rill_ml/preprocessing/
mean_imputer.rs

1//! Mean imputer for missing data.
2//!
3//! Replaces `NaN` values with the running mean of observed non-NaN
4//! values for each feature. Uses Welford's algorithm for numerical
5//! stability.
6
7use crate::error::{RillError, ensure_finite};
8use crate::traits::Transformer;
9
10/// Replaces `NaN` values with the per-feature running mean.
11///
12/// When a feature has seen zero non-NaN values, `NaN` is replaced
13/// with `0.0`. This transformer accepts `NaN` in its input.
14#[derive(Debug, Clone)]
15#[cfg_attr(feature = "serde", derive(serde::Serialize, serde::Deserialize))]
16pub struct MeanImputer {
17    feature_count: usize,
18    counts: Vec<u64>,
19    means: Vec<f64>,
20    samples_seen: u64,
21}
22
23impl MeanImputer {
24    /// Create a new imputer for `feature_count` features.
25    ///
26    /// # Errors
27    /// Returns [`RillError::EmptyFeatures`] if `feature_count` is `0`.
28    pub fn new(feature_count: usize) -> Result<Self, RillError> {
29        if feature_count == 0 {
30            return Err(RillError::EmptyFeatures);
31        }
32        Ok(Self {
33            feature_count,
34            counts: vec![0; feature_count],
35            means: vec![0.0; feature_count],
36            samples_seen: 0,
37        })
38    }
39
40    /// The per-feature running means of observed non-NaN values.
41    pub fn means(&self) -> &[f64] {
42        &self.means
43    }
44
45    /// The per-feature counts of observed non-NaN values.
46    pub fn counts(&self) -> &[u64] {
47        &self.counts
48    }
49
50    /// Validate only the dimension, allowing NaN values.
51    fn check_dimension(&self, features: &[f64]) -> Result<(), RillError> {
52        if features.is_empty() {
53            return Err(RillError::EmptyFeatures);
54        }
55        if features.len() != self.feature_count {
56            return Err(RillError::DimensionMismatch {
57                expected: self.feature_count,
58                actual: features.len(),
59            });
60        }
61        Ok(())
62    }
63
64    /// Update the running mean for feature `idx` using Welford's algorithm.
65    fn update_mean(&mut self, idx: usize, value: f64) {
66        let n = self.counts[idx] + 1;
67        self.counts[idx] = n;
68        let delta = value - self.means[idx];
69        self.means[idx] += delta / n as f64;
70    }
71}
72
73impl Transformer for MeanImputer {
74    fn input_dim(&self) -> usize {
75        self.feature_count
76    }
77
78    fn output_dim(&self) -> usize {
79        self.feature_count
80    }
81
82    fn transform(&self, features: &[f64]) -> Result<Vec<f64>, RillError> {
83        self.check_dimension(features)?;
84        Ok(features
85            .iter()
86            .enumerate()
87            .map(|(i, &x)| {
88                if x.is_nan() {
89                    if self.counts[i] == 0 {
90                        0.0
91                    } else {
92                        self.means[i]
93                    }
94                } else {
95                    x
96                }
97            })
98            .collect())
99    }
100
101    fn update(&mut self, features: &[f64]) -> Result<(), RillError> {
102        self.check_dimension(features)?;
103        for (i, &x) in features.iter().enumerate() {
104            if x.is_nan() {
105                continue;
106            }
107            ensure_finite("feature", x)?;
108            self.update_mean(i, x);
109        }
110        self.samples_seen += 1;
111        Ok(())
112    }
113
114    fn samples_seen(&self) -> u64 {
115        self.samples_seen
116    }
117
118    fn reset(&mut self) {
119        for c in &mut self.counts {
120            *c = 0;
121        }
122        for m in &mut self.means {
123            *m = 0.0;
