quantrs2_ml/anomaly_detection/

preprocessing.rs

//! Data preprocessing for quantum anomaly detection

use crate::error::{MLError, Result};
use ndarray::{Array1, Array2, Axis};
use rand::Rng;

use super::config::{
    DimensionalityReduction, FeatureSelection, MissingValueStrategy, NoiseFiltering,
    NormalizationType, PreprocessingConfig,
};

/// Data preprocessor
#[derive(Debug)]
pub struct DataPreprocessor {
    config: PreprocessingConfig,
    fitted: bool,
    normalization_params: Option<NormalizationParams>,
    feature_selector: Option<FeatureSelector>,
    dimensionality_reducer: Option<DimensionalityReducer>,
}

/// Normalization parameters
#[derive(Debug, Clone)]
pub struct NormalizationParams {
    pub means: Array1<f64>,
    pub stds: Array1<f64>,
    pub mins: Array1<f64>,
    pub maxs: Array1<f64>,
}

/// Feature selector
#[derive(Debug)]
pub struct FeatureSelector {
    pub selected_features: Vec<usize>,
    pub feature_scores: Array1<f64>,
}

/// Dimensionality reducer
#[derive(Debug)]
pub struct DimensionalityReducer {
    pub components: Array2<f64>,
    pub explained_variance: Array1<f64>,
    pub target_dim: usize,
}

impl DataPreprocessor {
    /// Create new preprocessor
    pub fn new(config: PreprocessingConfig) -> Self {
        DataPreprocessor {
            config,
            fitted: false,
            normalization_params: None,
            feature_selector: None,
            dimensionality_reducer: None,
        }
    }

    /// Fit and transform data
    pub fn fit_transform(&mut self, data: &Array2<f64>) -> Result<Array2<f64>> {
        self.fit(data)?;
        self.transform(data)
    }

    /// Fit preprocessor to data
    pub fn fit(&mut self, data: &Array2<f64>) -> Result<()> {
        // Compute normalization parameters
        self.normalization_params = Some(self.compute_normalization_params(data));

        let mut current_data = data.clone();

        // Apply normalization first
        if let Some(ref params) = self.normalization_params {
            current_data = self.apply_normalization(&current_data, params)?;
        }

        // Fit feature selector if configured
        if self.config.feature_selection.is_some() {
            self.feature_selector = Some(self.fit_feature_selector(&current_data)?);
            // Apply feature selection to get the reduced data
            if let Some(ref selector) = self.feature_selector {
                current_data = self.apply_feature_selection(&current_data, selector)?;
            }
        }

        // Fit dimensionality reducer if configured (on feature-selected data)
        if self.config.dimensionality_reduction.is_some() {
            self.dimensionality_reducer = Some(self.fit_dimensionality_reducer(&current_data)?);
        }

        self.fitted = true;
        Ok(())
    }

    /// Transform data
    pub fn transform(&self, data: &Array2<f64>) -> Result<Array2<f64>> {
        if !self.fitted {
            return Err(MLError::MLOperationError(
                "Preprocessor must be fitted before transform".to_string(),
            ));
        }

        let mut transformed = data.clone();

        // Apply normalization
        if let Some(ref params) = self.normalization_params {
            transformed = self.apply_normalization(&transformed, params)?;
        }

        // Apply feature selection
        if let Some(ref selector) = self.feature_selector {
            transformed = self.apply_feature_selection(&transformed, selector)?;
        }

        // Apply dimensionality reduction
        if let Some(ref reducer) = self.dimensionality_reducer {
            transformed = self.apply_dimensionality_reduction(&transformed, reducer)?;
        }

        Ok(transformed)
    }

    /// Compute normalization parameters
    fn compute_normalization_params(&self, data: &Array2<f64>) -> NormalizationParams {
        let n_features = data.ncols();
        let mut means = Array1::zeros(n_features);
        let mut stds = Array1::zeros(n_features);
        let mut mins = Array1::zeros(n_features);
        let mut maxs = Array1::zeros(n_features);

        for j in 0..n_features {
            let column = data.column(j);
            means[j] = column.mean().unwrap_or(0.0);
            stds[j] = column.std(0.0);
            mins[j] = column.fold(f64::INFINITY, |a, &b| a.min(b));
            maxs[j] = column.fold(f64::NEG_INFINITY, |a, &b| a.max(b));
        }

