Assessing a Mixture Model for Clustering with the Integrated Completed Likelihood

By Christophe Biernacki, Gilles Celeux, Gérard Govaert. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2000

1,616 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Epidemiology, Statistics and Probability

View Paper: https://www.semanticscholar.org/paper/45d0aa676a7418eab6dbeb33ec3908709b7dc79b

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Epidemiology

In Statistics and Probability

Citations in the corpus

Time-series clustering - A decade review (2015, background): “Additionally, there are a few articles which use model based clustering of time-series data which are composed of polynomial models [112], Gaussian mixed models [183], ARIMA [106], Markov chain [68] and Hidden Markov models [184,185].”
Hierarchical block structures and high-resolution model selection in large networks (2013, methodology): “Now, instead of ﬁnding the model parameters that maximize this quantity, we compute the inte-13 grated likelihood [38, 42, 55], P ( G, { b i }| B ) = d p rs d p r P ( G, { b i } , { p rs } , { p r }| B ) = d p rs P ( G”
Finite mixture models and model-based clustering (2010, background): “The authors suggest postprocessing the results using ICL in a second stage of their procedure to eliminate unnecessary components, and merging them hierarchically.”
Model-Based Clustering (2016, methodology): “To this end, alternatives such as the integrated completed likelihood (ICL; Biernacki, Celeux, and Govaert 2000) are sometimes considered.”
Community detection in graphs (2009, background): “The algorithm runs for a fixed number of clusters q, like that by Newman and Leicht; however, the optimal number of clusters can be determined by running the algorithm for a range of q-values and selecting the solution that maximizes the Integrated Classification Likelihood introduced by Biernacki e”
Clustering of time series data - a survey (2005, methodology): “[33] Multiple Gaussian mixture Parameters Log-likelihood EM algorithm Log-likelihood Non-specific”
LAF-Net: Locally Adaptive Fusion Networks for Stereo Confidence Estimation (2019, background): “Since there is no single conﬁdence feature that yields stably optimal performance, various approaches to beneﬁt from the feature combination among a different set of single conﬁdence measures have been proposed [8, 39] which trained a shallow classiﬁer such as random decision forest [1, 22].”
Finding the Number of Clusters in a Dataset (2003, methodology): “(1998), Frayley and Rafery (1998), and Biernacki et al. (2000), who developed Bayesian methods for choosing the number of components in a Gaussian mixture distribution. These approaches differ from ours both in that they are model-based and that they make no explicit use of the distortion curve. How”
Model-based clustering of categorical time series (2010, methodology): “BIC(H) is known to be consistent for the number of components, if the component density is correctly specified (Keribin 2000), although in small data sets it tends to choose models with too few components (Biernacki et al. 2000).”
Parsimonious Gaussian mixture models (2008, neutral): “But Ψ new g is a diagonal matrix, therefore ˆ Ψ new g = diag S g − 2 ˆ Λ newˆβ g S g + ˆ Λ new Θ g ( ˆ Λ new ) .”
Extending the Akaike Information Criterion to Mixture Regression Models (2007, methodology): “Presently, users can apply extant approaches (e.g., Biernacki, Celeux, and Govaert 2000; Tibshirani, Walther, and Hastie 2001; Tadesse, Sha, and Vannucci 2005) to determine the number of components (but not variables) or information criteria, such as the Akaike information criterion (AIC) (Akaike…”
Distributional assumptions of growth mixture models: implications for overextraction of latent trajectory classes. (2003, methodology): “Finally, the integrated completed likelihood criterion (ICL and ICL–BIC; Biernacki, Celeux, & Govaert, 2000 ) penalizes for both the number of parameters and the quality of classification, making it a more conservative selection criterion (see the Appendix for further details on these fit measures)”