Computes log-likelihood for clusters based on distance-based cohesion and repulsion. More...

#include <Natarajan_likelihood.hpp>

Inheritance diagram for Natarajan_likelihood:

Public Member Functions
	Natarajan_likelihood (const Data &data, const Params &param)
	Constructs a Likelihood object with precomputation.
double	cluster_loglikelihood (int cluster_index) const override final
	Computes the full log-likelihood for a cluster.
double	cluster_loglikelihood (int cluster_index, const Eigen::Ref< const Eigen::VectorXi > &cls_ass_k) const override final __attribute__((hot))
	Computes the full log-likelihood for a cluster with given assignments.
double	point_loglikelihood_cond (int point_index, int cluster_index) const override final __attribute__((hot))
	Computes the conditional log-likelihood of a point given a cluster.
Public Member Functions inherited from Likelihood
	Likelihood (const Data &data, const Params &param)
virtual	~Likelihood ()=default

Private Member Functions
double	compute_cohesion (int point_index, int cluster_index, const Eigen::Ref< const Eigen::VectorXi > &cls_ass_k, int n_k) const
	Computes the cohesion component of the log-likelihood.
double	compute_repulsion (int point_index, int cluster_index, const Eigen::Ref< const Eigen::VectorXi > &cls_ass_k, int n_k) const
	Computes the repulsion component of the log-likelihood.

Private Attributes
const double	lgamma_delta1
	Precomputed lgamma(delta1) for cohesion.
const double	log_beta_alpha
	Precomputed log(beta) * alpha - lgamma(alpha).
const double	lgamma_delta2
	Precomputed lgamma(delta2) for repulsion.
const double	log_gamma_zeta
	Precomputed log(gamma) * zeta - lgamma(zeta).
std::vector< double >	lgamma_zeta_mt_cache
	Cache for lgamma(zeta_mt) values.
std::vector< double >	lgamma_alpha_mh_cache
	Cache for lgamma(alpha_mh) values.
std::vector< double >	log_D_data
	Precomputed log distance matrix (flattened).
const int	D_cols
	Number of columns in distance matrix.

Additional Inherited Members
Protected Attributes inherited from Likelihood
const Data &	data
	Reference to Data object with distances and allocations.
const Params &	params
	Reference to model parameters.

Detailed Description

Computes log-likelihood for clusters based on distance-based cohesion and repulsion.

This class handles the computation of log-likelihoods for clustering models that incorporate both cohesion (within-cluster similarity) and repulsion (between-cluster dissimilarity) components. It precomputes values for efficiency and provides methods to evaluate both cluster-level and point-level conditional log-likelihoods. reference: Natarajan et al. (2023) "Cohesion and Repulsion in Bayesian Distance Clustering"

Constructor & Destructor Documentation

◆ Natarajan_likelihood()

Natarajan_likelihood::Natarajan_likelihood	(	const Data &	data,
		const Params &	param )

inline

Constructs a Likelihood object with precomputation.

Parameters

data	Reference to Data object with distances and allocations
param	Reference to model parameters

The constructor precomputes several values for computational efficiency:

Log-gamma values for delta parameters
Logarithmic combinations of hyperparameters
Logarithm of the entire distance matrix

Member Function Documentation

◆ cluster_loglikelihood() [1/2]

double Natarajan_likelihood::cluster_loglikelihood ( int cluster_index ) const

finaloverridevirtual

Computes the full log-likelihood for a cluster.

Parameters

cluster_index Index of the cluster to evaluate

Returns: Total log-likelihood (cohesion + repulsion)

This method computes both the within-cluster cohesion and the between-cluster repulsion contributions for the specified cluster.

Implements Likelihood.

◆ cluster_loglikelihood() [2/2]

double Natarajan_likelihood::cluster_loglikelihood	(	int	cluster_index,
		const Eigen::Ref< const Eigen::VectorXi > &	cls_ass_k ) const

finaloverridevirtual

Computes the full log-likelihood for a cluster with given assignments.

