acceptodds
Under review as a conference paper at ICLR 2027

Multi-Task Anti-Causal Learning for Reconstructing Urban Events from Residents’ Reports

Abstract

Many real-world machine learning tasks are anti-causal: they require inferring latent causes from observed effects. In practice, we often face multiple related tasks where the structural dependencies are a hybrid of task-invariant and task-specific mechanisms. We propose Multi-Task Anti-Causal learning (MTAC), a framework for estimating causes from outcomes and confounders by explicitly exploiting such cross-task invariances. MTAC learns a structural equation model (SEM) that factorizes the outcome-generation process into (i) a task-invariant mechanism and (ii) task-specific mechanisms via a shared backbone with task-specific deviations. Building on the learned forward model, MTAC performs maximum A posteriori (MAP) based inference to reconstruct causes by jointly optimizing latent mechanism variables and cause magnitudes under the learned structural model. We evaluate MTAC on the application of urban event reconstruction from resident reports, spanning three tasks: parking violations, abandoned properties, and unsanitary conditions. On real-world data collected from Manhattan and Newark, MTAC consistently improves reconstruction accuracy over strong baselines, achieving up to 33.04% MAE reduction and demonstrating the benefits of learning transferable mechanisms across tasks.

Primary area:
causal reasoning
Keywords:
Causal learning, Urban event prediction, Multi-task learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DbtwOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the problem of anti-causal learning by proposing the MTAC framework and applying it to the task of reconstructing the true distribution of urban events based on resident complaints. The authors argue that resident complaints result from urban events filtered through reporting behavior, which is influenced by socioeconomic status and is partially shared across event types. To address this, MTAC constructs a multi-task structural equation model that decomposes the underlying reporting preference mechanism into a task-shared backbone and task-specific biases. After learning a forward model, MTAC uses maximum a posteriori inference to jointly optimize the number of events and the latent mechanism variables, thereby reconstructing the true events from the biased complaints. The authors evaluated the model across three tasks, and the results show that MTAC outperforms multiple baselines.

Critical strengths
  1. This paper presents a novel research topic by formalizing the reconstruction of urban events as a multi-task anti-causal inference problem and explicitly modeling reporting behavior as a latent mediator influenced by socioeconomic factors.
  2. The proposed model uses parameters from a shared backbone, combined with a task-bias decomposition mechanism and a differentiated sparsity penalty, to structurally separate the cross-task invariant components from the task-specific components.
  3. MAP-based inversion using forward structural models and observed complaints, can distinguish between two possible explanations. E.g., a high number of incidents with few reports and a low number of incidents with many reports.
  4. The ground truth, such as parking tickets and hearing transcripts, is also influenced by the intensity of law enforcement and may not necessarily be an unbiased measure of actual events.
  5. The zero-shot transfer performance is very poor, indicating that the sharing mechanism alone is not sufficient.
Critical weaknesses
  1. The paper does not discuss distinguishability. Specifically, under what conditions the two scenarios, “many events but few reports” and “few events but many reports”, can truly be distinguished from one another. Therefore, the reliability of MAP inversion lacks theoretical support.
  2. The results presented in the paper are inconsistent. While the abstract and Section 5.2 state that the MAE was reduced by up to 33.04%, the list of contributions states 34.61%.
  3. The description of the baseline training methods in the paper is inconsistent. Section 5.1 states that all baselines were trained on data merged from the three tasks, while Appendix A.3 states that some baselines were trained separately for each task.
Questions for authors

please see weaknesses

Reviewer_kN5oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to reconstruct urban event counts from residents’ reports while accounting for differences in reporting preferences across neighborhoods. The authors propose Multi-Task Anti-Causal learning (MTAC), which uses a shared backbone and task-specific deviations to model common and event-specific reporting mechanisms. It then performs MAP inference to estimate event counts from report counts and SES factors. Experiments cover parking violations, abandoned properties, and unsanitary conditions. MTAC achieves lower reconstruction errors than the compared baselines, while joint-training and transfer experiments suggest that sharing information across tasks improves estimation accuracy.

Critical strengths

1. The paper recognizes that residents’ report counts reflect both event incidence and reporting preferences. Separating these factors could help allocate resources to areas with more underlying events. The problem is practically relevant and clearly motivated.

2. The method first models how urban events and socioeconomic status (SES) factors generate residents’ reports, then uses maximum a posteriori (MAP) inference to reconstruct event counts. The shared backbone and task-specific deviations provide a reasonable way to exploit common reporting mechanisms while retaining differences across tasks. The overall approach is clear.

3. The experiments are reasonably comprehensive, evaluating the method through single-task versus multi-task comparisons, cross-task transfer, component ablations, and hyperparameter sensitivity analyses.

