Parameter Estimation in DAGs from Incomplete Data via Optimal Transport

Authors: Vy Vo, Trung Le, Long Tung Vuong, He Zhao, Edwin V. Bonilla, Dinh Phung

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We further provide empirical evidence demonstrating the versatility of our method on various graphical structures, where OTP-DAG is shown to successfully recover the ground-truth parameters and achieve comparable or better performance than competing methods across a range of downstream applications. 5. Applications In this section, we illustrate the practical applications of the OTP-DAG algorithm. We consider various directed probabilistic models with different types of latent variables (continuous and discrete) and for different types of data (texts, images, and time series). In all tables, we report the average results over 5 random initializations and the best/second-best ones are bold/underlined. , indicate higher/lower performance is better, respectively.
Researcher Affiliation Collaboration 1Monash University, Australia 2CSIRO s Data61, Australia 3Vin AI Research, Vietnam.
Pseudocode Yes Algorithm 1 OTP-DAG Algorithm
Open Source Code Yes 1Our code is published at https://github.com/ is Vy08/OTP.
Open Datasets Yes We here use OTPDAG to infer the topics of 3 real-world datasets: 20 News Group3, BBC News (Greene & Cunningham, 2006) and DBLP4. We experiment with images in this application and compare OTP-DAG with VQ-VAE on CIFAR105, MNIST (Le Cun et al., 1998), SVHN (Netzer et al., 2011) and CELEBA datasets (Liu et al., 2015).
Dataset Splits No The paper does not explicitly provide specific training/validation/test dataset splits, only mentioning dataset generation and number of training epochs.
Hardware Specification Yes All models are run on 4 RTX 6000 GPU cores using Adam optimizer with a fixed learning rate of 1e 3.
Software Dependencies No The paper mentions software components like 'Adam optimizer' and implies the use of 'PythonOT' and neural network architectures (LSTM, perceptron), but does not provide specific version numbers for programming languages or libraries.
Experiment Setup Yes The hyperparameters are: D = M = 64, K = 512, η = 1e 3, ηr = 1.0, batch size of 32 and 100 training epochs.