acceptodds
Under review as a conference paper at ICLR 2027

From Prospective Actions to Events: Lifecycle Marked Temporal Point Processes

Abstract

A Marked Temporal Point Process (MTPP) is a stochastic process that models the timing and types of events in continuous time. MTPPs are widely used to characterize asynchronous event sequences in domains such as mobility, e-commerce, and online behavior. In many applications, event histories also contain prospective actions, such as destination searches or cart additions, that point to possible future targets but may never be realized. While existing models can represent these actions as part of event history, how an action's predictive status evolves as subsequent observations arrive has been under-explored. In this regime, we propose a prospective-action lifecycle MTPP (PAL-MTPP) that represents each prospective action as an evolving target-linked lifecycle state. PAL-MTPP updates these states using both subsequent events and no-event observations while preserving uncertainty about action realization. For events with multiple targets, we further introduce pairwise target interactions to refine target-set probabilities beyond independent target predictions. Experiments on real-world mobility and e-commerce datasets show that lifecycle-aware action modeling improves future-event prediction, while pairwise interactions provide additional gains for multi-target prediction.

Primary area:
learning on time series and dynamical systems
Keywords:
Marked Temporal Point Processes, Prospective Actions, Event Forecasting, Lifecycle Modeling, Set-Valued Prediction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_8UPtOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper focuses on extending the neural marked temporal point process (MTPP) in two aspects: (i) it also models observed auxiliary variables (referred to as "action" in the paper) in addition to continuous-time events, and (ii) it extends the categorical mark space (i.e., each event has a single "target") to "multi-target" events (i.e., each event is a set of multiple targets, so duplicates are not allowed). The high-level ideas are summarized below:

For (i), the marked intensity (for each "target") is computed by summing up (1) the base target-specific intensity computed from observed event history and (2) a function modeling the historical effects from the auxiliary actions; see Eq. 2 and Eq. 6.

For (ii), the probability of each multi-target set is factorized into the probability of set cardinality and the multi-target set probability conditioned on its cardinality (Eq. 10), with follow-up refinement (Eq. 11-14).

They performed experiments on three datasets: one with single-target events and two with multi-target events. Evaluation metrics include sequence negative log-likelihood (Eq. 8), target/set negative log-likelihood, and NDCG@10 (often used in recommendation systems).

Critical strengths
  1. This paper aims to address useful extensions of classical MTPPs by incorporating informative observed auxiliary action variables in real-world applications. For example, in e-commerce, the purchase event is related to other consumer behaviors, such as page views or adding to carts, that have not been explicitly modelled by prior MTPP work.
  2. While prior work looked into joint probabilistic modeling of event time and set-valued events (Chang et al., 2024), they only captured negative correlations between targets but not positive correlations; this work extends to modeling both positive and negative correlations between targets.
Critical weaknesses
  1. While the extensions are novel and well-motivated in their ideas, the methodology is not clearly explained, the terminology/notation is imprecise, and the modeling choices are not well-justified. For instance, Ht\mathcal{H}_t and Ht−\mathcal{H}_{t^-} are mixed and mm and vv are mixed in Sec 2 and Sec 3. It is unclear how the parameterization is optimized with training data, which architecture the paper uses to compute the logits lvl_v, why some design choices were made, or which component is conditioned on history and which is not; see "Questions For Authors" for more details.

  2. Experiments lack soundness or reproducibility with lots of essential details missing and no accompanying code for assessment. What hyperparameters did the experiments choose for each architecture (in addition to learning rate, batch sizes, and optimizer in Table 5 in Appendix)? Is Target NLL/Set NLL (or mark NLL) computed conditioned on the true event time or sampled event time (Eq. 14 that defines the final target-set distribution does not condition on time or history)? It is unclear how the standard neural MTPPs (NHP/THP/S2P2) are extended to multi-target events. Prior work on set-valued events (Chang et al., 2024) is more closely related to target-set modeling than NHP/THP/S2P2 but is not included in the experiments. How is NDCG@10 computed, since all prior work does not consider any ordering in targets? See "Questions For Authors" for more questions.

Questions for authors

Although I wish to support the novelty of this paper, the current manuscript needs some work to clarify technical details and justify its modeling choices. A lot more detail would be needed to assess the correctness and soundness of the experimental results.

Concrete questions that would be good to clarify in the next version of this manuscript and would impact the assessment of correctness and soundness:

Methodology:

  • Is the base intensity λvbase(t)\lambda_v^{\text{base}}(t) only a function of time, or is it also conditioned on event history?
  • As a core component, how does the binary lifecycle state Zi(t)Z_i(t) get updated between 0 and 1 when it sees more observations?
  • What is gig_i in Eq. 2, is it a function or a parameter? What does this imply about the effect of a previous action on future events over time?
  • How does an observed distribution QiQ_i look like (Sec 3.1)? In other words, how does such a distribution present in the training data?
  • What are the parameters in λvA(t)\lambda_v^A(t) and how to use the observations to optimize them?
  • What architecture does each component use?
  • Is the target-set distribution p0(V)p_0(V) (Eq. 9) conditioned on event time or event history?
  • How did the classifier for cardinality distribution p0(k)p_0(k) trained on observations?
  • How does sampling a target-set event work? First sample kk then VV conditioned on kk?
  • What does the sequence NLL (Eq. 8 for single-target events) for multi-target events look like if the set probability is normalized to non-empty sets (Eq. 9)?
  • How does the cross-target interaction modelled via Eq. 12 with diag(ω)\text{diag}(\omega)?
  • The inner product in Eq. 12 also depends on the magnitude of the vectors hu,hvh_u, h_v; how does this impact βuv\beta_{uv} that defines the interaction score? Why is there a factor of 2 when computing βuv\beta_{uv}?
  • Is Eq.15 the training objective for target-set events? Why would we want to refine qFq_F to be close to q0q_0 (which assumes conditional independence)?
  • If the interaction parameters are optimized using qFq_F in training, is qsafeq_{\text{safe}} only used in prediction, and is αk\alpha_k defined in Eq. 14 hyperparameters or learned from the training data?

