## Penalized Clustering of Large Scale Functional Data with Multiple Covariates

Citations: | 2 - 0 self |

### BibTeX

@MISC{Ma_penalizedclustering,

author = {Ping Ma and Wenxuan Zhong},

title = {Penalized Clustering of Large Scale Functional Data with Multiple Covariates},

year = {}

}

### OpenURL

### Abstract

In this article, we propose a penalized clustering method for large scale data with multiple covariates through a functional data approach. In the proposed method, responses and covariates are linked together through nonparametric multivariate functions (fixed effects), which have great flexibility in modeling a variety of function features, such as jump points, branching, and periodicity. Functional ANOVA is employed to further decompose multivariate functions in a reproducing kernel Hilbert space and provide associated notions of main effect and interaction. Parsimonious random effects are used to capture various corre-lation structures. The mixed-effect models are nested under a general mixture model, in which the heterogeneity of functional data is characterized. We pro-pose a penalized Henderson’s likelihood approach for model-fitting and design a rejection-controlled EM algorithm for the estimation. Our method selects

### Citations

8542 | Maximum likelihood from incomplete data via the EM algorithm
- Dempster, Laird, et al.
- 1977
(Show Context)
Citation Context ... Jik[(y i−µk(xi)−Zibi) T (y i−µk(xi)−Zibi)+σ 2 b T i B −1 k bi]+ where λk is the smoothing parameter for µk. K� NλkM(µk) Once the penalized Henderson’s likelihood (3.2) is obtained, the EM algorithm (=-=Dempster et al. 1977-=-, Green 1990) can be derived as follows. The E-step simply requires the calculation of wik = pkϕ(yi; µk(xi), Σk) �K l=1 plϕ(yi; µl(xi), Σl) where Σk = ZiBkZ T i + σ 2 I, and ϕ is the Gaussian density ... |

1436 | Generalized additive models
- Hastie, Tibshirani
- 1990
(Show Context)
Citation Context ...issue in using BIC in nonparametric settings is to determine the effective number of parameters. Here we use the trace of the smoothing matrix to approximate the number of parameters in each cluster (=-=Hastie and Tibshirani 1990-=-, Gu 2002). Thus BIC under our model is BIC = −2 n� K� K� log pkϕ(yi; µk(xi), Σk) + ( trAk(λk, Ωk) + P ) log N, (3.11) i=1 k=1 where Ak is the smoothing matrix for the kth cluster as defined in (2.9),... |

974 | Numerical Methods for Unconstrained Optimization and Nonlinear Equations - Dennis, Schnabel - 1983 |

972 | Finite Mixture Models - McLachlan, Peel - 2000 |

817 |
Comprehensive identification of cell cycle-regulated genes of the yeast saccharomyces cerevisiae by microarray hybridization
- SPELLMAN, SHERLOCK, et al.
- 1998
(Show Context)
Citation Context ...ns. A typical example is temporal gene expression studies, in which a series of micorarray experiments are conducted sequentially during a biological process, e.g., cell cycle microarray experiments (=-=Spellman et al. 1998-=-). At each time point, mRNA expression levels of thousands of genes are measured simultaneously. Collected over time, a gene’s “temporal expression profile” gives the scientist some clues on what role... |

550 | Comparing partitions - Hubert, Arabie - 1985 |

505 | Functional Data Analysis - Ramsay, Silverman - 1997 |

455 | Smoothing noisy data with spline functions: estimating the correct degree of smoothing by the method of generalized cross-validation - Craven, Wahba - 1979 |

283 | Model-based clustering, discriminant analysis, and density estimation
- Fraley, Raftery
- 2002
(Show Context)
Citation Context ...ervations in the measurements yield an unbalanced design, which requires imputation beforehand for application of mul2stivariate approaches, e.g., the multivariate Gaussian clustering method (MCLUST, =-=Fraley and Raftery 1990-=-). In addition to the time factor, such repeated measurements often contain other covariates, e.g., replicates at each time point, species in comparative genomics studies (McCarroll et al. 2004), and ... |

169 |
A Monte carlo implementation of the EM algorithm and the poor man’s data augmentation algorithms
- Wei, Tanner
- 1990
(Show Context)
Citation Context ...e wik right after the E-step (3.3). Note that when c = 0, the proposed algorithm is exactly the original EM algorithm, whereas the proposed algorithm reduces to a variant of Monte Carlo EM algorithm (=-=Wei and Tanner 1990-=-) when c = 1. In this way, it is possible to make accurate approximations during the E-step while greatly reducing the computation of the M-step. Finally, in order to avoid local optima, the rejection... |

