We study the problem of learning discrete latent variable causal structures from mixed-type observational data. Traditional methods, such as those based on the tensor rank condition, are designed to identify discrete latent structure models and provide robust identification bounds for discrete causal models. However, when observed variables—specifically, those representing the children of latent variables—are collected at various levels with continuous data types, the tensor rank condition is not applicable, limiting further causal structure learning for latent variables. In this paper, we consider a more general case where observed variables can be either continuous or discrete, and further allow for scenarios where multiple latent parents cause the same set of observed variables. We show that, under the completeness condition, it is possible to discretize the data in a way that satisfies the full-rank assumption required by the tensor rank condition. This enables the identifiability of discrete latent structure models within mixed-type observational data. Moreover, we introduce the two-sufficient measurement condition, a more general structural assumption under which the tensor rank condition holds and the underlying latent causal structure is identifiable by a proposed two-stage identification algorithm. Extensive experiments on both simulated and real-world data validate the effectiveness of our method.

With the development of deep neural networks, deep reinforcement learning (DRL) has shown great potential in solving complex decision-making problems in high-dimensional environments. Among various DRL methods, Q-value function approaches are well-grounded in theory and exhibit strong practical performance, but often suffer from overestimation of the value function. We find that this overestimation primarily stems from unobserved random factors (latent variables) within the environmental dynamics, while existing solutions fail to address this issue, relying mainly on low-noise and unbiased sampling environments. To address this problem, we introduce latent variables to model unobserved randomness behind the environmental dynamics and employ causal models to capture the underlying data generation mechanisms. Subsequently, we propose a causal Q-value function method based on latent causal models. Specifically, we first design a causal encoding-decoding framework for learning the latent causal models, then utilize the inferred latent variables to construct a causal Q-value function, effectively mitigating the Q-value function overestimation problem. Experimental results show that our method significantly improves the stability of policy learning and the speed of convergence, enhancing the robustness of reinforcement learning in complex dynamic environments.

Causal discovery from observational data is an important but challenging task in many scientific fields. A recent line of work formulates the structure learning problem as a continuous constrained optimization task using an algebraic characterization of directed acyclic graphs (DAGs) and the least-square loss function. Though the least-square loss function is well justified under the standard Gaussian noise assumption, it is limited if the assumption does not hold. In this work, we theoretically show that the violation of the Gaussian noise assumption will hinder the causal direction identification, making the causal orientation fully determined by the causal strength as well as the variances of noises in the linear case and by the strong non-Gaussian noises in the nonlinear case. Consequently, we propose a more general entropy-based loss that is theoretically consistent with the likelihood score under any noise distribution. We run extensive empirical evaluations on both synthetic data and real-world data to validate the effectiveness of the proposed method and show that our method achieves the best in structure Hamming distance, false discovery rate (FDR), and true-positive rate (TPR) matrices.

Conditional independence (CI) testing is an important problem, especially in causal discovery. Most testing methods assume that all variables are fully observable and then test the CI among the observed data. Such an assumption is often untenable beyond applications dealing with, e.g., psychological analysis about the mental health status and medical diagnosing (researchers need to consider the existence of latent variables in these scenarios); and typically adopted latent CI test schemes mainly suffer from robust or efficient issues. Accordingly, this article investigates the problem of testing CI between latent variables. To this end, we offer an auxiliary regression-based CI (AReCI) test by taking the measured variable as the surrogate variable of the latent variables to conduct the regression over the latent variables under the linear causal models, in which each latent variable has some certain measured variables. Specifically, given a pair of latent variables$L_X$and$L_Y$, and a corresponding latent variable set$\mathcal{L}_{O}$,$L_X \CI L_Y | \mathcal{L}_{O}$holds if and only if$A_{\{L_X\}}-\omega_1^\intercal A^{\prime}_{\{\mathcal{L}_{O}\}}$and$A_{\{L_Y\}}-\omega_2^\intercal A^{\prime\prime}_{\{\mathcal{L}_{O}\}}$are statistically independent, where$A^{\prime}$and$A^{\prime\prime}$are the two disjoint subset of the measured variable for the corresponding latent variables,$A^{\prime}_{\{\mathcal{L}_{O}\}} \cap A^{\prime\prime}_{\{\mathcal{L}_{O}\}} =\emptyset$, and$\omega_1$is a parameter vector characterized from the cross covariance between$A_{\{L_X\}}$and$A^{\prime}_{\{\mathcal{L}_{O}\}}$, and$\omega_{2}$is a parameter vector characterized from the cross covariance between$A_{\{L_Y\}}$and$A^{\prime\prime}_{\{\mathcal{L}_{O}\}}$. We theoretically show that the AReCI test is capable of addressing both Gaussian and non-Gaussian data. In addition, we find that the well-known partial correlation test can be seen as a special case of the AReCI test. Finally, we devise a causal discovery method by using the AReCI test as the CI test. The experimental results on synthetic and real-world data illustrate the effectiveness of our method.

