Campus Event Calendar

Event Entry

What and Who

Causal Inference on Discrete Data

Kailash Budhathoki
Cluster of Excellence - Multimodal Computing and Interaction - MMCI
AG 1, AG 2, AG 3, INET, AG 4, AG 5, SWS, RG1, MMCI  
Public Audience

Date, Time and Location

Monday, 6 July 2020
60 Minutes
E1 4


Causal inference is one of the fundamental problems in science. A particularly interesting setting is to tell cause from effect between a pair of random variables X and Y given a sample from their joint distribution. For a long period of time, it was thought to be impossible to distinguish between causal structures X → Y and Y → X from observational data as the factorisation of the joint distribution is the same in both directions. In the past decade, researchers have made a long stride in this direction by exploiting sophisticated properties of the joint distribution. Most of the existing methods, however, are for continuous real-valued data.

In the first part of the thesis, we consider bivariate causal inference on different discrete data settings—univariate i.i.d., univariate non-i.i.d., and multivariate i.i.d. pairs. To this end, we build upon the principle of algorithmic independence of conditionals (AIC), which states that marginal distribution of the cause is algorithmically independent of conditional distribution of the effect given the cause. However, as Kolmogorov complexity is not computable, we approximate the AIC from above through the statistically sound Minimum Description Length (MDL) principle. On univariate i.i.d. and non-i.i.d. pairs, where causal mechanisms are simple, we use refined MDL codes that are minimax optimal w.r.t. a model class. We resort to crude MDL codes on a pair of multivariate i.i.d. variables.
Although useful, saying that there exists a causal relationship from a set of variables towards a certain variable of interest does not always fully satisfy one’s curiosity; for a domain expert it is of particular interest to know those conditions that are most effective, such as the combinations of drugs and their dosages that are most effective towards recovery. Motivated by this problem, in the second part of this thesis, we consider discovering statistically reliable causal rules from observational data. Overall, extensive evaluations show that methods proposed in this thesis are highly accurate, and discover meaningful causations from real-world data.


Petra Schaaf
+49 681 9325 5000
--email hidden

Video Broadcast

E1 4
meeting ID: 968 9584 1029 password: Dr-1ng

Petra Schaaf, 06/26/2020 10:55 AM
Petra Schaaf, 06/26/2020 10:05 AM -- Created document.