Merge branch 'turnbull-emicm'

Change to EM-ICM algorithm to fit Turnbull estimator

Much more efficient - 19.7x speedup compared with old algorithm!

src/turnbull.rs

@@ -27,6 +27,7 @@ use prettytable::{Table, format, row};
 use rayon::prelude::*;
 use serde::{Serialize, Deserialize};
 
+use crate::pava::monotonic_regression_pava;
 use crate::term::UnconditionalTermLike;
 
 #[derive(Args)]
@@ -205,7 +206,7 @@ pub fn fit_turnbull(data_times: MatrixXx2<f64>, progress_bar: ProgressBar, max_i
 	progress_bar.set_style(ProgressStyle::with_template("[{elapsed_precise}] {bar:40} {msg}").unwrap());
 	progress_bar.set_length(u64::MAX);
 	progress_bar.reset();
-	progress_bar.println("Running iterative algorithm to fit Turnbull estimator");
+	progress_bar.println("Running EM-ICM algorithm to fit Turnbull estimator");
 	
 	let (s, ll) = fit_turnbull_estimator(&mut data, progress_bar.clone(), max_iterations, ll_tolerance, s);
 	
@@ -285,24 +286,38 @@ fn get_turnbull_intervals(data_times: &MatrixXx2<f64>) -> Vec<(f64, f64)> {
 	return intervals;
 }
 
-fn fit_turnbull_estimator(data: &mut TurnbullData, progress_bar: ProgressBar, max_iterations: u32, ll_tolerance: f64, mut s: Vec<f64>) -> (Vec<f64>, f64) {
+fn fit_turnbull_estimator(data: &mut TurnbullData, progress_bar: ProgressBar, max_iterations: u32, ll_tolerance: f64, mut p: Vec<f64>) -> (Vec<f64>, f64) {
-	// Get likelihood for each observation (denominator of μ_ij)
+	// Pre-compute S, the survival probability at the start of each interval
-	let mut likelihood_obs = get_likelihood_obs(data, &s);
+	let mut s = p_to_s(&p);
+	
+	// Get likelihood for each observation
+	let likelihood_obs = get_likelihood_obs(data, &s);
 	let mut ll_model: f64 = likelihood_obs.iter().map(|l| l.ln()).sum();
 	
 	let mut iteration = 1;
 	loop {
-		// Compute π_j to update s
+		// -------
-		let pi = compute_pi(data, &s, likelihood_obs);
+		// EM step
 		
-		let likelihood_obs_new = get_likelihood_obs(data, &pi);
+		let p_after_em = do_em_step(data, &p, &s);
-		let ll_model_new = likelihood_obs_new.iter().map(|l| l.ln()).sum();
+		let s_after_em = p_to_s(&p_after_em);
+		
+		let likelihood_obs_after_em = get_likelihood_obs(data, &s_after_em);
+		let ll_model_after_em: f64 = likelihood_obs_after_em.iter().map(|l| l.ln()).sum();
+		
+		p = p_after_em;
+		s = s_after_em;
+		
+		// --------
+		// ICM step
+		
+		let (p_new, s_new, ll_model_new) = do_icm_step(data, &p, &s, ll_model_after_em);
 		
 		let ll_change = ll_model_new - ll_model;
 		let converged = ll_change <= ll_tolerance;
 		
-		s = pi;
+		p = p_new;
-		likelihood_obs = likelihood_obs_new;
+		s = s_new;
 		ll_model = ll_model_new;
 		
 		// Estimate progress bar according to either the order of magnitude of the ll_change relative to tolerance, or iteration/max_iterations
@@ -324,48 +339,150 @@ fn fit_turnbull_estimator(data: &mut TurnbullData, progress_bar: ProgressBar, ma
 		}
 	}
 	