Experiments:

  • What are the statistics of each dataset? I see Table 4 in the appendix, but it does not include the number of possible targets or time granularity, and what are the "Risk intervals" (that is not mentioned elsewhere)?
  • What are the hyperparameters for baseline models, and is a hyperparameter search performed on each dataset for each architecture?
  • How were the models trained on each dataset, and are the reported results on validation or test data?
  • How did the experiments adapt NHP/THP/S2P2 for multi-target learning and prediction? Do the target-specific logits use the hidden state values in neural MTPPs, and if lvl_v at tt or tn−1t_{n-1} was used for NHP/THP/S2P2?
  • Is Target NLL/Set NLL (or mark NLL) computed conditioned on the true event time or sampled event time?
  • In Sec 6.3 lifecycle update analysis, how were the probabilities and predictions computed on a continuous timeline (e.g., is any grid point used)? Could the authors elaborate on why realization status becomes more uncertain over time, and does this relate to gi(t−si)g_i(t-s_i) in Eq. 2?
  • What are the T1 - T8 in Fig. 5a? They do not make sense to readers in terms of what they imply in e-commerce interactions in the Synerise dataset.
  • What kind of shuffling was done in Fig. 5b, and what does the experiment intend to tell the readers? Shuffling the target-pair assignments does not sound like what the model is optimized for, so I would not be surprised if the NLL is worse. Why is it evaluated on validation likelihood?
  • In Appendix B.3, the paper suggests a synthetic experiment where the ground-truth intensity is available. Could the authors elaborate on whether the events are sampled from a parametric Hawkes process, how the actions are sampled, or the parameters of the ground-truth model used for simulation, and how the intensity RMSE is defined for a continuous function?

Minor issue that does not affect assessment:

  • Typo at the beginning of Sec 3.3.
10 comments
AuthorsOpenReview ↗

## W1 — Method clarity and implementation details

While the extensions are novel and well-motivated in their ideas, the methodology is not clearly explained, the terminology/notation is imprecise, and the modeling choices are not well-justified. For instance, Ht\mathcal{H}_t and Ht−\mathcal{H}_{t^-} are mixed and mm and vv are mixed in Sec 2 and Sec 3. It is unclear how the parameterization is optimized with training data, which architecture the paper uses to compute the logits ℓv\ell_v, why some design choices were made, or which component is conditioned on history and which is not; see “Questions For Authors” for more details.

Thank you for your detailed review and constructive suggestions. We appreciate your recognition of the novelty and motivation of our work. We have revised Sections 2–4 to clarify notation, model definitions, and training, and expanded the appendices with derivations and implementation details. We address the specific points below.

Notation. We now consistently use Ht−\mathcal H_{t^-} and Ht+\mathcal H_{t^+} for histories before and after an observation. The symbol mm denotes an event mark, while vv denotes an individual target: m=vm=v for single-target events and m=Vm=V for set-valued events.

Parameter optimization. For mobility, we train the recursive model from scratch using user-averaged Joint NLL. For commerce, we first train the history encoder and prediction heads, then freeze the selected predictor and train the action-rate network. Both stages select checkpoints by validation Joint NLL. Pairwise refinement is trained subsequently with the recursive predictor frozen, using the objective in Section 4.4. No pending-state labels are required. Appendix E.3 specifies the optimization and selection settings.

Architecture for ℓv\ell_v. For commerce, each candidate item’s inclusion logit is produced by a network with one 32-unit GELU hidden layer. Inputs comprise the 32-dimensional history representation, the item representation, their elementwise product, and augmented candidate features. The item representation combines a hashed embedding with a linear projection of these features, which include cart status, action age, and a learned gate. Appendix E.3 provides the architecture details.

History dependence. The table below summarizes the conditioning of each component. History includes only observations available at the prediction point; commerce action rates use the interval-start history representation. The target distribution QiQ_i is fixed by the observed action, whereas the pending probability πi(t)\pi_i(t) is updated recursively.

ComponentMobilityCommerce
Baseline time modelHistoryHistory
Baseline mark modelHistory-derived destination featuresHistory + item features
Realization rateAction features + ageAction features + history + age
Expiry rateProportional to realizationAction features + history + age
Target distribution QiQ_iObserved actionObserved action
Action-source markFixed destinationBaseline basket conditioned on action target
Pairwise interactionsNot usedHistory + item features

We explain the individual modeling choices in our responses to the methodological questions below.

AuthorsOpenReview ↗

## W2 — Experimental details and reproducibility

Thank you for highlighting these concerns. We have expanded the experimental settings, clarified the baseline adaptations and metric definitions, and released anonymous code. We also added a comparison using Chang et al.’s released predictor and Instacart data. The main results now report means and sample standard deviations across three seeds.

Architectures and hyperparameters. Appendices E.2 and E.3 specify the architectures, training settings, and checkpoint selection. For mobility, NHP, THP, S2P2, and Action-Transformer use 16-dimensional representations and up to 32 available actions. Action-Transformer has one attention layer with two heads. These models use Adam at 0.010.01, with epoch budgets and regularization selected for Action-Transformer through inner-fold validation and shared with the other three encoders. Action-Feature uses a three-hour action window, Adam at 0.050.05, and its own inner-fold selection.

For commerce, encoders process up to 32 interaction records and produce 32-dimensional representations. Action-Transformer has two attention layers with four heads each; Action-Feature mean-pools interaction embeddings. All models use the same log-normal time-head and nonempty Bernoulli basket-head architectures, with independently trained parameters. The basket head has one 32-unit GELU hidden layer. Training uses Adam at 0.0010.001, batches of 512 intervals, five epochs, and validation Joint NLL selection. We do not conduct separate hyperparameter searches for NHP, THP, and S2P2.

Evaluation at observed event times. Target NLL and Set NLL use the observed event time and preceding history:

−log⁡p(mn∣tn,Htn−).-\log p(m_n\mid t_n,\mathcal H_{t_n^-}).

