Causal Effect Identifiability under Partial-Observability

ICML, pp. 5692-5701, 2020.

Cited by: 0|Bibtex|Views3
EI
Other Links: academic.microsoft.com|dblp.uni-trier.de
Weibo:
We developed a new algorithm that efficiently and systematically examines the identifiability of embedding factors and combines the identified minimum viable embedding factor to compose the expression for a given query

Abstract:

Causal effect identifiability is concerned with establishing the effect of intervening on a set of variables on another set of variables from observational or interventional distributions under causal assumptions that are usually encoded in the form of a causal graph. Most of the results of this literature implicitly assume that every var...More

Code:

Data:

0
Introduction
  • One of the central goals in data sciences, artificial intelligence, and machine learning is to discover cause and effect relationships.
  • If the scientific study is performed appropriately, and causal relations are eventually discovered, the corresponding effects are more likely to hold under a broader set of conditions.
  • Causal relations are usually more stable and generalizable across disparate conditions (Pearl, 2000; Pearl & Mackenzie, 2018).
  • Relevant data available to answer a question about a specific effect may be not usable in another, possibly very related study
Highlights
  • One of the central goals in data sciences, artificial intelligence, and machine learning is to discover cause and effect relationships
  • We introduced the general identifiability problem when the available distributions are only partially observable, which is named GID-PO
  • We investigated how a causal query can be factorized under different levels of projections, and introduced new constructs called embedding factors and co-identification. These constructs make explicit the connection of the factors required to identify the targeted query and the available observed distributions, which allows a systematic view of the problem of identifiability under different granularities
  • We introduced a new graphical structure called minimum viable embedding factor (MVEF) and studied its properties, including its uniqueness, disjointedness, and compositionality
  • We developed a new algorithm (GID-PO) that efficiently and systematically examines the identifiability of embedding factors and combines the identified minimum viable embedding factor to compose the expression for a given query
  • Since each of the factors cannot be identified from smaller marginal interventional distributions, we conjecture that the procedure is necessary
Conclusion
  • The authors introduced the general identifiability problem when the available distributions are only partially observable, which is named GID-PO.
  • The authors introduced a new graphical structure called minimum viable embedding factor (MVEF) and studied its properties, including its uniqueness, disjointedness, and compositionality
  • Putting these results together, the authors developed a new algorithm (GID-PO) that efficiently and systematically examines the identifiability of embedding factors and combines the identified MVEFs to compose the expression for a given query.
  • In practice, available datasets are measured inconsistently with respect to the variables they cover – they usually have different columns – the authors hope the results in this paper can help data scientists tackle more challenging identification instances and determine causal effects in more intricate and realistic scenarios
Summary
  • Introduction:

    One of the central goals in data sciences, artificial intelligence, and machine learning is to discover cause and effect relationships.
  • If the scientific study is performed appropriately, and causal relations are eventually discovered, the corresponding effects are more likely to hold under a broader set of conditions.
  • Causal relations are usually more stable and generalizable across disparate conditions (Pearl, 2000; Pearl & Mackenzie, 2018).
  • Relevant data available to answer a question about a specific effect may be not usable in another, possibly very related study
  • Conclusion:

    The authors introduced the general identifiability problem when the available distributions are only partially observable, which is named GID-PO.
  • The authors introduced a new graphical structure called minimum viable embedding factor (MVEF) and studied its properties, including its uniqueness, disjointedness, and compositionality
  • Putting these results together, the authors developed a new algorithm (GID-PO) that efficiently and systematically examines the identifiability of embedding factors and combines the identified MVEFs to compose the expression for a given query.
  • In practice, available datasets are measured inconsistently with respect to the variables they cover – they usually have different columns – the authors hope the results in this paper can help data scientists tackle more challenging identification instances and determine causal effects in more intricate and realistic scenarios
Funding
  • This research is supported in parts by grants from NSF (IIS-1704352 and IIS-1750807 (CAREER))
Reference
  • Bareinboim, E. and Pearl, J. Causal inference by surrogate experiments: z-identifiability. In de Freitas, N. and Murphy, K. (eds.), Proceedings of the Twenty-Eighth Conference on Uncertainty in Artificial Intelligence, pp. 113– 120, Corvallis, OR, 2012a. AUAI Press.
    Google ScholarLocate open access versionFindings
  • Bareinboim, E. and Pearl, J. Controlling selection bias in causal inference. In Girolami, M. and Lawrence, N. (eds.), Proceedings of The Fifteenth International Conference on Artificial Intelligence and Statistics (AISTATS 2012), pp. 100–108. JMLR (22), 2012b.
    Google ScholarLocate open access versionFindings
  • Bareinboim, E. and Pearl, J. Transportability from multiple environments with limited experiments: Completeness results. In Advances in Neural Information Processing Systems 27, pp. 280–288. Curran Associates, Inc., 2014.
    Google ScholarLocate open access versionFindings
  • Bareinboim, E. and Pearl, J. Causal inference and the datafusion problem. Proceedings of the National Academy of Sciences, 113:7345–7352, 2016.
    Google ScholarLocate open access versionFindings
  • Correa, J. and Bareinboim, E. Causal effect identification by adjustment under confounding and selection biases. In Proceedings of the 31st AAAI Conference on Artificial Intelligence, pp. 3740–3746. AAAI Press, 2017.
    Google ScholarLocate open access versionFindings
  • Huang, Y. and Valtorta, M. Identifiability in causal bayesian networks: A sound and complete algorithm. In Proceedings of the Twenty-First National Conference on Artificial Intelligence (AAAI 2006), pp. 1149–115AAAI Press, Menlo Park, CA, 2006a.
    Google ScholarLocate open access versionFindings
  • Huang, Y. and Valtorta, M. Pearl’s calculus of intervention is complete. In Dechter, R. and Richardson, T. (eds.), Proceedings of the Twenty-Second Conference on Uncertainty in Artificial Intelligence (UAI 2006), pp. 217–224. AUAI Press, Corvallis, OR, 2006b.
    Google ScholarLocate open access versionFindings
  • Jaber, A., Zhang, J., and Bareinboim, E. Causal identification under Markov equivalence: Completeness results. In Proceedings of the 36th International Conference on Machine Learning, volume 97, pp. 2981–2989. PMLR, 2019.
    Google ScholarLocate open access versionFindings
  • Karp, R. M. Reducibility among combinatorial problems. In Miller R.E., Thatcher J.W., B. J. (ed.), Complexity of Computer Computations, The IBM Research Symposia Series, pp. 85–103. Plenum Press, New York, 1972.
    Google ScholarLocate open access versionFindings
  • Lee, J. J. R. and Shpitser, I. Identification methods with arbitrary interventional distributions as inputs, 2020. arXiv:2004.01157.
    Findings
  • Lee, S., Correa, J. D., and Bareinboim, E. General identifiability with arbitrary surrogate experiments. In Proceedings of the Thirty-Fifth Conference Annual Conference on Uncertainty in Artificial Intelligence, Corvallis, OR, 2019. AUAI Press.
    Google ScholarLocate open access versionFindings
  • Lee, S., Correa, J. D., and Bareinboim, E. Generalized transportability: Synthesis of experiments from heterogeneous domains. In Proceedings of the 34th AAAI Conference on Artificial Intelligence. AAAI Press, 2020.
    Google ScholarLocate open access versionFindings
  • Mohan, K. and Pearl, J. On the testability of models with missing data. In Proceedings of The Seventeenth International Conference on Artificial Intelligence and Statistics. JMLR, 2014.
    Google ScholarLocate open access versionFindings
  • Pearl, J. Causal diagrams for empirical research. Biometrika, 82(4):669–688, 1995.
    Google ScholarLocate open access versionFindings
  • Pearl, J. Causality: Models, Reasoning, and Inference. Cambridge University Press, New York, 2000.
    Google ScholarFindings
  • Pearl, J. and Bareinboim, E. Transportability of causal and statistical relations: A formal approach. Technical Report Technical Report r-372, Cognitive Systems Laboratory, Department of Computer Science, UCLA, 2011.
    Google ScholarFindings
  • Pearl, J. and Mackenzie, D. The Book of Why: The New Science of Cause and Effect. Basic Books, 2018.
    Google ScholarFindings
  • Shpitser, I. and Pearl, J. Identification of joint interventional distributions in recursive semi-Markovian causal models. In Proceedings of The Twenty-First National Conference on Artificial Intelligence, pp. 1219–1226. AAAI Press, 2006a.
    Google ScholarLocate open access versionFindings
  • Shpitser, I. and Pearl, J. Identification of conditional interventional distributions. In Dechter, R. and Richardson, T. (eds.), Proceedings of the Twenty-Second Conference on Uncertainty in Artificial Intelligence (UAI 2006), pp. 437–444. AUAI Press, Corvallis, OR, 2006b.
    Google ScholarLocate open access versionFindings
  • Spirtes, P., Glymour, C., and Scheines, R. Causation, Prediction, and Search. MIT Press, Cambridge, MA, 2nd edition, 2001.
    Google ScholarFindings
  • Tian, J. Studies in Causal Reasoning and Learning. PhD thesis, Computer Science Department, University of California, Los Angeles, CA, November 2002.
    Google ScholarFindings
  • Tian, J. Identifying conditional causal effects. In Proceedings of the 20th Conference on Uncertainty in Artificial Intelligence, pp. 561–568. AUAI Press, 2004.
    Google ScholarLocate open access versionFindings
  • Tian, J. and Pearl, J. A general identification condition for causal effects. In Proceedings of the Eighteenth National Conference on Artificial Intelligence (AAAI 2002), pp. 567–573, Menlo Park, CA, 2002. AAAI Press/The MIT Press.
    Google ScholarLocate open access versionFindings
  • Tian, J. and Pearl, J. On the identification of causal effects. Technical Report R-290-L, Department of Computer Science, University of California, Los Angeles, CA, 2003.
    Google ScholarFindings
Full Text
Your rating :
0

 

Tags
Comments