-	return (s, ll_model);
+	return (p, ll_model);
+}
+
+fn p_to_s(p: &Vec<f64>) -> Vec<f64> {
+	let mut s = Vec::with_capacity(p.len() + 1);  // Survival probabilities
+	let mut survival = 1.0;
+	s.push(1.0);
+	for p_j in p.iter() {
+		survival -= p_j;
+		s.push(survival);
+	}
+	return s;
+}
+
+fn s_to_lambda(s: &Vec<f64>) -> Vec<f64> {
+	// S = 1 means Λ = -inf and S = 0 means Λ = inf so skip these
+	let mut lambda = Vec::with_capacity(s.len() - 2);  // Cumulative hazard
+	for s_j in &s[1..(s.len() - 1)] {
+		lambda.push((-s_j.ln()).ln());
+	}
+	return lambda;
 }
 
 fn get_likelihood_obs(data: &TurnbullData, s: &Vec<f64>) -> Vec<f64> {
 	return data.data_time_interval_indexes
 		.par_iter()
-		.map(|(idx_left, idx_right)| s[*idx_left..(*idx_right + 1)].iter().sum())
+		.map(|(idx_left, idx_right)| s[*idx_left] - s[*idx_right + 1])
-		.collect();
+		.collect();  // TODO: Return iterator directly
 }
 
-fn compute_pi(data: &TurnbullData, s: &Vec<f64>, likelihood_obs: Vec<f64>) -> Vec<f64> {
+fn do_em_step(data: &TurnbullData, p: &Vec<f64>, s: &Vec<f64>) -> Vec<f64> {
-	/*
+	// Compute contributions to m
-	let mut pi: Vec<f64> = vec![0.0; data.num_intervals()];
+	let mut m_contrib = vec![0.0; data.num_intervals()];
-	for ((idx_left, idx_right), likelihood_obs_i) in data.data_time_interval_indexes.iter().zip(likelihood_obs.iter()) {
+	for (idx_left, idx_right) in data.data_time_interval_indexes.iter() {
-		for j in *idx_left..(*idx_right + 1) {
+		let contrib = 1.0 / (s[*idx_left] - s[*idx_right + 1]);
-			pi[j] += s[j] / likelihood_obs_i / data.num_obs() as f64;
-		}
-	}
-	*/
 		
-	let pi = data.data_time_interval_indexes.par_iter().zip(likelihood_obs.par_iter())
+		// Adds to m for the first interval in the observation
-		.fold_with(
+		m_contrib[*idx_left] += contrib;
-			// Compute the contributions to pi[j] for each observation and sum them in parallel using fold_with
-			vec![0.0; data.num_intervals()],
-			|mut acc, ((idx_left, idx_right), likelihood_obs_i)| {
-				// Contributions to pi[j] for the i-th observation
-				for j in *idx_left..(*idx_right + 1) {
-					acc[j] += s[j] / likelihood_obs_i / data.num_obs() as f64;
-				}
-				acc
-			}
-		)
-		.reduce(
-			// Reduce all the sub-sums from fold_with into the total sum
-			|| vec![0.0; data.num_intervals()],
-			|mut acc, subsum| {
-				acc.iter_mut().zip(subsum.iter()).for_each(|(x, y)| *x += y);
-				acc
-			}
-		);
 		