        NormalizationParams {
            means,
            stds,
            mins,
            maxs,
        }
    }

    /// Apply normalization
    fn apply_normalization(
        &self,
        data: &Array2<f64>,
        params: &NormalizationParams,
    ) -> Result<Array2<f64>> {
        let mut normalized = data.clone();

        match self.config.normalization {
            NormalizationType::ZScore => {
                for j in 0..data.ncols() {
                    let mut column = normalized.column_mut(j);
                    if params.stds[j] > 1e-8 {
                        column.mapv_inplace(|x| (x - params.means[j]) / params.stds[j]);
                    }
                }
            }
            NormalizationType::MinMax => {
                for j in 0..data.ncols() {
                    let mut column = normalized.column_mut(j);
                    let range = params.maxs[j] - params.mins[j];
                    if range > 1e-8 {
                        column.mapv_inplace(|x| (x - params.mins[j]) / range);
                    }
                }
            }
            NormalizationType::Robust => {
                // Robust scaling using median and IQR
                for j in 0..data.ncols() {
                    let mut column_data: Vec<f64> = data.column(j).to_vec();
                    column_data.sort_by(|a, b| a.partial_cmp(b).unwrap());

                    let median = if column_data.len() % 2 == 0 {
                        (column_data[column_data.len() / 2 - 1]
                            + column_data[column_data.len() / 2])
                            / 2.0
                    } else {
                        column_data[column_data.len() / 2]
                    };

                    let q1 = column_data[column_data.len() / 4];
                    let q3 = column_data[3 * column_data.len() / 4];
                    let iqr = q3 - q1;

                    let mut column = normalized.column_mut(j);
                    if iqr > 1e-8 {
                        column.mapv_inplace(|x| (x - median) / iqr);
                    }
                }
            }
            NormalizationType::Quantum => {
                // Quantum normalization (placeholder - would use quantum circuits)
                for j in 0..data.ncols() {
                    let mut column = normalized.column_mut(j);
                    let norm = column.dot(&column).sqrt();
                    if norm > 1e-8 {
                        column.mapv_inplace(|x| x / norm);
                    }
                }
            }
        }

        Ok(normalized)
    }

    /// Fit feature selector
    fn fit_feature_selector(&self, data: &Array2<f64>) -> Result<FeatureSelector> {
        let n_features = data.ncols();

        let feature_scores = match &self.config.feature_selection {
            Some(FeatureSelection::Variance) => self.compute_variance_scores(data),
            Some(FeatureSelection::Correlation) => self.compute_correlation_scores(data),
            Some(FeatureSelection::MutualInformation) => {
                self.compute_mutual_information_scores(data)
            }
            Some(FeatureSelection::QuantumInformation) => {
                self.compute_quantum_information_scores(data)
            }
            None => Array1::zeros(n_features),
        };

        // Select top features
        let mut indexed_scores: Vec<(usize, f64)> = feature_scores
            .iter()
            .enumerate()
            .map(|(i, &score)| (i, score))
            .collect();

        indexed_scores.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap());

        let num_selected = (n_features / 2).max(1);
        let selected_features: Vec<usize> = indexed_scores
            .into_iter()
            .take(num_selected)
            .map(|(idx, _)| idx)
            .collect();

        Ok(FeatureSelector {
            selected_features,
            feature_scores,
        })
    }

    /// Apply feature selection
    fn apply_feature_selection(
        &self,
        data: &Array2<f64>,
        selector: &FeatureSelector,
    ) -> Result<Array2<f64>> {
        let selected_data = data.select(Axis(1), &selector.selected_features);
        Ok(selected_data)
    }