Deep neural networks (DNNs) have been demonstrated to be vulnerable to well-crafted adversarial examples, which are generated through either well-conceived L_p-norm restricted or unrestricted attacks. Nevertheless, the majority of those approaches assume that adversaries can modify any features as they wish, and neglect the causal generating process of the data, which is unreasonable and unpractical. For instance, a modification in income would inevitably impact features like the debt-to-income ratio within a banking system. By considering the underappreciated causal generating process, first, we pinpoint the source of the vulnerability of DNNs via the lens of causality, then give theoretical results to answer where to attack. Second, considering the consequences of the attack interventions on the current state of the examples to generate more realistic adversarial examples, we propose CADE, a framework that can generate Counterfactual ADversarial Examples to answer how to attack. The empirical results demonstrate CADE's effectiveness, as evidenced by its competitive performance across diverse attack scenarios, including white-box, transfer-based, and random intervention attacks.

Learning Granger causality from event sequences is a challenging but essential task across various applications. Most existing methods rely on the assumption that event sequences are independent and identically distributed (i.i.d.). However, this i.i.d. assumption is often violated due to the inherent dependencies among the event sequences. Fortunately, in practice, we find these dependencies can be modeled by a topological network, suggesting a potential solution to the non-i.i.d. problem by introducing the prior topological network into Granger causal discovery. This observation prompts us to tackle two ensuing challenges: 1) how to model the event sequences while incorporating both the prior topological network and the latent Granger causal structure, and 2) how to learn the Granger causal structure. To this end, we devise a unified topological neural Poisson auto-regressive model with two processes. In the generation process, we employ a variant of the neural Poisson process to model the event sequences, considering influences from both the topological network and the Granger causal structure. In the inference process, we formulate an amortized inference algorithm to infer the latent Granger causal structure. We encapsulate these two processes within a unified likelihood function, providing an end-to-end framework for this task. Experiments on simulated and real-world data demonstrate the effectiveness of our approach.

Missing data are an unavoidable complication frequently encountered in many causal discovery tasks. When a missing process depends on the missing values themselves (known as self-masking missingness), the recovery of the joint distribution becomes unattainable, and detecting the presence of such self-masking missingness remains a perplexing challenge. Consequently, due to the inability to reconstruct the original distribution and to discern the underlying missingness mechanism, simply applying existing causal discovery methods would lead to wrong conclusions. In this work, we found that the recent advances additive noise model has the potential for learning causal structure under the existence of the self-masking missingness. With this observation, we aim to investigate the identification problem of learning causal structure from missing data under an additive noise model with different missingness mechanisms, where the `no self-masking missingness' assumption can be eliminated appropriately. Specifically, we first elegantly extend the scope of identifiability of causal skeleton to the case with weak self-masking missingness (i.e., no other variable could be the cause of self-masking indicators except itself). We further provide the sufficient and necessary identification conditions of the causal direction under additive noise model and show that the causal structure can be identified up to an IN-equivalent pattern. We finally propose a practical algorithm based on the above theoretical results on learning the causal skeleton and causal direction. Extensive experiments on synthetic and real data demonstrate the efficiency and effectiveness of the proposed algorithms.

