Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
VIME
Variational Imputation and Mask Estimation, a missing value imputation framework that uses variational inference to jointly model the data distribution and the missingness mask.
Missingness Mask
Binary matrix indicating the presence (1) or absence (0) of data for each observation and variable, whose distribution is modeled in VIME.
Variational Autoencoder (VAE)
Generative neural network architecture that learns a probabilistic latent representation of the data, used in VIME to model the complete data distribution.
MAR Mechanism (Missing At Random)
Assumption where the probability of a value being missing depends only on observed values, conditionally satisfied by the VIME framework.
MNAR Mechanism (Missing Not At Random)
Scenario where the probability of missingness depends on the missing values themselves, which VIME can model through its joint estimation of the mask.
Importance Resampling
Monte Carlo estimation technique used in VIME to approximate the expectation under the variational distribution with reduced variance.
Posterior Predictive
Distribution of future data that incorporates uncertainty about the model parameters, which VIME leverages to generate realistic imputations.
Amortized Inference
Strategy where variational parameters are produced by a neural network (encoder) rather than optimized individually, enabling efficient processing of large datasets.
Decomposition of the Variational Bound
Analysis of the ELBO in distinct terms (reconstruction, latent regularization, mask prediction) that guides the architecture and training of the VIME model.
Autoregressive Imputation
Method where missing values are imputed sequentially by conditioning on previous imputations, an alternative to VIME's variational approach.
Deep Generative Model
Class of models that learn the probability distribution of data, of which VIME is a specialized example for structured data with missing values.