They do not use sampled event times. For basket prediction, pp is the recursive distribution p0p_0 before refinement and prefp_{\mathrm{ref}} after refinement. Both retain time and history conditioning, as clarified in Section 4.1 and Appendix E.1.

Multi-target baseline adaptation. We retain NHP, THP, and S2P2 as history encoders and replace their original heads with the shared time and basket-head architectures. At purchase time tt, NHP decays its state to t−t^-, THP uses a causally masked query at tt, and S2P2 uses its left-limit state before the event impulse. The basket head combines this pre-event representation with item representations and candidate features. The current basket is used only as the label. Each encoder and its heads are trained jointly using Joint NLL.

Comparison using Chang et al. (2024). We use the authors’ released conditional-independence predictor and processed Instacart department data. We train on 8,192 official training users, select checkpoints using 2,048 official validation users, and evaluate on the complete official test split: 26,192 users and 449,861 events. Additive and pairwise refinement use the same frozen predictor and matched training and selection settings.

Test Set NLL is 6.970±0.0076.970\pm0.007 without refinement, 6.962±0.0076.962\pm0.007 with additive refinement, and 6.961±0.0076.961\pm0.007 with pairwise refinement across three seeds. Both refinements yield small improvements, while ranking scores change little. This comparison uses training and validation subsets, rather than reproducing the original full-data training. Instacart has no prospective-action records, so it tests set prediction rather than lifecycle inference. Appendix F.6 provides the settings and results.

NDCG@10 for unordered baskets. We rank items by their predicted marginal inclusion probabilities:

sv(t)=∑V⊆Ct:v∈Vp(V∣t,Ht−).s_v(t)= \sum_{V\subseteq\mathcal C_t:v\in V} p(V\mid t,\mathcal H_{t^-}).

Observed basket items have binary relevance; their recorded order is not used. We exclude Other from ranking and compute ideal DCG using the original basket size, capped at ten. Thus, NDCG evaluates the predicted ranking without requiring an ordered ground-truth basket. Appendix E.1 gives the definition.

Code and uncertainty reporting. Code and configurations are available in our anonymous repository. Three-seed standard deviations measure training variability. Paired test comparisons additionally use 2,000 user-cluster bootstrap resamples, conditional on the fitted models.

AuthorsOpenReview ↗

## M1 — M4

M1 — History dependence of the baseline intensity

Is the base intensity λvbase(t)\lambda_v^{\mathrm{base}}(t) only a function of time, or is it also conditioned on event history?

Thank you for this question. The baseline intensity depends on both time and observed history. Explicitly, this dependence is:

λmbase(t∣Ht−)=λbase(t∣Ht−)qb(m∣t,Ht−).\lambda_m^{\mathrm{base}}(t\mid\mathcal H_{t^-}) = \lambda^{\mathrm{base}}(t\mid\mathcal H_{t^-}) q_{\mathrm b}(m\mid t,\mathcal H_{t^-}).

The event rate uses a history-conditioned log-normal hazard. The mark model uses history-derived destination features for mobility and history and candidate-item features for commerce. Section 3.1 and Appendix E.3 specify these components.

M2 — Lifecycle state updates

As a core component, how does the binary lifecycle state Zi(t)Z_i(t) get updated between 0 and 1 when it sees more observations?

Thank you for this question. We infer the posterior pending probability πi(t)\pi_i(t) rather than directly observing the binary state Zi(t)Z_i(t). New actions start with πi=1\pi_i=1, and explicit cancellations set πi=0\pi_i=0.

No-event update. For an at-risk interval (a,b)(a,b) without an event, let Ri(a,u)=∫au[ri(s)+ξi(s)],dsR_i(a,u)=\int_a^u[r_i(s)+\xi_i(s)]\\,ds. Then

Si(a,b)=1−πi(a+)∫abri(u)e−Ri(a,u),du,πi(b−)=πi(a+)e−Ri(a,b)Si(a,b).\begin{aligned} S_i(a,b) &=1-\pi_i(a^+)\int_a^b r_i(u)e^{-R_i(a,u)}\\,du,\\\\ \pi_i(b^-) &=\frac{\pi_i(a^+)e^{-R_i(a,b)}}{S_i(a,b)}. \end{aligned}

This accounts for hidden expiry and the evidence provided by observing no event. Outside at-risk intervals, only expiry is propagated.

Event update. For an event with mark mm at time tt,

γi(t,m)=πi(t−)ri(t)qi(m∣t)λmbase(t)+∑jπj(t−)rj(t)qj(m∣t),πi(t+)=πi(t−)[1−γi(t,m)].\begin{aligned} \gamma_i(t,m) &= \frac{\pi_i(t^-)r_i(t)q_i(m\mid t)} {\lambda_m^{\mathrm{base}}(t) +\sum_j\pi_j(t^-)r_j(t)q_j(m\mid t)},\\\\ \pi_i(t^+) &=\pi_i(t^-)[1-\gamma_i(t,m)]. \end{aligned}

All responsibilities use the same pre-event state, and all probabilities are updated simultaneously. We retain these marginals in a factorized Bernoulli approximation. Section 3.2 and Appendix B give the derivations; Section 6.3 and Appendices F.1–F.3 evaluate state recovery and inference accuracy.

M3 — Action-age kernel and temporal influence

What is gig_i in Eq. 2, is it a function or a parameter? What does this imply about the effect of a previous action on future events over time?

Thank you for this question. gig_i is a function of action age with learnable parameters. For mobility,

ri(t)=gi(t−si)=αiLogNormal⁡(t−si;μi,σ),t>si.r_i(t)=g_i(t-s_i) =\alpha_i\operatorname{LogNormal} (t-s_i;\mu_i,\sigma),\qquad t > s_i.