Count data naturally arise in many fields, such as finance, neuroscience, and epidemiology, and discovering causal structure among count data is a crucial task in various scientific and industrial scenarios. One of the most common characteristics of count data is the inherent branching structure described by a binomial thinning operator and an independent Poisson distribution that captures both branching and noise. For instance, in a population count scenario, mortality and immigration contribute to the count, where survival follows a Bernoulli distribution, and immigration follows a Poisson distribution. However, causal discovery from such data is challenging due to the non-identifiability issue: a single causal pair is Markov equivalent, i.e., X->Y and Y->X are distributed equivalent. Fortunately, in this work, we found that the causal order from X to its child Y is identifiable if X is a root vertex and has at least two directed paths to Y, or the ancestor of X with the most directed path to X has a directed path to Y without passing X. Specifically, we propose a Poisson Branching Structure Causal Model (PB-SCM) and perform a path analysis on PB-SCM using high-order cumulants. Theoretical results establish the connection between the path and cumulant and demonstrate that the path information can be obtained from the cumulant. With the path information, causal order is identifiable under some graphical conditions. A practical algorithm for learning causal structure under PB-SCM is proposed and the experiments demonstrate and verify the effectiveness of the proposed method.

Causal effect estimation under networked interference is an important but challenging problem. Available parametric methods are limited in their model space, while previous semiparametric methods, e.g., leveraging neural networks to fit only one single nuisance function, may still encounter misspecification problems under networked interference without appropriate assumptions on the data generation process. To mitigate bias stemming from misspecification, we propose a novel doubly robust causal effect estimator under networked interference, by adapting the targeted learning technique to the training of neural networks. Specifically, we generalize the targeted learning technique into the networked interference setting and establish the condition under which an estimator achieves double robustness. Based on the condition, we devise an end-to-end causal effect estimator by transforming the identified theoretical condition into a targeted loss. Moreover, we provide a theoretical analysis of our designed estimator, revealing a faster convergence rate compared to a single nuisance model. Extensive experimental results on two real-world networks with semisynthetic data demonstrate the effectiveness of our proposed estimators.

Unobserved discrete data are ubiquitous in many scientific disciplines, and how to learn the causal structure of these latent variables is crucial for uncovering data patterns. Most studies focus on the linear latent variable model or impose strict constraints on latent structures, which fail to address cases in discrete data involving non-linear relationships or complex latent structures. To achieve this, we explore a tensor rank condition on contingency tables for an observed variable set $\mathbf{X}_p$, showing that the rank is determined by the minimum support of a specific conditional set (not necessary in $\mathbf{X}_p$) that d-separates all variables in $\mathbf{X}_p$. By this, one can locate the latent variable through probing the rank on different observed variables set, and further identify the latent causal structure under some structure assumptions. We present the corresponding identification algorithm and conduct simulated experiments to verify the effectiveness of our method. In general, our results elegantly extend the identification boundary for causal discovery with discrete latent variables and expand the application scope of causal discovery with latent variables.

Causal discovery from observational data, especially for count data, is essential across scientific and industrial contexts, such as biology, economics, and network operation maintenance. For this task, most approaches model count data using Bayesian networks or ordinal relations. However, they overlook the inherent branching structures that are frequently encountered, e.g., a browsing event might trigger an adding cart or purchasing event. This can be modeled by a binomial thinning operator (for branching) and an additive independent Poisson distribution (for noising), known as Poisson Branching Structure Causal Model (PB-SCM). There is a provably sound cumulant-based causal discovery method that allows the identification of the causal structure under a branching structure. However, we show that there still remains a gap in that there exist causal directions that are identifiable while the algorithm fails to identify them. In this work, we address this gap by exploring the identifiability of PB-SCM using the Probability Generating Function (PGF). By developing a compact and exact closed-form solution for the PGF of PB-SCM, we demonstrate that each component in this closed-form solution uniquely encodes a specific local structure, enabling the identification of the local structures by testing their corresponding component appearances in the PGF. Building on this, we propose a practical algorithm for learning causal skeletons and identifying causal directions of PB-SCM using PGF. The effectiveness of our method is demonstrated through experiments on both synthetic and real datasets.

Learning causal structure among event types on multitype event sequences is an important but challenging task. Existing methods, such as the Multivariate Hawkes processes, mostly assumed that each sequence is independent and identically distributed. However, in many real-world applications, it is commonplace to encounter a topological network behind the event sequences such that an event is excited or inhibited not only by its history but also by its topological neighbors. Consequently, the failure in describing the topological dependency among the event sequences leads to the error detection of the causal structure. By considering the Hawkes processes from the view of temporal convolution, we propose a topological Hawkes process (THP) to draw a connection between the graph convolution in the topology domain and the temporal convolution in time domains. We further propose a causal structure learning method on THP in a likelihood framework. The proposed method is featured with the graph convolution-based likelihood function of THP and a sparse optimization scheme with an Expectation-Maximization of the likelihood function. Theoretical analysis and experiments on both synthetic and real-world data demonstrate the effectiveness of the proposed method.

