Local and global causal discovery methods for observational data with discrete variables

Vetró, Mihály and Hullám, Gábor (2021) Local and global causal discovery methods for observational data with discrete variables. In: Proceedings of the 28th PhD Mini-Symposium. BME MIT, Budapest, pp. 40-43.

Preview

Text Vetro_Hullam_CausalDiscovery_MiniSympMIT2021.pdf Download (2MB) | Preview

Abstract

Identifying causal relationships based on observational data is challenging, because in the absence of interventions (i.e. experiments), causal discovery algorithms have to tackle the problem of observational equivalence. The latter means that causal relationships cannot be inferred from dependency relationships unambiguously, except for certain cases. Thus, the causal interpretation of dependency patterns between variables is limited, and in most cases it is only possible with adequate background knowledge, such as temporal information. To address the problem, a number of methods have emerged that follow either a global or a local approach. The former seek to reconstruct the entire dependency structure based on the available data and apply a causal interpretation if possible, while the latter examine small local units of the whole dependency structure and infer local causal relationships. In this paper, we investigate selected local and global causal discovery algorithms and compare their performance on artificial datasets. Results indicate that the accuracy of local causal discovery methods is lower than that of global methods, which is inline with expectations as not all types of causal relationships can be identified locally. On the other hand, the causal direction of discovered relationships appear to be reliable.

Actions (login required)

Edit Item