Shared learned mappings of action features produce αi\alpha_i and μi\mu_i; σ\sigma is learned and shared across actions. The rate rises to a peak and then declines. The action’s predictive contribution is

πi(t−),ri(t),qi(m∣t),\pi_i(t^-)\\,r_i(t)\\,q_i(m\mid t),

so its influence depends on both action age and the inferred pending probability.

For commerce, rates are piecewise constant over four action-age bins and depend on action features and interval-start history. Section 3.1 and Appendix E.3 specify the parameterizations.

M4 — Observed action targets

How does an observed distribution QiQ_i look like (Sec 3.1)? In other words, how does such a distribution present in the training data?

Thank you for this question. Each action records a single target viv_i, represented by the point-mass distribution

Qi(v)=1v=vi.Q_i(v)=\mathbf 1\\{v=v_i\\}.

It is the specified destination for mobility or the item added to the cart for commerce. Thus, QiQ_i is observed rather than learned.

The action-source mark distribution is qi(v∣t)=Qi(v)q_i(v\mid t)=Q_i(v) for mobility. For commerce, an event may contain several items, so

qi(V∣t)=qb(V∣t,v~i∈V),q_i(V\mid t) =q_{\mathrm b}(V\mid t,\tilde v_i\in V),

where v~i\tilde v_i is the recorded item if represented in the candidate set, and Other otherwise. The pending probability πi(t)\pi_i(t) is inferred separately. Sections 3.1 and 4.1 define these quantities.

Reviewer_U5g2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PAL-MTPP for future-event prediction with prospective actions, such as destination searches and cart additions. It models each action with a lifecycle state that is updated based on subsequent events and periods without events. The paper also introduces pairwise target interactions for multi-target prediction. Experiments are conducted on mobility and e-commerce datasets.

Critical strengths
  1. The problem of modeling uncertain prospective actions is interesting and practically relevant.

  2. The paper is generally well organized. The experiments include action-based baselines, lifecycle analyses, and synthetic intensity recovery.

Critical weaknesses
  1. The paper compares PAL-MTPP with Action-Feature and Action-Transformer, but not with simpler models that explicitly handle time decay and action completion. It is unclear whether the probabilistic lifecycle model is needed to achieve the reported improvements.

  2. The experiments mainly evaluate event prediction and intensity recovery. Although the lifecycle probabilities are visualized, there is no direct evaluation of how accurately the model infers whether an action is still pending. This is an important part of the proposed method.

  3. The model assumes an independent posterior for the lifecycle states of different actions. However, multiple actions may refer to the same target or to competing targets. The effect of this approximation is not examined.

  4. The pairwise interaction module uses a relatively standard approach to modeling target co-occurrence and is trained separately from the lifecycle model. Its novelty and connection to the main contribution are not very clear.

Questions for authors
  1. How does PAL-MTPP compare with a simpler action-aware TPP using learned time decay and event-based updates?
  2. Can the authors evaluate lifecycle-state recovery on synthetic data with known ground-truth states?
  3. How much does the factorized posterior approximation affect performance when several actions are related?
6 comments
AuthorsOpenReview ↗

## W1 and Q1 — Comparison with a simpler lifecycle model

The paper compares PAL-MTPP with Action-Feature and Action-Transformer, but not with simpler models that explicitly handle time decay and action completion. It is unclear whether the probabilistic lifecycle model is needed to achieve the reported improvements.

Thank you for this suggestion. We added a simpler lifecycle baseline, Hard completion. It performs similarly to PAL-MTPP on the real datasets, while controlled experiments show the benefit of probabilistic lifecycle inference when matching events leave some actions unfinished. These comparisons are included in Section 6 and Appendix F.1.

Hard completion learns age-dependent action contributions and deterministically retires matching actions after an observed event. It uses a learned action-age kernel for mobility and learned expiry rates for commerce. We compare Hard completion with the PAL-MTPP recursive predictor before pairwise refinement, denoted as Full recursion in the tables. We use five-fold user-level evaluation for mobility and chronological test sets for commerce, averaged over three seeds. Hard completion obtains slightly lower Joint NLL, while PAL-MTPP obtains lower Set NLL on Synerise, with similar ranking performance.

Real-data prediction

DatasetMethodJoint NLL ↓\downarrowTarget/Set NLL ↓\downarrowTop-1/NDCG@10 (%) ↑\uparrow
LBS-IntentBenchHard completion3.7951.58745.27
LBS-IntentBenchFull recursion3.7991.59045.20
SyneriseHard completion0.3971.60987.04
SyneriseFull recursion0.3981.60587.06
RetailrocketHard completion0.1480.83496.38
RetailrocketFull recursion0.1490.83696.37

Mobility reports Target NLL and Top-1; commerce reports Set NLL and NDCG@10. Results are averaged over three seeds. Bold values indicate the better mean.

To identify when probabilistic updates help, we compare the two methods on controlled single- and multi-target sequences under partial, mixed, and common completion. Partial completion ends only the generating action, if any; common completion ends all matching pending actions; mixed completion applies either rule with equal probability. Both models learn from observed actions and events without pending-state labels.

Full recursion improves state recovery and prediction under partial completion. Under mixed completion, it improves prediction with similar Brier scores. Hard completion performs better under common completion. These results show why retaining lifecycle uncertainty matters: a matching event may complete one action while leaving others available to explain subsequent events.

Single-target synthetic events

MechanismMethodBrier score ↓\downarrowTop-1 (%) ↑\uparrowMRR ↑\uparrow
PartialHard completion0.08134.710.55
PartialFull recursion0.06541.760.60
MixedHard completion0.04836.210.56
MixedFull recursion0.04838.770.58
CommonHard completion0.02837.020.56
CommonFull recursion0.03636.960.56

Means over eight conditions and three seeds. Brier score measures pending-state recovery; Top-1 and MRR measure target prediction. Bold values indicate the better mean.