We study the problem of learning hierarchical causal structure among latent variables from measured variables. While some existing methods are able to recover the latent hierarchical causal structure, they mostly suffer from restricted assumptions, including the tree-structured graph constraint, no ``triangle" structure, and non-Gaussian assumptions. In this paper, we relax these restrictions above and consider a more general and challenging scenario where the beyond tree-structured graph, the ``triangle" structure, and the arbitrary noise distribution are allowed. We investigate the identifiability of the latent hierarchical causal structure and show that by using second-order statistics, the latent hierarchical structure can be identified up to the Markov equivalence classes over latent variables. Moreover, some directions in the Markov equivalence classes of latent variables can be further identified using partially non-Gaussian data. Based on the theoretical results above, we design an effective algorithm for learning the latent hierarchical causal structure. The experimental results on synthetic data verify the effectiveness of the proposed method.

Learning causal structure among event types from discrete-time event sequences is a particularly important but challenging task. Existing methods, such as the multivariate Hawkes processes based methods, mostly boil down to learning the so-called Granger causality which assumes that the cause event happens strictly prior to its effect event. Such an assumption is often untenable beyond applications, especially when dealing with discrete-time event sequences in low-resolution; and typical discrete Hawkes processes mainly suffer from identifiability issues raised by the instantaneous effect, i.e., the causal relationship that occurred simultaneously due to the low-resolution data will not be captured by Granger causality. In this work, we propose Structure Hawkes Processes (SHPs) that leverage the instantaneous effect for learning the causal structure among events type in discrete-time event sequence. The proposed method is featured with the Expectation-Maximization of the likelihood function and a sparse optimization scheme. Theoretical results show that the instantaneous effect is a blessing rather than a curse, and the causal structure is identifiable under the existence of the instantaneous effect. Experiments on synthetic and real-world data verify the effectiveness of the proposed method.

An important problem across multiple disciplines is to infer and understand meaningful latent variables. One strategy commonly used is to model the measured variables in terms of the latent variables under suitable assumptions on the connectivity from the latents to the measured (known as measurement model). Furthermore, it might be even more interesting to discover the causal relations among the latent variables (known as structural model). Recently, some methods have been proposed to estimate the structural model by assuming that the noise terms in the measured and latent variables are non-Gaussian. However, they are not suitable when some of the noise terms become Gaussian. To bridge this gap, we investigate the problem of identification of the structural model with arbitrary noise distributions. We provide necessary and sufficient condition under which the structural model is identifiable: it is identifiable iff for each pair of adjacent latent variables Lx, Ly, (1) at least one of Lx and Ly has non-Gaussian noise, or (2) at least one of them has a non-Gaussian ancestor and is not d-separated from the non-Gaussian component of this ancestor by the common causes of Lx and Ly. This identifiability result relaxes the non-Gaussianity requirements to only a (hopefully small) subset of variables, and accordingly elegantly extends the application scope of the structural model. Based on the above identifiability result, we further propose a practical algorithm to learn the structural model. We verify the correctness of the identifiability result and the effectiveness of the proposed method through empirical studies.

Localizing fault root causes is challenging but critical for wireless network operation and maintenance. Though supervised methods have shown promising results in training samples, most of the existing approaches assume that the training and the testing samples are independent and identical distributed. Such an i.i.d assumption usually does not hold due to network faults that may occur in different devices across different domains (well known as the distribution shift). Thus, it is necessary to align distributions between the training and test data set. Motivated by the stability of the causal mechanism across the domains, a Causal Alignment based Root Cause Localization (CARCL) framework, including the causal alignment and the multi-stage classifier, is proposed. CARCL first offers to align the distributions locally for each causal module but not globally on the complete variable set. We further develop a multi-stage classifier to determine the root causes with the help of predicted pseudo labels. The experiments demonstrate a superior performance of our method.