Parameters

cluster_index	Index of the cluster to evaluate
cls_ass_k	Vector of point indices in the cluster

Returns: Total log-likelihood (cohesion + repulsion)

This overload allows computing the likelihood with a custom set of cluster assignments without modifying the data structure.

Implements Likelihood.

◆ compute_cohesion()

double Natarajan_likelihood::compute_cohesion	(	int	point_index,
		int	cluster_index,
		const Eigen::Ref< const Eigen::VectorXi > &	cls_ass_k,
		int	n_k ) const

private

Computes the cohesion component of the log-likelihood.

Parameters

point_index	Index of the point being evaluated
cluster_index	Index of the cluster
cls_ass_k	Vector of point indices in the cluster
n_k	Number of points in the cluster

Returns: Cohesion log-likelihood contribution

◆ compute_repulsion()

double Natarajan_likelihood::compute_repulsion	(	int	point_index,
		int	cluster_index,
		const Eigen::Ref< const Eigen::VectorXi > &	cls_ass_k,
		int	n_k ) const

private

Computes the repulsion component of the log-likelihood.

Parameters

point_index	Index of the point being evaluated
cluster_index	Index of the cluster
cls_ass_k	Vector of point indices in the cluster
n_k	Number of points in the cluster

Returns: Repulsion log-likelihood contribution

◆ point_loglikelihood_cond()

double Natarajan_likelihood::point_loglikelihood_cond	(	int	point_index,
		int	cluster_index ) const

finaloverridevirtual

Computes the conditional log-likelihood of a point given a cluster.

Parameters

point_index	Index of the point to evaluate
cluster_index	Index of the target cluster

Returns: Conditional log-likelihood of assigning the point to the cluster

This method evaluates how well a point fits into a specific cluster, considering both its cohesion with points in that cluster and its repulsion from points in other clusters.

Implements Likelihood.

Member Data Documentation

◆ D_cols

const int Natarajan_likelihood::D_cols

private

Number of columns in distance matrix.

◆ lgamma_alpha_mh_cache

std::vector<double> Natarajan_likelihood::lgamma_alpha_mh_cache

private

Cache for lgamma(alpha_mh) values.

◆ lgamma_delta1

const double Natarajan_likelihood::lgamma_delta1

private

Precomputed lgamma(delta1) for cohesion.

◆ lgamma_delta2

const double Natarajan_likelihood::lgamma_delta2

private

Precomputed lgamma(delta2) for repulsion.

◆ lgamma_zeta_mt_cache

std::vector<double> Natarajan_likelihood::lgamma_zeta_mt_cache

private

Cache for lgamma(zeta_mt) values.

◆ log_beta_alpha

const double Natarajan_likelihood::log_beta_alpha

private

Precomputed log(beta) * alpha - lgamma(alpha).

◆ log_D_data

std::vector<double> Natarajan_likelihood::log_D_data

private

Precomputed log distance matrix (flattened).

◆ log_gamma_zeta

const double Natarajan_likelihood::log_gamma_zeta

private

Precomputed log(gamma) * zeta - lgamma(zeta).

The documentation for this class was generated from the following files:

src/likelihoods/Natarajan_likelihood.hpp
src/likelihoods/Natarajan_likelihood.cpp

Public Member Functions

Private Member Functions

Private Attributes

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

◆ Natarajan_likelihood()

Member Function Documentation

◆ cluster_loglikelihood() [1/2]

◆ cluster_loglikelihood() [2/2]

◆ compute_cohesion()

◆ compute_repulsion()

◆ point_loglikelihood_cond()

Member Data Documentation

◆ D_cols

◆ lgamma_alpha_mh_cache

◆ lgamma_delta1

◆ lgamma_delta2

◆ lgamma_zeta_mt_cache

◆ log_beta_alpha

◆ log_D_data

◆ log_gamma_zeta