Multi-target synthetic events

MechanismMethodBrier score ↓\downarrowSet NLL ↓\downarrowNDCG@3 (%) ↑\uparrow
PartialHard completion0.0963.43350.48
PartialFull recursion0.0733.36454.21
MixedHard completion0.0473.39051.44
MixedFull recursion0.0473.37652.90
CommonHard completion0.0223.34452.29
CommonFull recursion0.0353.36652.28

The experimental settings are the same as above. Brier score measures pending-state recovery; Set NLL and NDCG@3 measure basket prediction. Bold values indicate the better mean.

AuthorsOpenReview ↗

## W2 and Q2 — Recovery of pending action states

The experiments mainly evaluate event prediction and intensity recovery. Although the lifecycle probabilities are visualized, there is no direct evaluation of how accurately the model infers whether an action is still pending. This is an important part of the proposed method.

Thank you for highlighting this important aspect. We added a direct evaluation of pending-state recovery. Full recursion achieves the lowest mean Brier score under partial completion without using state labels for training. Hard completion performs best under common completion. This pattern holds for both single- and multi-target events across all three seeds.

We use synthetic sequences with known pending states. Partial completion ends only the generating action, if any, while common completion ends all matching pending actions. We compare Full recursion and Hard completion with synthetic adaptations of Action-Feature and Action-Transformer equipped with per-action state heads. These adaptations use true state labels for training and model selection, whereas Full recursion and Hard completion use only observed actions and events. All methods are evaluated on the same test sequences.

For each test sequence, we compute

Brier=1∣I∣∑(i,t)∈I(πi(t)−Zi(t))2,\mathrm{Brier} = \frac{1}{|\mathcal I|} \sum_{(i,t)\in\mathcal I} \bigl(\pi_i(t)-Z_i(t)\bigr)^2,

where πi(t)\pi_i(t) is the predicted pending probability and Zi(t)Z_i(t) is the true binary state. The set I\mathcal I includes observed actions evaluated on a shared 0.25-hour grid. Scores are averaged across test sequences, eight conditions, and three seeds.

Pending-state recovery (Brier score ↓\downarrow)

MethodSingle-target / PartialSingle-target / CommonMulti-target / PartialMulti-target / Common
Full recursion0.0650.0360.0730.035
Hard completion0.0810.0280.0960.022
Action-Feature + state head0.0760.0470.0830.043
Action-Transformer + state head0.0780.0450.0860.041

Means over eight conditions and three seeds. All four methods are scored on the same test sequences.

AuthorsOpenReview ↗

## W3 and Q3 — Effect of posterior factorization

The model assumes an independent posterior for the lifecycle states of different actions. However, multiple actions may refer to the same target or to competing targets. The effect of this approximation is not examined.

Thank you for raising this concern. Our comparison with exact joint inference shows that factorized filtering introduces small average differences in the tested histories, although larger local differences occur. Mean absolute differences in pending probabilities are below 0.02 percentage points on each dataset, and mean absolute Joint NLL differences per interval are at most 1.78×10−41.78\times10^{-4}.

We compare the two filters on 351 real histories across three datasets, using identical model parameters and observations across three seeds. Exact joint inference retains dependencies between action states, whereas factorized filtering maintains a product of Bernoulli distributions. The table below reports overall results and results grouped by historical action count. Overall results also appear in Table 3 of Section 6.3; Appendix F.2 provides the historical-action groups and error analysis.

Exact joint inference versus factorized filtering

DatasetHistorical actionsHistoriesMean posterior gap (pp)Max posterior gap (pp)
LBS-IntentBenchOverall400.017411.32
LBS-IntentBench2--10130.002900.33
LBS-IntentBench11--17140.023241.32
LBS-IntentBench18--24130.025641.27
SyneriseOverall1800.000940.59
Synerise2--4600.001800.24
Synerise5--8600.000530.59
Synerise9--12600.000490.52
RetailrocketOverall1310.006594.97
Retailrocket2--4600.008581.68
Retailrocket5--8510.005693.29
Retailrocket9--12200.002944.97

Posterior gaps are absolute differences in marginal pending probabilities, expressed in percentage points (pp). Means are averaged within each history and then across histories and three seeds. Maxima cover evaluated endpoints and interior query points, not continuous-time bounds.

Local posterior differences reach 4.97 percentage points. To examine when larger differences occur, we group intervals by event-source competition, measured by the second-largest action-source responsibility in preceding events. A higher value means that multiple actions plausibly explain the same event.

As shown below, larger posterior gaps are more frequent under stronger competition on all three datasets. In the largest-gap case, the filters agree immediately after the event but diverge during subsequent no-event updates. This illustrates how discarding posterior dependencies can affect later state estimates, even when the immediate marginal update agrees.

Approximation errors by event-source competition

Second-largest action responsibility in prior eventsLBS-IntentBenchSyneriseRetailrocket
Below 1%0.00%0.00%0.00%
1% to $ < 10\%$12.99%1.98%0.48%
At least 10%50.12%11.98%5.41%

Entries are percentages of intervals with maximum sampled posterior gaps of at least 0.1 pp. Only intervals following multi-action matching events are included. Groups use the highest second-largest action-source responsibility in preceding events. Results pool three seeds.

Reviewer_d8mbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies marked temporal point processes whose histories include prospective actions, such as searches or cart additions, that may or may not be realized by later events. PAL-MTPP associates each action with a binary pending state and maintains a factorized posterior pending probability. A target's intensity combines a baseline event-history term with contributions from pending actions. The posterior is said to update after events, cancellations, and event-free intervals. For set-valued purchase events, a second-stage model preserves a baseline cardinality distribution while exponentially reweighting same-cardinality target sets using learned pairwise interactions. Experiments on one mobility and two commerce datasets report lower joint NLL than neural TPP and action-feature baselines. Additional plots examine lifecycle updates, a synthetic intensity experiment, and shuffled pair assignments. The problem is relevant and the high-level decomposition is promising, but crucial parts of the filtering algorithm and target-set computation are not specified well enough to establish correctness or reproducibility.