Transcriptomic and epigenetic alterations during early embryo development have been proven to play essential roles in regulating the cell fate. Nowadays, advances in single-cell transcriptomics and epigenomics profiling techniques provide large volumes of data for understanding the molecular regulatory mechanisms in early embryos and facilitate the investigation of assisted reproductive technology as well as preimplantation genetic testing. However, the lack of integrated data collection and unified analytic procedures greatly limits their usage in scientific research and clinical application. Hence, it is necessary to establish a database integrating the regulatory information of human and mouse early embryos with unified analytic procedures. Here, we introduce DevOmics (http://devomics.cn/), which contains normalized gene expression, DNA methylation, histone modifications (H3K4me3, H3K9me3, H3K27me3, H3K27ac), chromatin accessibility and 3D chromatin architecture profiles of human and mouse early embryos spanning six developmental stages (zygote, 2cell, 4cell, 8cell, morula and blastocyst (ICM, TE)). The current version of DevOmics provides Search and Advanced Search for retrieving genes a researcher is interested in, Analysis Tools including the differentially expressed genes (DEGs) analysis for acquiring DEGs between different types of samples, allelic explorer for displaying allele-specific gene expression as well as epigenetic modifications and correlation analysis for showing the dynamic changes in different layers of data across developmental stages, as well as Genome Browser and Ortholog for visualization. DevOmics offers a user-friendly website for biologists and clinicians to decipher molecular regulatory mechanisms of human and mouse early embryos.

Identification of causal direction between a causal-effect pair from observed data has recently attracted much attention. Various methods based on functional causal models have been proposed to solve this problem, by assuming the causal process satisfies some (structural) constraints and showing that the reverse direction violates such constraints. The nonlinear additive noise model has been demonstrated to be effective for this purpose, but the model class does not allow any confounding or intermediate variables between a cause pair–even if each direct causal relation follows this model. However, omitting the latent causal variables is frequently encountered in practice. After the omission, the model does not necessarily follow the model constraints. As a consequence, the nonlinear additive noise model may fail to correctly discover causal direction. In this work, we propose a confounding cascade nonlinear additive noise model to represent such causal influences–each direct causal relation follows the nonlinear additive noise model but we observe only the initial cause and final effect. We further propose a method to estimate the model, including the unmeasured confounding and intermediate variables, from data under the variational auto-encoder framework. Our theoretical results show that with our model, the causal direction is identifiable under suitable technical conditions on the data generation process. Simulation results illustrate the power of the proposed method in identifying indirect causal relations across various settings, and experimental results on real data suggest that the proposed model and method greatly extend the applicability of causal discovery based on functional causal models in nonlinear cases.

BACKGROUND: Haplotyping reveals chromosome blocks inherited from parents to in vitro fertilized (IVF) embryos in preimplantation genetic diagnosis (PGD), enabling the observation of the transmission of disease alleles between generations. However, the methods of haplotyping that are suitable for single cells are limited because a whole genome amplification (WGA) process is performed before sequencing or genotyping in PGD, and true haplotype profiles of embryos need to be constructed based on genotypes that can contain many WGA artifacts. RESULTS: Here, we offer scHaplotyper as a genetic diagnosis tool that reconstructs and visualizes the haplotype profiles of single cells based on the Hidden Markov Model (HMM). scHaplotyper can trace the origin of each haplotype block in the embryo, enabling the detection of carrier status of disease alleles in each embryo. We applied this method in PGD in two families affected with genetic disorders, and the result was the healthy live births of two children in the two families, demonstrating the clinical application of this method. CONCLUSION: Next generation sequencing (NGS) of preimplantation embryos enable genetic screening for families with genetic disorders, avoiding the birth of affected babies. With the validation and successful clinical application, we showed that scHaplotyper is a convenient and accurate method to screen out embryos. More patients with genetic disorder will benefit from the genetic diagnosis of embryos. The source code of scHaplotyper is available at GitHub repository: https://github.com/yzqheart/scHaplotyper.

