Entropic Causal Inference
Authors: Murat Kocaoglu, Alexandros Dimakis, Sriram Vishwanath, Babak Hassibi
AAAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we test the performance of our algorithms on real and artificial data. First, we test the greedy entropy minimization algorithm and show that it performs close to the trivial lower bound. Then, we test our conjecture of identifiability using entropy. Lastly, we test our entropy-minimization based causal identification technique on real data. |
| Researcher Affiliation | Academia | Murat Kocaoglu, Alexandros G. Dimakis, Sriram Vishwanath Department of Electrical and Computer Engineering, The University of Texas at Austin, USA Babak Hassibi Department of Electrical Engineering California Institute of Technology, USA |
| Pseudocode | Yes | Algorithm 1 Joint Entropy Minimization Algorithm 1: Input: Marginal distributions of m variables each with n states, in matrix form M = [p T 1 ; p T 2 ; ..., p T m]. 2: e = [ ] 3: Sort each row of M in decreasing order. 4: Find minimum of maximum of each row: r mini(pi(1)) 5: while r > 0 do 6: e [e, r] 7: Update maximum of each row: pi(1) pi(1) r, i 8: Sort each row of M in decreasing order. 9: r mini(pi(1)) 10: end while 11: return e. |
| Open Source Code | No | The paper does not provide concrete access to source code. It only links to an arXiv preprint for a version with proofs in a footnote: '1For a version with proofs, see https://arxiv.org/abs/1611.04035'. |
| Open Datasets | Yes | We test our entropy-based causal inference algorithm on the Cause Effect Pairs repository (Mooij et al. 2016a). |
| Dataset Splits | No | The paper mentions testing and real data, but does not provide specific dataset split information (exact percentages, sample counts, or detailed splitting methodology) for training, validation, or testing. |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types, or computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment. |
| Experiment Setup | Yes | The challenge in applying our framework on this dataset is choosing the correct quantization. Small number of quantization levels may result in loss of information regarding the joint distribution, and a very large number of states might be computationally hard to work with. We pick the same number of states for both X and Y , and use a uniform quantization that assures each state of the variables has 10 samples on average. From the samples, we estimate the conditonal transition matrices Y|X and X|Y and feed the columns to the greedy entropy minimization algorithm (Algorithm 1), which outputs an approximate of the smallest entropy exogenous variable. Later we compare H(X, E) and H(Y, E) and declare the model with smallest input entropy to be the true model, based on Conjecture 1. For a causal pair, we invoke the algorithm if |H(X, E) H(Y, E)| t log(n) for threshold parameter t, which determines the decision rate. |