embedding with one computed from correct labels.
References
- Concept Bottleneck Models
- Cumulative Effect in Information Diffusion: Empirical Study on a Microblogging Network
- Explanations for Temporal Recommendations
- Extracting Attentive Social Temporal Excitation for Sequential Recommendation
- NFTracer: Tracing NFT Impact Dynamics in Transaction-flow Substitutive Systems with Visual Analytics
- Concepts in Motion: Temporal Concept Bottleneck Model for Interpretable Video Classification
- Influence Propagation in Social Networks: A Data Mining Perspective
- Herding Effect Based Attention for Personalized Time-Sync Video Recommendation
- Temporal scaling in information propagation
- StructInf: Mining Structural Influence from Social Streams
- Fast Budgeted Influence Maximization Over Multi-Action Event Logs
- Graph Concept Bottleneck Models
- Influence Functions in Deep Learning Are Fragile