152 |
Spline Models for Observational Data, Volume 59
- Wahba
- 1990
(Show Context)
Citation Context ...dicity. Functional ANOVA is employed to further decompose multivariate functions (fixed effects) in a reproducing kernel Hilbert space and provide associated notions of main effect 3sand interaction (=-=Wahba 1990-=- and Gu 2002). Parsimonious random effects, complementing the fixed effects, are used to capture various correlation structures. The mixed-effect models are nested under a general mixture model, in wh... |

121 |
Smoothing Spline ANOVA Models
- Gu
- 2002
(Show Context)
Citation Context ...nal ANOVA is employed to further decompose multivariate functions (fixed effects) in a reproducing kernel Hilbert space and provide associated notions of main effect 3sand interaction (Wahba 1990 and =-=Gu 2002-=-). Parsimonious random effects, complementing the fixed effects, are used to capture various correlation structures. The mixed-effect models are nested under a general mixture model, in which the hete... |

94 | Bayesian confidence intervals for the cross-validated smoothing spline - WAHBA - 1983 |

86 |
That BLUP is a good thing: the estimation of random effects
- Robinson
- 1991
(Show Context)
Citation Context ...ss-of-fit and the smoothness of µ. (2.3) is also referred to as penalized Henderson’s likelihood since the first two terms are proportional to the joint density (Henderson’s likelihood) of (y i, bi) (=-=Robinson 1991-=-). To minimize (2.3), we only need to consider smooth functions in the space {µ : M(µ) < ∞} or subspace therein. As a abstract generalization of the vector spaces used extensively in multivariate anal... |

82 | Applied functional data analysis: methods and case studies - JO, BW |

79 | Gene expression during the life cycle of Drosophila melanogaster - Arbeitman, Furlong, et al. - 2002 |

77 | Clustering of time-course gene expression data using a mixed-effects model with B-spline - Luan, Li - 2003 |

68 |
On use of the EM algorithm for penalized likelihood estimation
- Green
- 1990
(Show Context)
Citation Context ... T (y i−µk(xi)−Zibi)+σ 2 b T i B −1 k bi]+ where λk is the smoothing parameter for µk. K� NλkM(µk) Once the penalized Henderson’s likelihood (3.2) is obtained, the EM algorithm (Dempster et al. 1977, =-=Green 1990-=-) can be derived as follows. The E-step simply requires the calculation of wik = pkϕ(yi; µk(xi), Σk) �K l=1 plϕ(yi; µl(xi), Σl) where Σk = ZiBkZ T i + σ 2 I, and ϕ is the Gaussian density function. − ... |

59 |
FunSpec: a web-based cluster interpreter for yeast. BMC Bioinformatics 3:35
- Robinson, Grigull, et al.
- 2002
(Show Context)
Citation Context ...ic condition, bi ∼ N(0, σ2 b ) is the gene specific random effect. We fit the model using the penalty (2.5) with a = 2. In total, 2388 genes were clustered into 28 clusters using our method. FunSpec (=-=Robinson et al. 2002-=-) was used for gene annotation and biological function enrichment analysis. We found 26 clusters out of 28 clusters discovered have overrepresented biological functions. The estimated mean gene expres... |

48 | Clustering for sparsely sampled functional data
- James, Sugar
- 2003
(Show Context)
Citation Context ...timation with roughness penalty M(µ) = ∫ 1 0 (d2 µ1/dt 2 ) 2 dt. We compared our method with MCLUST (Fraley and Raftery 1990), FCM classification likelihood (FCMc), and FCM mixture likelihood (FCMm) (=-=James and Sugar 2003-=-). 19−4 −2 0 2 4 −4 −2 0 2 4 −4 −2 0 2 4 −4 −2 0 2 4 Figure 4.1: The estimated mean curves (dash lines) and 95% Bayesian confidence intervals for one simulated dataset. The true functions are superim... |

48 | Bayesian ‘Confidence’ Intervals for Smoothing Splines - Nychka - 1988 |

41 | A quantative study of gene regulation involved in the immune response of anopheline mosquitoes: An application of Bayesian hierarchical clustering of curves - Heard, Holmes, et al. - 2006 |

41 | Genome-wide analysis of developmental and sex-regulated gene expression profiles in Caenorhabditis elegans - Jiang, Ryu, et al. - 2001 |

