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

Liangkai Zhou; Susu Xu; Shuqi Zhong; Shan Lin

cs.LGMar 12, 2026

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

Authors: Liangkai Zhou, Susu Xu, Shuqi Zhong, Shan Lin

Organizations: Stony Brook University Stony Brook, New York, United States · Johns Hopkins University Baltimore, Maryland, United States

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.

Figures & tables

Figure 1: Multi-task causal graph with shared causal mechanism.

Figure 2: MTAC Framework. The white nodes represent parameterized deterministic neural networks, gray nodes represents drawing samples from the respective distribution. The left panel shows the multi-task structural equation model, in which the mechanism variables W and confounders Z are shared across tasks while causes

X_{k}

and outcomes

Y_{k}

are task-dependent. The orange box details the multi-task mechanism module for a single mediator

W_{1}

. The right panel illustrates the MAP inference procedure.

Figure 3: Multi-task SEM for mechanism variable

\mathbf{W}_{i}

. The norm of first-layer weights decides whether input variables are parents.

Model	Parking Violation		Abandoned Property		Unsanitary Condition
	MAE	MSE	MAE	MSE	MAE	MSE
MTAC	0.2971	0.1616	0.4163	0.1910	0.3755	0.2501
CEVAE	$0.3235$	$0.1925$	$0.4867$	$0.2607$	$0.4091$	$0.2884$
TEDVAE	$0.4315$	$0.2636$	$0.4932$	$0.2583$	$0.4188$	$0.2700$
BSM-UR	$0.4437$	$0.2976$	$0.5438$	$0.3175$	$0.3982$	$0.3026$
PLE	$0.3337$	$0.2174$	$0.4533$	$0.2351$	$0.4101$	$0.2697$

Table 1: MAE and MSE estimation for MTAC and baselines across three tasks.

	Parking Violation		Abandoned Property		Unsanitary Condition
Model	MAE	MSE	MAE	MSE	MAE	MSE
Multi-task	0.2971	0.1616	0.4163	0.1910	0.3755	0.2501
Single-task	0.3693	0.2947	0.4974	0.2517	0.4486	0.2989

Table 2: Performance comparison for MTAC between multi-task training and single-task training.

	UC+PV $\to$ AP		UC+AP $\to$ PV		PV+AP $\to$ UC
Model	MAE	MSE	MAE	MSE	MAE	MSE
Full fine-tuned	0.4507	0.2313	0.3475	0.2502	0.3977	0.2656
Deviation-only fine-tuned	0.4631	0.2297	0.3407	0.2586	0.4106	0.2694
Zero-shot	0.7702	0.4323	0.7699	0.5472	0.8965	0.7518
Single-task trained	0.4974	0.2517	0.3693	0.2947	0.4486	0.2989

Table 3: Prediction Error of transferred MTAC. The transferred models are trained on two tasks and transferred to the remaining task. PV: parking violation; AP: abandoned property; UC: unsanitary condition.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Factors
Finance	mean income, unemployment rate, mortgage ratio, poverty rate, housing cost
Education Attainment	% less than high school, % high school, % bachelor or higher
Race & Culture	% Hispanic, % white, % black, % Asian
Access to Technology	% has computer, % has smartphone, % has internet
Social Environment	population, % multifamily, % owner occupied, % renter occupied, median year built, % room occupation $\geq 0.5$ , mobility rate

Appendix

Table 4: An overview of the socioeconomic status factors.

Model	Parking Violation		Abandoned Property		Unsanitary Condition
	MAE	MSE	MAE	MSE	MAE	MSE
Complete MTAC	0.2971	0.1616	0.4163	0.1910	0.3755	0.2501
w/o MAP	0.5841	0.6977	0.7841	0.6006	0.5957	0.7426
w/o Causality	0.3916	0.1811	0.4967	0.2842	0.4287	0.2938

Appendix

Table 5: Ablation study. For MTAC without MAP, the cause is estimated directly from the forward causal model.

Panel of Figure 4 — Figure 4: Sensitivity of reconstruction performance on model parameters.

Explore similar work

Causal Learning with the Invariance Principle
May 13, 2026Francesco Montagna, Francesco LocatelloCausal Discovery Methods Causal Graph
Institute of Science and Technology Austria
Abstract Event Causal Rules: Induction and Application
Aug 5, 2026Ziwei Zheng, Peiqiong Chen, Bang WangCausal Attention Induction
School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan, China
FoundCause: Causal Discovery with Latent Confounders from Observational Data
Jun 16, 2026Patrick Blöbaum, Krishnakumar Balasubramanian, Shiva Prasad KasiviswanathanCausal Discovery Methods Causal Modeling
Amazon Web Services · Department of Statistics, University of California, Davis

May 13, 2026, stat.MLJ/K move · Enter open · S save

Causal Learning with the Invariance Principle

Francesco Montagna, Francesco Locatello

Institute of Science and Technology Austria

Causal Discovery Methods Causal Graph Causal Causal Reasoning Principles Nonlinearity

Open paper AskPDF arXiv

Causal discovery, the problem of inferring the direction of causality, is generally ill-posed. We use the language of structural causal models (SCM) to show that assuming that the causal relations are acyclic and invariant across multiple environments (e.g., the way minimum wage affects employment rate is stable across different geographical regions), \textit{only} two auxiliary environments are sufficient to infer the causal graph for arbitrary nonlinear mechanisms. Moreover, we demonstrate that this implies identifiability of the SCM functional mechanisms: as a corollary, we show that \textit{two} auxiliary environments are sufficient to guarantee correct counterfactual inference. We empirically support our theoretical results on synthetic data.

Abstract

Explore similar work

Causal Learning with the Invariance Principle

Abstract Event Causal Rules: Induction and Application

FoundCause: Causal Discovery with Latent Confounders from Observational Data