libsvm_rs/

predict.rs

//! Prediction functions matching the original LIBSVM.
//!
//! Provides `predict`, `predict_values`, and `predict_probability` for
//! all SVM types:
//! - Classification (C-SVC, ν-SVC): one-vs-one voting + optional probability
//! - One-class SVM: sign of decision value + optional density probability
//! - Regression (ε-SVR, ν-SVR): continuous output

use crate::kernel::k_function;
use crate::probability::{multiclass_probability, predict_one_class_probability, sigmoid_predict};
use crate::types::{SvmModel, SvmNode, SvmType};

/// Compute decision values and return the predicted label/value.
///
/// For classification, `dec_values` receives `nr_class * (nr_class - 1) / 2`
/// pairwise decision values. For regression/one-class, a single value.
///
/// Returns the predicted label (classification) or function value (regression).
///
/// Matches LIBSVM's `svm_predict_values`.
pub fn predict_values(model: &SvmModel, x: &[SvmNode], dec_values: &mut [f64]) -> f64 {
    match model.param.svm_type {
        SvmType::OneClass | SvmType::EpsilonSvr | SvmType::NuSvr => {
            let sv_coef = &model.sv_coef[0];
            let mut sum = 0.0;
            for (i, sv) in model.sv.iter().enumerate() {
                sum += sv_coef[i] * k_function(x, sv, &model.param);
            }
            sum -= model.rho[0];
            dec_values[0] = sum;

            if model.param.svm_type == SvmType::OneClass {
                if sum > 0.0 {
                    1.0
                } else {
                    -1.0
                }
            } else {
                sum
            }
        }
        SvmType::CSvc | SvmType::NuSvc => {
            let nr_class = model.nr_class;
            let l = model.sv.len();

            // Compute kernel values for all SVs
            let kvalue: Vec<f64> = model
                .sv
                .iter()
                .map(|sv| k_function(x, sv, &model.param))
                .collect();

            // Compute start indices for each class's SVs
            let mut start = vec![0usize; nr_class];
            for i in 1..nr_class {
                start[i] = start[i - 1] + model.n_sv[i - 1];
            }

            // One-vs-one voting
            let mut vote = vec![0usize; nr_class];
            let mut p = 0;
            for i in 0..nr_class {
                for j in (i + 1)..nr_class {
                    let mut sum = 0.0;
                    let si = start[i];
                    let sj = start[j];
                    let ci = model.n_sv[i];
                    let cj = model.n_sv[j];

                    let coef1 = &model.sv_coef[j - 1];
                    let coef2 = &model.sv_coef[i];

                    for k in 0..ci {
                        sum += coef1[si + k] * kvalue[si + k];
                    }
                    for k in 0..cj {
                        sum += coef2[sj + k] * kvalue[sj + k];
                    }
                    sum -= model.rho[p];
                    dec_values[p] = sum;

                    if sum > 0.0 {
                        vote[i] += 1;
                    } else {
                        vote[j] += 1;
                    }
                    p += 1;
                }
            }

            // Find class with most votes
            let vote_max_idx = vote
                .iter()
                .enumerate()
                .max_by_key(|&(_, &v)| v)
                .map(|(i, _)| i)
                .unwrap_or(0);

            let _ = l; // suppress unused warning
            model.label[vote_max_idx] as f64
        }
    }
}

/// Predict the label/value for a single instance.
///
/// Convenience wrapper around `predict_values` that allocates the
/// decision values buffer internally. Matches LIBSVM's `svm_predict`.
///
/// ```
/// use libsvm_rs::predict::predict;
/// use libsvm_rs::train::svm_train;
/// use libsvm_rs::{KernelType, SvmNode, SvmParameterBuilder, SvmProblem, SvmType};
///
/// let problem = SvmProblem {
///     labels: vec![-1.0, -1.0, 1.0, 1.0],
///     instances: vec![
///         vec![SvmNode { index: 1, value: -2.0 }],
///         vec![SvmNode { index: 1, value: -1.0 }],
///         vec![SvmNode { index: 1, value: 1.0 }],
///         vec![SvmNode { index: 1, value: 2.0 }],
///     ],
/// };
/// let param = SvmParameterBuilder::new()
///     .svm_type(SvmType::CSvc)
///     .kernel_type(KernelType::Linear)
///     .build()?;
/// let model = svm_train(&problem, &param);
///
/// assert_eq!(predict(&model, &[SvmNode { index: 1, value: 1.5 }]), 1.0);
/// # Ok::<(), libsvm_rs::SvmError>(())
/// ```
pub fn predict(model: &SvmModel, x: &[SvmNode]) -> f64 {
    let n = match model.param.svm_type {
        SvmType::OneClass | SvmType::EpsilonSvr | SvmType::NuSvr => 1,
        SvmType::CSvc | SvmType::NuSvc => model.nr_class * (model.nr_class - 1) / 2,
    };
    let mut dec_values = vec![0.0; n];
    predict_values(model, x, &mut dec_values)
}