Critical weaknesses

1. Shared backbones with task-specific deviations, neural structural equation models (SEMs), and maximum a posteriori (MAP) inference are all well-established approaches. The paper should explain more clearly what new technical contribution arises from combining them.

2. The three event types span different cities and time periods, where reporting behavior may differ substantially. Better joint-training performance does not establish whether the gains come from sharing task-invariant mechanisms or simply from access to more training data.

3. The main text states that the baselines are trained on concatenated data from all tasks, whereas the appendix says that some baselines are trained separately on each task. This inconsistency needs clarification, as it affects the fairness and interpretation of the comparisons.

Questions for authors

1. What distinguishes MTAC from existing shared–private multi-task models and generative-model-based anti-causal inference? Please clarify the methodological advances separately from the application contribution to urban event reconstruction.

2. The main text states that the baselines are trained on concatenated multi-task data with a task indicator, whereas Appendix A.3 says that some baselines are trained separately on each task. Please clarify which setting was used for the reported results.

3. The experiments compare joint training across all three tasks with single-task training, but do not separately report two-task joint-training results. Parking violations and unsanitary conditions come from the same city, while abandoned properties come from another. Does adding the abandoned-property task consistently help the other tasks, or can it lead to negative transfer?

Reviewer_XV66OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses urban event reconstruction from resident reports, where observed report counts reflect both event occurrence and heterogeneous reporting behavior. The proposed Multi-Task Anti-Causal Learning framework (MTAC) models event counts, latent reporting mechanisms, and observed reports through a neural structural equation model. Shared parameters and task-specific deviations encode partially shared reporting mechanisms across event types. At inference time, the method jointly optimizes event counts and latent mechanism variables through MAP estimation.

Experiments cover parking violations, abandoned properties, and unsanitary conditions in Manhattan and Newark. The paper reports improvements over the evaluated baselines and includes single-task comparisons, transfer experiments, ablations, and sensitivity analyses. These results provide preliminary evidence for the predictive usefulness of structured sharing, although the stronger interpretations concerning mechanism invariance and causal robustness remain insufficiently supported.

Critical strengths

A meaningful problem formulation and structured sharing strategy. The paper recognizes that resident reports are filtered measurements of urban events and proposes sharing parts of the reporting mechanism across event categories. Locating the shared representation within the forward report-generation model provides a meaningful inductive bias beyond ordinary direct prediction.

Consistent reported predictive improvements across three tasks. In Table 1, MTAC reduces MAE relative to the strongest listed baseline for each task by approximately 8.16%, 8.16%, and 5.70%, respectively. Table 2 also reports improvements from multi-task training. Together with the transfer and direct-prediction comparisons, these results provide preliminary evidence that the proposed architecture has practical value.

Critical weaknesses
  1. My main concern is the limited conceptual contribution. MTAC combines a shared-specific multi-task model with standard MAP inference, which is somewhat normal. The causal interpretation also needs stronger support. The paper provides neither an identification argument nor independent validation that its latent factors capture causal reporting mechanisms. Prediction and transfer gains can arise from regularization and cross-task information sharing, and do not by themselves establish the factors’ causal interpretation.

  2. The evaluation needs correction and the causal claims need qualification.

All three zero-shot entries in Table 3 and the abandoned-property “w/o MAP” entry in Table 5 violate MSE≥MAE2\mathrm{MSE}\geq\mathrm{MAE}^2 when both metrics use the same weighted errors on the same scale. Neither rounding nor fold averaging explains these discrepancies. These auxiliary results need correction but do not directly invalidate the main comparisons. Section 5 describes pooled baseline training while Appendix A.3 states that each task uses separate training. The authors should confirm which protocol they used so readers can assess comparison fairness.

Freezing θˉ\bar\theta does not fix the full mechanism because the model still adapts Δk\Delta_k. The transfer gains support reuse but do not establish invariant effects of SES on reporting. Appendix A.5 evaluates SES subgroups within the existing test set without testing a specified distribution shift. The “w/o MAP” ablation removes report information and therefore cannot isolate the benefit of MAP. The administrative labels may themselves depend on reporting or enforcement. Predicting these records accurately does not establish recovery of true event incidence.

Questions for authors
  1. What exact training objective is optimized when WW is unobserved? Is the mechanism network trained through reparameterized samples and the downstream report loss, or another approximation? Please clarify the sampling procedure, gradient path, and the wording that no SEM is fitted for WW.

  2. Can you explain or correct the MAE/MSE inconsistencies in Table 3 and the abandoned-property "w/o MAP" entry in Table 5? Please specify the metric scales, sample weighting, and aggregation across runs or folds.

  3. Were the baselines trained on pooled data or separately by task? Do their test-time inference procedures condition on the same observed reports and SES variables as MTAC? Please also clarify the spatial or temporal unit used for data splitting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.