34 |
Genemerge: post-genomic analysis, data-mining and hypothesis testing
- Castillo-Davis, Hartl
- 2003
(Show Context)
Citation Context ...al functions of genes in each cluster were annotated using Gene Ontology, and Bonferroni corrected P-values of biological function enrichment were calculated based on the hypergeometric distribution (=-=Castillo-Davis and Hartl 2003-=-). Of the 34 clusters discovered, 21 clusters exhibit significant biological functions over-representation (P-value < 0.05). The estimated mean gene expression curves of three clusters and their 95% B... |

34 |
Comparing genomic expression patterns across species identifies shared transcriptional profile in aging." Nat Genet 2004 36(2
- Mccarroll, Murphy, et al.
(Show Context)
Citation Context ...UST, Fraley and Raftery 1990). In addition to the time factor, such repeated measurements often contain other covariates, e.g., replicates at each time point, species in comparative genomics studies (=-=McCarroll et al. 2004-=-), and treatment groups in case-control studies (Storey et al. 2005), as well as many factors in a factorial designed experiment. Incorporation of multiple covariates adds another layer of complexity.... |

31 |
Rejection control for sequential importance sampling
- Liu, Chen, et al.
- 1998
(Show Context)
Citation Context ... Henderson’s likelihood approach for model-fitting and design a rejection-controlled EM algorithm for estimation. In this EM algorithm, the E-step is followed by a rejection-controlled sampling step (=-=Liu et al. 1998-=-) to eliminate a significant number of functional observations, whose posterior probabilities of belonging to a particular cluster is negligible, from calculation in the subsequent M-step. The M-step ... |

27 |
A data-driven clustering method for time course gene expression
- Ma, Castillo-Davis, et al.
- 2006
(Show Context)
Citation Context ... the M-step (3.9) is expensive, unstable and sometimes even infeasible. To alleviate the computation and stabilize the algorithm, we propose to add a rejection control step (Liu, Chen, and Wong 1998, =-=Ma et al. 2006-=-) in the EM algorithm and refer to the modified algorithm as rejection controlled EM algorithm. Firstly, we set up a threshold value c (e.g., c = 0.05). Given this threshold value, we introduce the fo... |

26 | Smoothing spline Gaussian regression: More scalable computation via efficient approximation - Kim, Gu - 2004 |

16 | Mixed-effects smoothing spline ANOVA - Wang - 1998 |

9 | Model-based methods for identifying periodically regulated genes based on the time course microarray gene expression data - Luan, Li - 2004 |

9 |
The nonparametric estimation of branching curves
- Silverman, Wood
- 1987
(Show Context)
Citation Context ...s the gene specific random effect. The penalized Henderson’s likelihood was employed with roughness penalty M of the form (2.5). Sex differentiation of the fruitfly was modeled by a branching spline (=-=Silverman and Wood 1987-=-), the general analytic form of which with two branches on the right is ⎧ �m ⎪⎨ ν=1 µ(t) = ⎪⎩ dνφν(t) + �k i=1 ciRM(si, t) if t ≤ sk �m ν=1 dνφν(t) + �k i=1 ciRM(si, t) + �T i=k+1 c1iRM(si − sk, t − s... |

5 |
Significance of time course microarray experiments
- Storey, Xiao, et al.
- 2005
(Show Context)
Citation Context ...peated measurements often contain other covariates, e.g., replicates at each time point, species in comparative genomics studies (McCarroll et al. 2004), and treatment groups in case-control studies (=-=Storey et al. 2005-=-), as well as many factors in a factorial designed experiment. Incorporation of multiple covariates adds another layer of complexity. Clustering methods taking all these factors into account are still... |

3 | Imaging of structure at and near the core mantle boundary using a generalized Radon transform: 2. Statistical inference of singularities - Ma, Wang, et al. |

2 | Optimal smoothing in nonparametric mixed effect models - Gu, Ma - 2005 |

2 | Smoothing Spline ANOVA Models - unknown authors - 2002 |

1 | Kwast (2006). Metabolic-statedependent remodeling of the transcriptome in response to anoxia and subsequent reoxygenation in saccharomyces cerevisiae - Lai, Kosorukoff, et al. |

1 | der Hilst (2007). Imaging of structure at and near the core mantle boundary using a generalized Radon transform: 2. statistical inference of singularities - Ma, Wang, et al. |