124        }
125        self.samples_seen = 0;
126    }
127}
128
129#[cfg(test)]
130mod tests {
131    use super::*;
132
133    #[test]
134    fn nan_replaced_with_mean_after_update() {
135        let mut imp = MeanImputer::new(2).unwrap();
136        // feature 0: observed [2.0, 4.0] -> mean 3.0
137        // feature 1: observed [10.0] -> mean 10.0
138        imp.update(&[2.0, 10.0]).unwrap();
139        imp.update(&[4.0, f64::NAN]).unwrap();
140        let out = imp.transform(&[f64::NAN, f64::NAN]).unwrap();
141        assert!((out[0] - 3.0).abs() < 1e-12);
142        assert!((out[1] - 10.0).abs() < 1e-12);
143    }
144
145    #[test]
146    fn nan_replaced_with_zero_when_no_data() {
147        let imp = MeanImputer::new(2).unwrap();
148        let out = imp.transform(&[f64::NAN, f64::NAN]).unwrap();
149        assert_eq!(out, vec![0.0, 0.0]);
150    }
151
152    #[test]
153    fn non_nan_passed_through() {
154        let mut imp = MeanImputer::new(2).unwrap();
155        imp.update(&[5.0, 6.0]).unwrap();
156        let out = imp.transform(&[1.5, -2.0]).unwrap();
157        assert_eq!(out, vec![1.5, -2.0]);
158    }
159
160    #[test]
161    fn mean_updates_correctly() {
162        let mut imp = MeanImputer::new(1).unwrap();
163        imp.update(&[1.0]).unwrap();
164        imp.update(&[2.0]).unwrap();
165        imp.update(&[3.0]).unwrap();
166        assert!((imp.means()[0] - 2.0).abs() < 1e-12);
167        assert_eq!(imp.counts()[0], 3);
168    }
169
170    #[test]
171    fn nan_skipped_in_update() {
172        let mut imp = MeanImputer::new(2).unwrap();
173        // feature 0: [1.0, NaN, 3.0] -> mean 2.0 (NaN skipped)
174        // feature 1: [NaN, NaN, NaN] -> count 0, mean 0.0
175        imp.update(&[1.0, f64::NAN]).unwrap();
176        imp.update(&[f64::NAN, f64::NAN]).unwrap();
177        imp.update(&[3.0, f64::NAN]).unwrap();
178        assert!((imp.means()[0] - 2.0).abs() < 1e-12);
179        assert_eq!(imp.counts()[0], 2);
180        assert_eq!(imp.counts()[1], 0);
181        assert!((imp.means()[1] - 0.0).abs() < 1e-12);
182    }
183
184    #[test]
185    fn dimension_mismatch_rejected() {
186        let imp = MeanImputer::new(3).unwrap();
187        assert!(matches!(
188            imp.transform(&[1.0, 2.0]),
189            Err(RillError::DimensionMismatch { .. })
190        ));
191        let mut imp = imp;
192        assert!(matches!(
193            imp.update(&[1.0, 2.0, 3.0, 4.0]),
194            Err(RillError::DimensionMismatch { .. })
195        ));
196    }
197
198    #[test]
199    fn reset_clears_state() {
200        let mut imp = MeanImputer::new(2).unwrap();
201        imp.update(&[1.0, 2.0]).unwrap();
202        imp.update(&[3.0, 4.0]).unwrap();
203        assert_eq!(imp.samples_seen(), 2);
204        assert_eq!(imp.counts()[0], 2);
205        imp.reset();
206        assert_eq!(imp.samples_seen(), 0);
207        assert_eq!(imp.counts()[0], 0);
208        assert!((imp.means()[0] - 0.0).abs() < 1e-12);
209    }
210
211    #[test]
212    #[cfg(feature = "serde")]
213    fn serde_roundtrip() {
214        let mut imp = MeanImputer::new(2).unwrap();
215        imp.update(&[1.0, f64::NAN]).unwrap();
216        imp.update(&[3.0, 5.0]).unwrap();
217        let json = serde_json::to_string(&imp).unwrap();
218        let restored: MeanImputer = serde_json::from_str(&json).unwrap();
219        assert_eq!(restored.input_dim(), imp.input_dim());
220        assert_eq!(restored.output_dim(), imp.output_dim());
221        assert_eq!(restored.samples_seen(), imp.samples_seen());
222        assert_eq!(restored.counts(), imp.counts());
223        assert!((restored.means()[0] - imp.means()[0]).abs() < 1e-12);
224    }
225}
rill_ml/preprocessing/mean_imputer.rs

rill_ml/preprocessing/
mean_imputer.rs