-	return pi;
+		// Subtracts from m for the first interval beyond the observation
+		if *idx_right + 1 < data.num_intervals() {
+			m_contrib[*idx_right + 1] -= contrib;
+		}
+	}
+	
+	// Compute m
+	let mut m = Vec::with_capacity(data.num_intervals());
+	let mut m_last = 0.0;
+	for m_contrib_j in m_contrib {
+		let m_next = m_last + m_contrib_j / (data.num_obs() as f64);
+		m.push(m_next);
+		m_last = m_next;
+	}
+	
+	// Update p
+	// p := p * m
+	let p_new = p.par_iter().zip(m.into_par_iter()).map(|(p_j, m_j)| p_j * m_j).collect();
+	
+	return p_new;
+}
+
+fn do_icm_step(data: &TurnbullData, _p: &Vec<f64>, s: &Vec<f64>, ll_model: f64) -> (Vec<f64>, Vec<f64>, f64) {
+	// Compute Λ, the cumulative hazard
+	// Since Λ = -inf when survival is 1, and Λ = inf when survival is 0, these are omitted
+	// The entry at lambda[j] corresponds to the survival immediately before time point j + 1
+	let lambda = s_to_lambda(&s);
+	
+	// Compute gradient and diagonal of Hessian
+	let mut gradient = vec![0.0; data.num_intervals() - 1];
+	let mut hessdiag = vec![0.0; data.num_intervals() - 1];
+	for (idx_left, idx_right) in data.data_time_interval_indexes.iter() {
+		let denom = s[*idx_left] - s[*idx_right + 1];
+		
+		// Add to gradient[j] when j + 1 == idx_right + 1
+		// Add to hessdiag[j] when j + 1 == idx_right + 1
+		if *idx_right < gradient.len() {
+			let j = *idx_right;
+			gradient[j] += (-lambda[j].exp() + lambda[j]).exp() / denom;
+			
+			let a = ((lambda[j] - lambda[j].exp()).exp() * (1.0 - lambda[j].exp())) / denom;
+			let b = (2.0 * lambda[j] - 2.0 * lambda[j].exp()).exp() / denom.powi(2);
+			hessdiag[j] += a - b;
+		}
+		
+		// Subtract from gradient[j] when j + 1 == idx_left
+		// Add to hessdiag[j] when j + 1 == idx_left
+		if *idx_left > 0 {
+			let j = *idx_left - 1;
+			gradient[j] -= (-lambda[j].exp() + lambda[j]).exp() / denom;
+			
+			let a = ((lambda[j] - lambda[j].exp()).exp() * (1.0 - lambda[j].exp())) / denom;
+			let b = (2.0 * lambda[j] - 2.0 * lambda[j].exp()).exp() / denom.powi(2);
+			hessdiag[j] += -a - b;
+		}
+	}
+	
+	// Description in Anderson-Bergman (2017) is slightly misleading
+	// Since we are maximising the likelihood, the second derivatives will be negative
+	// And we will move in the direction of the gradient
+	// So there are a few more negative signs here than suggested
+	
+	let weights = -DVector::from_vec(hessdiag.clone()) / 2.0;
+	let gradient_over_hessdiag = DVector::from_vec(gradient.par_iter().zip(hessdiag.par_iter()).map(|(g, h)| g / h).collect());
+	
+	let mut s_new;
+	let mut p_new;
+	let mut ll_model_new: f64;
+	
+	// Take as large a step as possible while the log-likelihood increases
+	let mut step_size_exponent: i32 = 0;
+	loop {
+		let step_size = 0.5_f64.powi(step_size_exponent);
+		let lambda_target = -gradient_over_hessdiag.clone() * step_size + DVector::from_vec(lambda.clone());
+		
+		let lambda_new = monotonic_regression_pava(lambda_target, weights.clone());
+		
+		// Convert Λ to S to p
+		s_new = Vec::with_capacity(data.num_intervals() + 1);
+		p_new = Vec::with_capacity(data.num_intervals());
+		
+		let mut survival = 1.0;
+		s_new.push(1.0);
+		for lambda_j in lambda_new.iter() {
+			let next_survival = (-lambda_j.exp()).exp();
+			s_new.push(next_survival);
+			p_new.push(survival - next_survival);
+			survival = next_survival;
+		}
+		s_new.push(0.0);
+		p_new.push(survival);
+		
+		let likelihood_obs_new = get_likelihood_obs(data, &s_new);
+		ll_model_new = likelihood_obs_new.iter().map(|l| l.ln()).sum();
+		
+		if ll_model_new > ll_model {
+			return (p_new, s_new, ll_model_new);
+		}
+		
+		step_size_exponent += 1;
+		
+		if step_size_exponent > 10 {
+			panic!("ICM fails to increase log-likelihood");
+		}
+	}
 }
 
 fn compute_hessian(data: &TurnbullData, s: &Vec<f64>) -> DMatrix<f64> {