sklears_feature_selection/automl/
method_selector.rs

1//! Method Selection Module for AutoML Feature Selection
2//!
3//! Intelligently selects appropriate feature selection methods based on data characteristics.
4//! All implementations follow the SciRS2 policy using scirs2-core for numerical computations.
5
6use super::automl_core::{AutoMLMethod, DataCharacteristics, TargetType};
7use sklears_core::error::Result as SklResult;
8
9type Result<T> = SklResult<T>;
10
11/// Method selector for choosing appropriate feature selection methods
12#[derive(Debug, Clone)]
13pub struct MethodSelector;
14
15impl MethodSelector {
16    pub fn new() -> Self {
17        Self
18    }
19
20    pub fn select_methods(
21        &self,
22        characteristics: &DataCharacteristics,
23    ) -> Result<Vec<AutoMLMethod>> {
24        let mut selected_methods = Vec::new();
25
26        // Rule-based method selection
27        match characteristics.target_type {
28            TargetType::BinaryClassification | TargetType::MultiClassification => {
29                // Always include univariate filtering for classification
30                selected_methods.push(AutoMLMethod::UnivariateFiltering);
31
32                if characteristics.n_features > 100 {
33                    selected_methods.push(AutoMLMethod::LassoBased);
34                }
35
36                if characteristics.computational_budget.allow_complex_methods {
37                    selected_methods.push(AutoMLMethod::TreeBased);
38                }
39            }
40            TargetType::Regression => {
41                selected_methods.push(AutoMLMethod::CorrelationBased);
42
43                if characteristics.n_features > 50 {
44                    selected_methods.push(AutoMLMethod::LassoBased);
45                }
46
47                if characteristics.computational_budget.allow_complex_methods {
48                    selected_methods.push(AutoMLMethod::WrapperBased);
49                }
50            }
51            TargetType::MultiLabel => {
52                selected_methods.push(AutoMLMethod::UnivariateFiltering);
53                selected_methods.push(AutoMLMethod::EnsembleBased);
54            }
55            TargetType::Survival => {
56                selected_methods.push(AutoMLMethod::UnivariateFiltering);
57                selected_methods.push(AutoMLMethod::CorrelationBased);
58            }
59        }
60
61        // Add correlation-based if high correlation detected
62        if characteristics.correlation_structure.high_correlation_pairs
63            > characteristics.n_features / 4
64            && !selected_methods.contains(&AutoMLMethod::CorrelationBased)
65        {
66            selected_methods.push(AutoMLMethod::CorrelationBased);
67        }
68
69        // Add ensemble method for complex scenarios
70        if characteristics.feature_to_sample_ratio > 0.5
71            && characteristics.computational_budget.allow_complex_methods
72        {
73            selected_methods.push(AutoMLMethod::EnsembleBased);
74        }
75
76        // Hybrid approach for very complex data
77        if characteristics.n_features > 1000 && characteristics.n_samples > 1000 {
78            selected_methods.push(AutoMLMethod::Hybrid);
79        }
80
81        // Neural Architecture Search for very high-dimensional data
82        if characteristics.feature_to_sample_ratio > 2.0
83            && characteristics.computational_budget.allow_complex_methods
84            && !characteristics.computational_budget.prefer_speed
85        {
86            selected_methods.push(AutoMLMethod::NeuralArchitectureSearch);
87        }
88
89        // Transfer Learning when we have moderate complexity but good compute budget
90        if characteristics.n_features > 100
91            && characteristics.n_samples >= 500
92            && characteristics.computational_budget.allow_complex_methods
93        {
94            selected_methods.push(AutoMLMethod::TransferLearning);
95        }
96
97        // Meta-Learning Ensemble for complex scenarios with sufficient data
98        if characteristics.n_features > 50
99            && characteristics.n_samples > 200
100            && characteristics.computational_budget.allow_complex_methods
101            && selected_methods.len() >= 2
102        {
103            selected_methods.push(AutoMLMethod::MetaLearningEnsemble);
104        }
105
106        if selected_methods.is_empty() {
107            // Fallback to univariate filtering
108            selected_methods.push(AutoMLMethod::UnivariateFiltering);
109        }
110
111        Ok(selected_methods)
112    }
113}
114
115impl Default for MethodSelector {
116    fn default() -> Self {
117        Self::new()
118    }
119}
sklears_feature_selection/automl/method_selector.rs

sklears_feature_selection/automl/
method_selector.rs