Chromosome enumeration is an essential but tedious procedure in karyotyping analysis. To automate the enumeration process, we develop a chromosome enumeration framework, DeepACEv2, based on the region based object detection scheme. The framework is developed following three steps. Firstly, we take the classical ResNet-101 as the backbone and attach the Feature Pyramid Network (FPN) to the backbone. The FPN takes full advantage of the multiple level features, and we only output the level of feature map that most of the chromosomes are assigned to. Secondly, we enhance the region proposal network's ability by adding a newly proposed Hard Negative Anchors Sampling to extract unapparent but essential information about highly confusing partial chromosomes. Next, to alleviate serious occlusion problems, besides the traditional detection branch, we novelly introduce an isolated Template Module branch to extract unique embeddings of each proposal by utilizing the chromosome's geometric information. The embeddings are further incorporated into the No Maximum Suppression (NMS) procedure to improve the detection of overlapping chromosomes. Finally, we design a Truncated Normalized Repulsion Loss and add it to the loss function to avoid inaccurate localization caused by occlusion. In the newly collected 1375 metaphase images that came from a clinical laboratory, a series of ablation studies validate the effectiveness of each proposed module. Combining them, the proposed DeepACEv2 outperforms all the previous methods, yielding the Whole Correct Ratio(WCR)(%) with respect to images as 71.39, and the Average Error Ratio(AER)(%) with respect to chromosomes as about 1.17.

Domain adaptation is an important but challenging task. Most of the existing domain adaptation methods struggle to extract the domain-invariant representation on the feature space with entangling domain information and semantic information. Different from previous efforts on the entangled feature space, we aim to extract the domain invariant semantic information in the latent disentangled semantic representation (DSR) of the data. In DSR, we assume the data generation process is controlled by two independent sets of variables, i.e., the semantic latent variables and the domain latent variables. Under the above assumption, we employ a variational auto-encoder to reconstruct the semantic latent variables and domain latent variables behind the data. We further devise a dual adversarial network to disentangle these two sets of reconstructed latent variables. The disentangled semantic latent variables are finally adapted across the domains. Experimental studies testify that our model yields state-of-the-art performance on several domain adaptation benchmark datasets.

Identification of causal direction between a causal-effect pair from observed data has recently attracted much attention. Various methods based on functional causal models have been proposed to solve this problem, by assuming the causal process satisfies some (structural) constraints and showing that the reverse direction violates such constraints. The nonlinear additive noise model has been demonstrated to be effective for this purpose, but the model class is not transitive--even if each direct causal relation follows this model, indirect causal influences, which result from omitted intermediate causal variables and are frequently encountered in practice, do not necessarily follow the model constraints; as a consequence, the nonlinear additive noise model may fail to correctly discover causal direction. In this work, we propose a cascade nonlinear additive noise model to represent such causal influences--each direct causal relation follows the nonlinear additive noise model but we observe only the initial cause and final effect. We further propose a method to estimate the model, including the unmeasured intermediate variables, from data, under the variational auto-encoder framework. Our theoretical results show that with our model, causal direction is identifiable under suitable technical conditions on the data generation process. Simulation results illustrate the power of the proposed method in identifying indirect causal relations across various settings, and experimental results on real data suggest that the proposed model and method greatly extend the applicability of causal discovery based on functional causal models in nonlinear cases.

Causal discovery without intervention is well recognized as a challenging yet powerful data analysis tool, boosting the development of other scientific areas, such as biology, astronomy, and social science. The major technical difficulty behind the observation-based causal discovery is to effectively and efficiently identify causes and effects from correlated variables given the existence of significant noises. Previous studies mostly employ two very different methodologies under Bayesian network framework, namely global likelihood maximization and locally complexity analysis over marginal distributions. While these approaches are effective in their respective problem domains, in this paper, we show that they can be combined to formulate a new global optimization model with local statistical significance, called structural equational likelihood framework (or SELF in short). We provide thorough analysis on the soundness of the model under mild conditions and present efficient heuristic-based algorithms for scalable model training. Empirical evaluations using XGBoost validate the superiority of our proposal over state-of-the-art solutions, on both synthetic and real world causal structures.

Causal discovery from a set of observations is one of the fundamental problems across several disciplines. For continuous variables, recently a number of causal discovery methods have demonstrated their effectiveness in distinguishing the cause from effect by exploring certain properties of the conditional distribution, but causal discovery on categorical data still remains to be a challenging problem, because it is generally not easy to find a compact description of the causal mechanism for the true causal direction. In this paper we make an attempt to find a way to solve this problem by assuming a two-stage causal process: the first stage maps the cause to a hidden variable of a lower cardinality, and the second stage generates the effect from the hidden representation. In this way, the causal mechanism admits a simple yet compact representation. We show that under this model, the causal direction is identifiable under some weak conditions on the true causal mechanism. We also provide an effective solution to recover the above hidden compact representation within the likelihood framework. Empirical studies verify the effectiveness of the proposed approach on both synthetic and real-world data.