Critical strengths
  1. The paper identifies a practically important distinction between an observed action and a realized event. Treating the action's relevance as an evolving latent state is more faithful than retaining a static action embedding indefinitely, and using event-free time as evidence is a sensible point-process treatment.
  2. The evaluation spans mobility and commerce, includes conventional and action-aware baselines, and goes beyond a single aggregate score. The synthetic intensity study, completion-aware analysis, and signed pair visualization are useful attempts to inspect the proposed mechanisms rather than treating PAL-MTPP as a black box.
Critical weaknesses
  1. The core event update is never defined mathematically. Equations 4 and 5 specify only the no-event evolution. For a target-matching event, the paper merely says that baseline and action-realization explanations are marginalized; Appendix C.1 repeats that statement without giving responsibilities, the marginal projection, or an algorithm for multiple pending actions with the same target. The parameterizations of the baseline intensity, realization rate, expiry rate, and action target distribution are also omitted. Consequently, the likelihood in Equation 8 cannot be independently reconstructed or checked.
  2. The multi-target module is computationally underspecified. Computing the cardinality normalizer and q0(V given k) appears to require sums over all subsets, and the paper does not state the candidate-set construction, focal-set size, enumeration or approximation method, or its error. The interaction-specific evidence uses validation NLL rather than a held-out test ablation, while Table 1 combines lifecycle and interaction changes. Main results omit dispersion, and three seeds do not compensate for the absence of a test-set mechanism ablation or detailed baseline configurations.
Questions for authors
  1. What is the exact event-time posterior update for each pending action when an observed event can arise from the baseline or any of several matching actions? Please provide the responsibility equations, factorized projection, cancellation handling, and complete parameterizations needed to evaluate and train Equation 8.
  2. How are q0(k), q0(V given k), and the refined normalizer computed without enumerating the power set? Please report candidate and focal-set sizes, complexity, approximation error if any, and a held-out test ablation separating lifecycle filtering from pairwise refinement with uncertainty across independent data splits.
5 comments
AuthorsOpenReview ↗

## W1 and Q1 — Event updates and likelihood computation

Thank you for pointing out these missing details. We update all pending actions simultaneously using their probabilities of generating the observed event, then project the joint posterior onto a product of Bernoulli distributions. We have revised Sections 3–4 and expanded the appendices to make these updates, parameterizations, and likelihood calculations explicit.

Event update and posterior projection. An event may come from the baseline or one pending action. For an event with mark mm at time tt, the marked intensities and event-source responsibilities are

λmbase(t)=λbase(t)qb(m∣t),λˉm(t)=λmbase(t)+∑j:sj<tπj(t−)rj(t)qj(m∣t),γi(t,m)=πi(t−)ri(t)qi(m∣t)λˉm(t),γ0(t,m)=λmbase(t)λˉm(t),\begin{aligned} \lambda_m^{\mathrm{base}}(t) &=\lambda^{\mathrm{base}}(t)q_{\mathrm b}(m\mid t),\\\\ \bar\lambda_m(t) &=\lambda_m^{\mathrm{base}}(t) +\sum_{j:s_j < t}\pi_j(t^-)r_j(t)q_j(m\mid t),\\\\ \gamma_i(t,m) &=\frac{\pi_i(t^-)r_i(t)q_i(m\mid t)} {\bar\lambda_m(t)},\qquad \gamma_0(t,m) =\frac{\lambda_m^{\mathrm{base}}(t)} {\bar\lambda_m(t)}, \end{aligned}

where πi(t−)\pi_i(t^-) is the pending probability before the event, ri(t)r_i(t) is the realization rate, and qi(m∣t)q_i(m\mid t) is the action-source mark probability. The responsibilities satisfy γ0+∑iγi=1\gamma_0+\sum_i\gamma_i=1.

All responsibilities use the same pre-event state, including when several actions match the event. We update their marginals simultaneously and form the factorized posterior:

πi(t+)=πi(t−)[1−γi(t,m)],p^(z+∣Ht+)=∏iBern⁡(zi+;πi(t+)).\begin{aligned} \pi_i(t^+) &=\pi_i(t^-)[1-\gamma_i(t,m)],\\\\ \hat p(\mathbf z^+\mid\mathcal H_{t^+}) &=\prod_i\operatorname{Bern}(z_i^+;\pi_i(t^+)). \end{aligned}

The marginal update is exact under the factorized pre-event distribution; replacing the resulting joint posterior with this product is the approximation. New actions initialize with πi=1\pi_i=1, and explicit cancellations set the affected probability to zero. Section 3.2 and Appendix B.2 provide the derivation.

Model parameterizations. The baseline event-time intensity uses a history-conditioned log-normal hazard. For mobility, its mark distribution is a softmax over destination scores. For commerce, candidate logits give Bernoulli inclusion parameters ρv(t)=σ(ℓv(t))\rho_v(t)=\sigma(\ell_v(t)) before conditioning on nonemptiness. The resulting basket distribution is:

qb(V∣t)=∏v∈Vρv(t)∏u∈Ct∖V(1−ρu(t))1−∏u∈Ct(1−ρu(t)),∅≠V⊆Ct.q_{\mathrm b}(V\mid t) = \frac{ \prod_{v\in V}\rho_v(t) \prod_{u\in\mathcal C_t\setminus V}(1-\rho_u(t)) }{ 1-\prod_{u\in\mathcal C_t}(1-\rho_u(t)) }, \qquad \varnothing\ne V\subseteq\mathcal C_t.