    /// Fit dimensionality reducer
    fn fit_dimensionality_reducer(&self, data: &Array2<f64>) -> Result<DimensionalityReducer> {
        let n_features = data.ncols();
        let target_dim = (n_features / 2).max(1);

        match &self.config.dimensionality_reduction {
            Some(DimensionalityReduction::PCA) => self.fit_pca(data, target_dim),
            Some(DimensionalityReduction::ICA) => self.fit_ica(data, target_dim),
            Some(DimensionalityReduction::UMAP) => self.fit_umap(data, target_dim),
            Some(DimensionalityReduction::QuantumPCA) => self.fit_quantum_pca(data, target_dim),
            Some(DimensionalityReduction::QuantumManifold) => {
                self.fit_quantum_manifold(data, target_dim)
            }
            None => {
                // Fallback to identity
                let components = Array2::eye(n_features);
                let explained_variance = Array1::ones(n_features);
                Ok(DimensionalityReducer {
                    components,
                    explained_variance,
                    target_dim: n_features,
                })
            }
        }
    }

    /// Apply dimensionality reduction
    fn apply_dimensionality_reduction(
        &self,
        data: &Array2<f64>,
        reducer: &DimensionalityReducer,
    ) -> Result<Array2<f64>> {
        let reduced = data.dot(&reducer.components.t());
        Ok(reduced)
    }

    // Helper methods for feature selection

    fn compute_variance_scores(&self, data: &Array2<f64>) -> Array1<f64> {
        let n_features = data.ncols();
        let mut scores = Array1::zeros(n_features);

        for j in 0..n_features {
            let column = data.column(j);
            scores[j] = column.var(0.0);
        }

        scores
    }

    fn compute_correlation_scores(&self, data: &Array2<f64>) -> Array1<f64> {
        // Placeholder: compute feature correlations
        let n_features = data.ncols();
        Array1::from_vec(
            (0..n_features)
                .map(|_| rand::thread_rng().gen::<f64>())
                .collect(),
        )
    }

    fn compute_mutual_information_scores(&self, data: &Array2<f64>) -> Array1<f64> {
        // Placeholder: compute mutual information
        let n_features = data.ncols();
        Array1::from_vec(
            (0..n_features)
                .map(|_| rand::thread_rng().gen::<f64>())
                .collect(),
        )
    }

    fn compute_quantum_information_scores(&self, data: &Array2<f64>) -> Array1<f64> {
        // Placeholder: compute quantum information scores
        let n_features = data.ncols();
        Array1::from_vec(
            (0..n_features)
                .map(|_| rand::thread_rng().gen::<f64>())
                .collect(),
        )
    }

    // Helper methods for dimensionality reduction

    fn fit_pca(&self, data: &Array2<f64>, target_dim: usize) -> Result<DimensionalityReducer> {
        // Placeholder PCA implementation
        let n_features = data.ncols();
        let components =
            Array2::from_shape_fn(
                (target_dim, n_features),
                |(i, j)| {
                    if i == j {
                        1.0
                    } else {
                        0.0
                    }
                },
            );

        let explained_variance =
            Array1::from_vec((0..target_dim).map(|i| 1.0 / (i + 1) as f64).collect());

        Ok(DimensionalityReducer {
            components,
            explained_variance,
            target_dim,
        })
    }

    fn fit_ica(&self, data: &Array2<f64>, target_dim: usize) -> Result<DimensionalityReducer> {
        // Placeholder ICA implementation
        self.fit_pca(data, target_dim)
    }

    fn fit_umap(&self, data: &Array2<f64>, target_dim: usize) -> Result<DimensionalityReducer> {
        // Placeholder UMAP implementation
        self.fit_pca(data, target_dim)
    }

    fn fit_quantum_pca(
        &self,
        data: &Array2<f64>,
        target_dim: usize,
    ) -> Result<DimensionalityReducer> {
        // Placeholder Quantum PCA implementation
        self.fit_pca(data, target_dim)
    }

    fn fit_quantum_manifold(
        &self,
        data: &Array2<f64>,
        target_dim: usize,
    ) -> Result<DimensionalityReducer> {
        // Placeholder Quantum Manifold implementation
        self.fit_pca(data, target_dim)
    }
}