Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Causal Mixture Models: Characterization and Discovery
Authors: Sarah Mameche, Janis Kalofolias, Jilles Vreeken
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 6 Evaluation We evaluate our approach on two main questions, (i) Discovering Mixing Structure: Given a causal graph, can our approach accurately discover the underlying mixing structure, including the number of mixing variables, the number of their components, assignments, and sets of targeted observed variables? (ii) Discovering Causal Structure: When the causal graph is unknown, can our approach recover the mixing structure as well as the causal graph? We address the above questions on synthetic and real-world data. |
| Researcher Affiliation | Academia | Sarah Mameche CISPA Helmholtz Center for Information Security EMAIL Janis Kalofolias CISPA Helmholtz Center for Information Security EMAIL Jilles Vreeken CISPA Helmholtz Center for Information Security EMAIL |
| Pseudocode | Yes | Algorithm 1: DISCOVER A CAUSAL MIXTURE MODEL (CMM) Input: Dataset X, max. number of mixture components Kmax Output: Set of latent variables Z, causal graph GZ 1 Initialize Z , GZ , G , T [ ]; // Discover local mixing and graph 2 while not all nodes are ordered do 3 Xj INFERSOURCE(T, G); 4 G EDGEADDITIONS(Xj, G); 5 G EDGEPRUNING(Xj, G); 6 for each k in 1, . . . , Kmax do 7 Using EM, fit (Xj|Paj = y) MLR Bj, γj, σ2 with k components; 8 Zj mixing assignment with best BICEM Z = max1 K Kmax BICEM Z (K); 9 T.APPEND(Xj); // Infer global mixing 10 Z, GZ INFERMIXING(G, {Zj}); 11 return G, Z; |
| Open Source Code | Yes | Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: The code for running the proposed algorithm, the baselines shown, as well as for reproducing all experiments shown in the evaluation will be included in the supplementary material, with reasonable guidelines (scripts for reproducing the data used in each Figure). |
| Open Datasets | Yes | Case Study: Flow Cytometry Data Finally, we investigate the real-world flow cytometry dataset curated by Sachs et al., 2005. ... Case Study: Mixtures of Interventions Next, we replicate the experimental setup of Kumar et al., 2024 using their data generators to evaluate performance in the mixtures-of-interventions setting. |
| Dataset Splits | Yes | For example, for K=2, we uniformly at random draw S SZ samples belong class 0, otherwise assign class 1. ... As in previous analyses [Wang et al., 2017b], we combine the data from all experimental conditions into a larger dataset of size 5846, and do not disclose their origin to the algorithm. ... For the mixture of interventions, we have NZ = 1 with K = NX + 1 classes which defines a split into one observational and K interventional datasets. |
| Hardware Specification | Yes | We ran the evaluations on an 11th Gen Intel Core i9 CPU. |
| Software Dependencies | No | We apply all baselines without optimization of their hyperparameters using their implementations available in the causal-learn, causal discovery toolbox (cdt), causal Disco and dodiscover Python libraries. |
| Experiment Setup | Yes | The experiments address the effect of several parameters: the number of observed NX {5, . . . 20} and latent variables NZ {0, . . . , 10}, number of latent classes K {2, . . . , 5}, fraction of observed variables p Z [0, 1] affected by mixing, dag density p G [0, 1], sample size S {200, ...1000} and a parameter controlling the size of the samples in each group SZ {2, . . . , 10}; for example, for K=2, we uniformly at random draw S SZ samples belong class 0, otherwise assign class 1. By default, we show results for NX = 10, NZ = 4, K = 2, p Z = 0.5, p G = 0.4, S = 500, SZ = 5. |