For mobility, realization follows a scaled log-normal action-age kernel, with proportional expiry:

ri(t)=αiLogNormal⁡(t−si;μi,σ),t>si,αi=softplus⁡(bα+wα⊤xi)+10−6,μi=bμ+wμ⊤xi,σ=softplus⁡(sσ)+0.05,ξi(t)=κri(t),κ=softplus⁡(sκ)+10−8,\begin{aligned} r_i(t) &=\alpha_i\operatorname{LogNormal} (t-s_i;\mu_i,\sigma),\qquad t > s_i,\\\\ \alpha_i &=\operatorname{softplus} (b_\alpha+w_\alpha^\top x_i)+10^{-6},\\\\ \mu_i &=b_\mu+w_\mu^\top x_i,\qquad \sigma=\operatorname{softplus}(s_\sigma)+0.05,\\\\ \xi_i(t) &=\kappa r_i(t),\qquad \kappa=\operatorname{softplus}(s_\kappa)+10^{-8}, \end{aligned}

where xix_i contains observed action features, action age is measured in hours, and the weights and biases are shared across actions. For commerce, a network with one 16-unit tanh⁡\tanh hidden layer takes action features and the interval-start history representation. Eight softplus outputs give realization and expiry rates over four action-age bins, with boundaries at 1, 6, and 24 hours. Rates are constant within each bin during an interval and recomputed at the next interval origin.

Each observed action identifies one target, so Qi(v)=1v=viQ_i(v)=\mathbf 1\\{v=v_i\\}. The action-source mark distribution is

qi(v∣t)=Qi(v)for mobility,qi(V∣t)=qb(V∣t,v~i∈V)for commerce,q_i(v\mid t)=Q_i(v) \quad\text{for mobility},\qquad q_i(V\mid t) =q_{\mathrm b}(V\mid t,\tilde v_i\in V) \quad\text{for commerce},

where v~i\tilde v_i is the represented action target. Sections 3.1 and 4.1 and Appendix E.3 provide the implementation and training details.

Likelihood computation. The total predictive intensity and interval negative log-likelihood are

λˉ(t)=∑mλˉm(t)=λbase(t)+∑i:si<tπi(t−)ri(t),L(a,b]=∫abλˉ(u),du−δblog⁡λˉmb(b),\begin{aligned} \bar\lambda(t) &=\sum_m\bar\lambda_m(t) =\lambda^{\mathrm{base}}(t) +\sum_{i:s_i < t}\pi_i(t^-)r_i(t),\\\\ \mathcal L_{(a,b]} &=\int_a^b\bar\lambda(u)\\,du -\delta_b\log\bar\lambda_{m_b}(b), \end{aligned}

where δb\delta_b indicates whether an event occurs at bb. The integral propagates pending probabilities along the no-event path. The event term uses πi(b−)\pi_i(b^-) before the event update. Right-censored intervals contribute only the integral term. Section 3.3 and Appendix C give the likelihood derivation.

AuthorsOpenReview ↗

## W2 and Q2 — Set normalization and test-set ablations

The multi-target module is computationally underspecified. Computing the cardinality normalizer and q0(V given k) appears to require sums over all subsets, and the paper does not state the candidate-set construction, focal-set size, enumeration or approximation method, or its error. The interaction-specific evidence uses validation NLL rather than a held-out test ablation, while Table 1 combines lifecycle and interaction changes. Main results omit dispersion, and three seeds do not compensate for the absence of a test-set mechanism ablation or detailed baseline configurations.

Thank you for pointing out these missing details. The set probabilities and refinement normalizer are computed exactly under the focal interaction model, without enumerating all candidate subsets. We have clarified this computation in Section 4 and Appendix D, added separate test-set ablations for lifecycle inference and refinement, and reported training variability and results across disjoint user groups.

Set probabilities and normalization. We select the eight candidates with the highest baseline logits as focal targets, giving 28=2562^8=256 focal subsets. Interactions are restricted to these targets; dynamic programming (DP) aggregates the remaining candidates by set size.

Let p0(V)p_0(V) be the recursive predictor’s set distribution before refinement. For each focal subset AA and set size kk, define

Bk(A)=∑U⊆Ct∖Ft∣U∣=k−∣A∣p0(A∪U).B_k(A)= \sum_{\substack{ U\subseteq\mathcal C_t\setminus\mathcal F_t\\\\ |U|=k-|A| }} p_0(A\cup U).

This is the total probability of size-kk sets whose focal targets are exactly AA. We compute

p0(k)=∑A⊆FtBk(A),q0(V∣k)=p0(V)p0(k),Zk=∑A⊆FtBk(A)eF(A)p0(k),\begin{aligned} p_0(k) &=\sum_{A\subseteq\mathcal F_t}B_k(A),\\\\ q_0(V\mid k) &=\frac{p_0(V)}{p_0(k)},\\\\ Z_k &=\frac{\sum_{A\subseteq\mathcal F_t} B_k(A)e^{F(A)}}{p_0(k)}, \end{aligned}

where F(A)F(A) is the refinement score. Thus, evaluating q0(V∣k)q_0(V\mid k) requires the probability of that set and the DP-computed p0(k)p_0(k), rather than enumeration of all size-kk sets.

The background DP maintains two arrays over set size: one for Bernoulli-odds products and one incorporating action-source weights. They are computed once up to the required size and reused across focal subsets. For fixed kk, focal masses cost O(Ck+2ff)O(Ck+2^f f) and interaction scores cost O(2ff2)O(2^f f^2), where CC is the candidate count and f=8f=8. Normalization is exact up to floating-point precision. Restricting interactions to focal targets limits the modeled pairs, rather than approximating the normalizer. Appendix D.2 gives the DP recurrences.

Candidates and baseline configurations. Synerise candidates combine the 20 most frequently purchased training items, the user’s previously purchased items, and items in either of the two cart states. Retailrocket candidates combine the 64 most recently observed distinct items and the 32 most frequently purchased training items. Both include an Other token. In the full-cohort test evaluation, median candidate counts (95th percentiles), including Other, are 24 (54) for Synerise and 36 (97) for Retailrocket.

Appendices E.1–E.3 specify candidate construction, baseline architectures, prediction heads, hyperparameters, and checkpoint selection. Code and configurations are available in our anonymous repository.

Separate test-set ablations. We isolate lifecycle inference and refinement using separate controlled comparisons.