/// Predict with probability estimates.
///
/// Returns `Some((label, probs))` where `probs[i]` is the estimated
/// probability of class `model.label[i]`. Returns `None` when the
/// model was not trained with probability support.
///
/// - **C-SVC / ν-SVC**: requires `model.prob_a` and `model.prob_b`.
///   Uses Platt scaling on pairwise decision values, then
///   `multiclass_probability` for k > 2.
/// - **One-class**: requires `model.prob_density_marks`.
///   Returns `[p, 1-p]` via density-mark lookup.
/// - **SVR**: probability prediction is not supported (returns `None`).
///
/// Matches LIBSVM's `svm_predict_probability`.
#[allow(clippy::needless_range_loop)]
pub fn predict_probability(model: &SvmModel, x: &[SvmNode]) -> Option<(f64, Vec<f64>)> {
    match model.param.svm_type {
        SvmType::CSvc | SvmType::NuSvc if !model.prob_a.is_empty() && !model.prob_b.is_empty() => {
            let nr_class = model.nr_class;
            let n_pairs = nr_class * (nr_class - 1) / 2;
            let mut dec_values = vec![0.0; n_pairs];
            predict_values(model, x, &mut dec_values);

            let min_prob = 1e-7;

            // Build pairwise probability matrix
            let mut pairwise = vec![vec![0.0; nr_class]; nr_class];
            let mut k = 0;
            for i in 0..nr_class {
                for j in (i + 1)..nr_class {
                    let p = sigmoid_predict(dec_values[k], model.prob_a[k], model.prob_b[k])
                        .max(min_prob)
                        .min(1.0 - min_prob);
                    pairwise[i][j] = p;
                    pairwise[j][i] = 1.0 - p;
                    k += 1;
                }
            }

            let mut prob_estimates = vec![0.0; nr_class];
            if nr_class == 2 {
                prob_estimates[0] = pairwise[0][1];
                prob_estimates[1] = pairwise[1][0];
            } else {
                multiclass_probability(nr_class, &pairwise, &mut prob_estimates);
            }

            // Find class with highest probability
            let mut best = 0;
            for i in 1..nr_class {
                if prob_estimates[i] > prob_estimates[best] {
                    best = i;
                }
            }

            Some((model.label[best] as f64, prob_estimates))
        }

        SvmType::OneClass if !model.prob_density_marks.is_empty() => {
            let mut dec_value = [0.0];
            let pred_result = predict_values(model, x, &mut dec_value);
            let p = predict_one_class_probability(&model.prob_density_marks, dec_value[0]);
            Some((pred_result, vec![p, 1.0 - p]))
        }

        _ => None,
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::io::load_model;
    use crate::io::load_problem;
    use std::path::PathBuf;

    fn data_dir() -> PathBuf {
        PathBuf::from(env!("CARGO_MANIFEST_DIR"))
            .join("..")
            .join("..")
            .join("data")
    }

    #[test]
    fn predict_heart_scale() {
        // Load model trained by C LIBSVM and predict on training data
        let model = load_model(&data_dir().join("heart_scale.model")).unwrap();
        let problem = load_problem(&data_dir().join("heart_scale")).unwrap();

        let mut correct = 0;
        for (i, instance) in problem.instances.iter().enumerate() {
            let pred = predict(&model, instance);
            if pred == problem.labels[i] {
                correct += 1;
            }
        }

        let accuracy = correct as f64 / problem.labels.len() as f64;
        // C LIBSVM gets ~86.67% accuracy on training set with default params
        assert!(
            accuracy > 0.85,
            "accuracy {:.2}% too low (expected >85%)",
            accuracy * 100.0
        );
    }

    #[test]
    fn predict_values_binary() {
        let model = load_model(&data_dir().join("heart_scale.model")).unwrap();
        let problem = load_problem(&data_dir().join("heart_scale")).unwrap();

