Operationalizing Complex Causes: A Pragmatic View of Mediation
Authors: Limor Gultchin, David Watson, Matt Kusner, Ricardo Silva
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate, on a range of simulated and real-world-inspired examples, that our approach allows us to efficiently estimate the effect of crude interventions with limited data from new treatment regimes. and 4. Experiments In this section, we demonstrate our method in a variety of domains. |
| Researcher Affiliation | Academia | 1Department of Computer Science, University of Oxford, Oxford, UK 2The Alan Turing Institute, London, UK 3Department of Statistical Science, University of College London, London, UK 4Department of Computer Science, University of College London, London, UK. |
| Pseudocode | Yes | Algorithm 1 Causal Response Prediction and Algorithm 2 Pragmatic Mediation Selection |
| Open Source Code | Yes | The code to reproduce results can be found at https: //github.com/limorigu/Complex Causes. |
| Open Datasets | Yes | As a second example, we consider a dataset from a computational humor experiment. Participants were given news headlines and asked to make single entity changes such that the resulting headline would be humorous (Hossain et al., 2019). This work was further extended into a Sem Eval2020 task, and full datasets were made publicly available.7 and Dialogue for Reverse Engineering Assessments and Methods (DREAM) challenge (Marbach et al., 2012). and GloVe vector representations (Pennington et al., 2014) |
| Dataset Splits | No | We assume access to m 1 mutually independent historic training regimes with corresponding labeled datasets Dl1, . . . , Dlm 1 and In this regime, we are given access to limited labeled training data Dlw = {(W i , Xi, Yi, Zi)lw } |Dlw | i=1 and more unlabeled training data Duw = {(W i , Xi, Zi)uw } |Duw | i=1 , where |Duw | |Dlw |. This captures settings where measurements for Y are expensive, delayed, or simply unrecorded. All methods are evaluated on a test dataset Tw = {(W i , Yi, Zi)tw }|Tw | i=1 . The term validation is not used for a dataset split. |
| Hardware Specification | No | No specific hardware details (like GPU/CPU models or memory) are mentioned for the experimental setup. |
| Software Dependencies | No | We use GENIE3 (Huynh Thu et al., 2010), a leading gene regulatory network inference algorithm based on random forests, to fit the 4177 structural equations that govern this system. and We performed a k-means clustering on GloVe vector representations (Pennington et al., 2014) and lasso regression (linear), support vector regression (SVR), random forest (RF), and gradient boosting (GB). No version numbers are provided. |
| Experiment Setup | Yes | Significance levels for all tests were fixed at α = 0.01, with p-values adjusted for multiple testing via Holm (1979) s method. and Default hyperparameters are used throughout; see Appendix for details. |