For lifecycle inference, we compare Bayesian no-event conditioning with expiry-only propagation. Both variants fix event updates to Hard completion and are retrained separately with matched initialization and training settings. Bayesian conditioning improves Synerise Joint NLL, with a paired difference of −0.001985-0.001985 (95% CI: [−0.002134,−0.001834][-0.002134,-0.001834]); differences on mobility and Retailrocket are small. Appendix F.3 provides the protocol, and Appendix F.1 evaluates pending-state recovery on controlled synthetic sequences.

No-event conditioning on real data

DatasetPropagationJoint NLL ↓\downarrowMark NLL ↓\downarrowRanking (%) ↑\uparrow
LBS-IntentBenchExpiry only3.795±0.0003.795\pm0.0001.588±0.0001.588\pm0.00045.18±0.0045.18\pm0.00
LBS-IntentBenchBayesian no-event3.795±0.0003.795\pm0.0001.588±0.0001.588\pm0.00045.17±0.0145.17\pm0.01
SyneriseExpiry only0.397±0.0110.397\pm0.0111.609±0.1281.609\pm0.12887.04±0.2287.04\pm0.22
SyneriseBayesian no-event0.395±0.0130.395\pm0.0131.608±0.1311.608\pm0.13187.05±0.2187.05\pm0.21
RetailrocketExpiry only0.148±0.0030.148\pm0.0030.834±0.0090.834\pm0.00996.38±0.0896.38\pm0.08
RetailrocketBayesian no-event0.148±0.0040.148\pm0.0040.835±0.0090.835\pm0.00996.39±0.0896.39\pm0.08

Means ±\pm sample standard deviations across three seeds. For mobility, Mark NLL and ranking denote Target NLL and Top-1; for commerce, they denote Set NLL and NDCG@10. Displayed zero standard deviations reflect rounding.

For refinement, we compare no refinement, additive refinement, and pairwise refinement using the same frozen recursive predictor. The two heads have matched parameter counts, training budgets, and selection settings. Evaluation includes multi-target test events with at least eight candidates.

Both refinements reduce Set NLL on both datasets. On Synerise, pairwise refinement improves Set NLL over additive refinement (paired difference, pairwise minus additive: −0.0315-0.0315; 95% user-bootstrap CI: [−0.0405,−0.0227][-0.0405,-0.0227]). On Retailrocket, the two refinements perform similarly, with slightly higher mean ranking scores for pairwise refinement. Appendix F.4 reports the paired comparisons.

Refinement on multi-target test events

DatasetMethodSet NLL ↓\downarrowNDCG@3 (%) ↑\uparrowRecall@3 (%) ↑\uparrow
SyneriseNo refinement5.13687.5777.14
SyneriseAdditive refinement5.11787.6577.23
SynerisePairwise refinement5.08587.6277.18
RetailrocketNo refinement3.73288.1181.55
RetailrocketAdditive refinement3.71988.2381.68
RetailrocketPairwise refinement3.72288.3381.82

Means over three seeds on multi-target test events with at least two represented targets and eight candidates: 1,461 for Synerise and 457 for Retailrocket. Additive refinement uses item-level adjustments without pairwise interactions. Bold values indicate the best mean.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to Reviewers and Summary of Revisions

We thank all three reviewers for their constructive comments and for recognizing the novelty and motivation of our work. Their feedback helped us improve the manuscript in three areas:

Method clarity. (d8mb, 8UPt) We reorganized Sections 3 and 4 to clarify the model formulation and inference procedure. We now explicitly define event-source responsibilities, event updates, and the factorized posterior projection, and explain the parameterizations and likelihood computation. We also clarify candidate-set construction, focal-target selection, and exact target-set normalization through dynamic programming, with supporting derivations in the appendices.

Experimental validation. We added the following experiments to address the reviewers' concerns:

  • Comparison with a simpler lifecycle model. (U5g2) We added Hard completion to examine whether learned time decay and deterministic completion explain the gains. Both lifecycle models outperform the five original baselines on the real datasets, with comparable overall predictive performance (Section 6.2).

  • Direct evaluation of pending-state recovery. (U5g2) Synthetic experiments with known action states show that Full recursion improves state recovery and prediction under partial completion, where matching events leave some actions unfinished (Section 6.3 and Appendix F.1).

  • Accuracy of factorized filtering. (U5g2) Comparisons with exact joint inference show mean posterior differences below 0.02 percentage points on all three datasets, supporting the accuracy of factorized filtering on the evaluated histories (Section 6.3 and Appendix F.2).

  • Benefits of lifecycle updates on real data. (U5g2, d8mb) Separate analyses show that no-event conditioning improves prediction on Synerise, while retaining matched actions benefits subsequent return visits on LBS-IntentBench (Appendix F.3).

  • Contribution of pairwise refinement. (U5g2, d8mb, 8UPt) Controlled test-set comparisons use the same frozen recursive predictor. Basket-size analyses show larger Set NLL reductions relative to additive refinement for baskets with at least three represented targets than for two-target baskets (Section 6.4).

  • Evaluation across user groups. (d8mb) To examine whether refinement gains persist across different user populations, we independently preprocess and train on three disjoint user groups in each commerce dataset. Pairwise refinement achieves lower mean multi-target Set NLL than additive refinement in all six groups. These results are reported in the reviewer response.

  • Transfer to an existing set-valued predictor. (8UPt) Instacart experiments show that refinement also improves an existing set-valued predictor without prospective actions (Appendix F.5).

  • Recovery of event intensity. (8UPt) Synthetic experiments show that Full recursion improves total-intensity and action-contribution recovery over expiry-only propagation, with lower sequence NLL (Appendix F.6).

The main results now report means and sample standard deviations across three seeds.

Reproducibility. (d8mb, 8UPt) We clarified data processing, baseline adaptations, model architectures, and training and evaluation protocols. Code and reproduction instructions are available in our anonymous repository.

We have carefully addressed the reviewers‘ major concerns through detailed responses, additional experiments, and substantial revisions. However, we have not yet received follow-up feedback from two reviewers. We would greatly appreciate further assessment of our responses and welcome further discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.