        // For binary classification, there's exactly 1 decision value
        let mut dec_values = vec![0.0; 1];
        let label = predict_values(&model, &problem.instances[0], &mut dec_values);

        // Decision value should be non-zero
        assert!(dec_values[0].abs() > 1e-10);
        // Label should match what predict returns
        assert_eq!(label, predict(&model, &problem.instances[0]));
    }

    #[test]
    fn predict_probability_binary() {
        let problem = load_problem(&data_dir().join("heart_scale")).unwrap();
        let param = crate::types::SvmParameter {
            svm_type: SvmType::CSvc,
            kernel_type: crate::types::KernelType::Rbf,
            gamma: 1.0 / 13.0,
            c: 1.0,
            cache_size: 100.0,
            eps: 0.001,
            shrinking: true,
            probability: true,
            ..Default::default()
        };

        let model = crate::train::svm_train(&problem, &param);
        assert!(!model.prob_a.is_empty());

        for instance in &problem.instances {
            let result = predict_probability(&model, instance);
            assert!(result.is_some(), "should return probability");
            let (label, probs) = result.unwrap();
            assert!(label == 1.0 || label == -1.0);
            assert_eq!(probs.len(), 2);
            let sum: f64 = probs.iter().sum();
            assert!(
                (sum - 1.0).abs() < 1e-6,
                "probs sum to {}, expected 1.0",
                sum
            );
            for &p in &probs {
                assert!((0.0..=1.0).contains(&p), "prob {} out of [0,1]", p);
            }
        }
    }

    #[test]
    fn predict_probability_multiclass() {
        let problem = load_problem(&data_dir().join("iris.scale")).unwrap();
        let param = crate::types::SvmParameter {
            svm_type: SvmType::CSvc,
            kernel_type: crate::types::KernelType::Rbf,
            gamma: 0.25,
            c: 1.0,
            cache_size: 100.0,
            eps: 0.001,
            shrinking: true,
            probability: true,
            ..Default::default()
        };

        let model = crate::train::svm_train(&problem, &param);
        assert_eq!(model.nr_class, 3);
        assert_eq!(model.prob_a.len(), 3); // 3 pairs

        for instance in problem.instances.iter().take(10) {
            let result = predict_probability(&model, instance);
            assert!(result.is_some());
            let (_label, probs) = result.unwrap();
            assert_eq!(probs.len(), 3);
            let sum: f64 = probs.iter().sum();
            assert!(
                (sum - 1.0).abs() < 1e-6,
                "probs sum to {}, expected 1.0",
                sum
            );
        }
    }

    #[test]
    fn predict_matches_c_svm_predict() {
        // Run C svm-predict and compare outputs
        // First, let's verify our predictions match by checking a few specific instances
        let model = load_model(&data_dir().join("heart_scale.model")).unwrap();
        let problem = load_problem(&data_dir().join("heart_scale")).unwrap();

        // Run C svm-predict to get reference predictions
        let c_predict = std::path::Path::new(env!("CARGO_MANIFEST_DIR"))
            .join("..")
            .join("..")
            .join("vendor")
            .join("libsvm")
            .join("svm-predict");

        if !c_predict.exists() {
            // Skip if C binary not compiled
            return;
        }

        // Write predictions to a temp file
        let output_path = data_dir().join("heart_scale.predict_test");
        let status = std::process::Command::new(&c_predict)
            .args([
                data_dir().join("heart_scale").to_str().unwrap(),
                data_dir().join("heart_scale.model").to_str().unwrap(),
                output_path.to_str().unwrap(),
            ])
            .output();

        if let Ok(output) = status {
            if output.status.success() {
                let c_preds: Vec<f64> = std::fs::read_to_string(&output_path)
                    .unwrap()
                    .lines()
                    .filter(|l| !l.is_empty())
                    .map(|l| l.trim().parse().unwrap())
                    .collect();

                assert_eq!(c_preds.len(), problem.labels.len());

                let mut mismatches = 0;
                for (i, instance) in problem.instances.iter().enumerate() {
                    let rust_pred = predict(&model, instance);
                    if rust_pred != c_preds[i] {
                        mismatches += 1;
                    }
                }

                assert_eq!(
                    mismatches, 0,
                    "{} predictions differ between Rust and C",
                    mismatches
                );

                // Clean up
                let _ = std::fs::remove_file(&output_path);
            }
        }
    }
}