naivebayes 0.1.2

use std::collections::HashMap;
use std::collections::HashSet;
use std::collections::hash_map::Keys;
use std::iter::FromIterator;
use std::vec::Vec;

struct Attributes {
    attributes: HashMap<String, HashMap<String, i64>>,
}

impl Attributes {
    pub fn new() -> Attributes {
        Attributes {
            attributes: HashMap::new(),
        }
    }

    fn add(&mut self, attribute: &String, label: &String) {
        let labels = self.attributes
            .entry(attribute.to_string())
            .or_insert(HashMap::new());
        let value = labels.entry((*label).to_string()).or_insert(0);
        *value += 1;
    }

    fn get_frequency(&mut self, attribute: &String, label: &String) -> (Option<&i64>, bool) {
        match self.attributes.get(attribute) {
            Some(labels) => match labels.get(label) {
                Some(value) => return (Some(value), true),
                None => return (None, true),
            },
            None => return (None, false),
        }
    }
}

struct Labels {
    counts: HashMap<String, i64>,
}

impl Labels {
    pub fn new() -> Labels {
        Labels {
            counts: HashMap::new(),
        }
    }

    fn add(&mut self, label: &String) {
        let value = self.counts.entry(label.to_string()).or_insert(0);
        *value += 1;
    }

    fn get_count(&mut self, label: &String) -> Option<&i64> {
        return self.counts.get(label);
    }

    fn get_labels(&mut self) -> Keys<String, i64> {
        return self.counts.keys();
    }

    fn get_total(&mut self) -> i64 {
        return self.counts.values().fold(0, |acc, x| acc + x);
    }
}

struct Model {
    labels: Labels,
    attributes: Attributes,
}

impl Model {
    pub fn new() -> Model {
        Model {
            labels: Labels::new(),
            attributes: Attributes::new(),
        }
    }
    fn train(&mut self, data: &Vec<String>, label: &String) {
        self.labels.add(label);
        for attribute in data {
            self.attributes.add(attribute, label);
        }
    }
}

pub struct NaiveBayes {
    model: Model,
    minimum_probability: f64,
    minimum_log_probability: f64
}

impl NaiveBayes {
    /// creates a new instance of a `NaiveBayes` classifier.
    pub fn new() -> NaiveBayes {
        NaiveBayes {
            model: Model::new(),
            minimum_probability: 1e-9,
            minimum_log_probability: -100.0,
        }
    }

    fn prior(&mut self, label: &String) -> Option<f64> {
        let total = *(&self.model.labels.get_total()) as f64;
        let label = &self.model.labels.get_count(label);
        if label.is_some() && total > 0.0 {
            return Some(*label.unwrap() as f64 / total);
        } else {
            return None;
        }
    }

    fn log_prior(&mut self, label: &String) -> Option<f64> {
        let total = *(&self.model.labels.get_total()) as f64;
        let label = &self.model.labels.get_count(label);
        if label.is_some() && total > 0.0 {
            return Some((*label.unwrap() as f64).ln() - total.ln());
        } else {
            return None;
        }
    }

    fn calculate_attr_prob(&mut self, attribute: &String, label: &String) -> Option<f64> {
        match self.model.attributes.get_frequency(attribute, label) {
            (Some(frequency), true) => match self.model.labels.get_count(label) {
                Some(count) => return Some((*frequency as f64) / (*count as f64)),
                None => return None,
            },
            (None, true) => return Some(self.minimum_probability),
            (None, false) => return None,
            (Some(_), false) => None,
        }
    }

    fn calculate_attr_log_prob(&mut self, attribute: &String, label: &String) -> Option<f64> {
        match self.model.attributes.get_frequency(attribute, label) {
            (Some(frequency), true) => match self.model.labels.get_count(label) {
                Some(count) => return Some((*frequency as f64).ln() - (*count as f64).ln()),
                None => return None,
            },
            (None, true) => return Some(self.minimum_log_probability),
            (None, false) => return None,
            (Some(_), false) => None,
        }
    }

    fn label_prob(&mut self, label: &String, attrs: &HashSet<String>) -> Vec<f64> {
        let mut probs: Vec<f64> = Vec::new();
        for attr in attrs {
            match self.calculate_attr_prob(attr, label) {
                Some(p) => {
                    probs.push(p);
                }
                None => {}
            }
        }
        return probs;
    }

    fn label_log_prob(&mut self, label: &String, attrs: &HashSet<String>) -> Vec<f64> {
        let mut probs: Vec<f64> = Vec::new();
        for attr in attrs {
            match self.calculate_attr_log_prob(attr, label) {
                Some(p) => {
                    probs.push(p);
                }
                None => {}
            }
        }
        return probs;
    }

    /// trains the model with a `Vec<String>` of tokens, associating it with a `String` label.
    pub fn train(&mut self, data: &Vec<String>, label: &String) {
        self.model.train(data, label);
    }

    /// classify a `Vec<String>` of tokens returning a map of tokens and probabilities
    /// as keys and values, respectively.
    pub fn classify(&mut self, data: &Vec<String>) -> HashMap<String, f64> {
        let attribute_set: HashSet<String> = HashSet::from_iter(data.iter().cloned());
        let mut result: HashMap<String, f64> = HashMap::new();
        let labels: HashSet<String> =
            HashSet::from_iter(self.model.labels.get_labels().into_iter().cloned());
        for label in labels {
            let p = self.label_prob(&label, &attribute_set);
            let p_iter = p.into_iter().fold(1.0, |acc, x| acc * x);
            let _value = result
                .entry(label.to_string())
                .or_insert(p_iter * self.prior(&label).unwrap());
        }

        return result;
    }

    /// classify a `Vec<String>` of tokens returning a map of tokens and log-probabilities
    /// as keys and values, respectively. Using `log_classify` may prevent underflows.
    pub fn log_classify(&mut self, data: &Vec<String>) -> HashMap<String, f64> {
        let attribute_set: HashSet<String> = HashSet::from_iter(data.iter().cloned());
        let mut result: HashMap<String, f64> = HashMap::new();
        let labels: HashSet<String> =
            HashSet::from_iter(self.model.labels.get_labels().into_iter().cloned());
        for label in labels {
            let p = self.label_log_prob(&label, &attribute_set);
            let max = p.iter().cloned().fold(-1./0. /* inf */, f64::max);
            let p_iter = p.into_iter().fold(0.0, |acc, x| acc + (x - max).exp());
            let _value = result
                .entry(label.to_string())
                .or_insert(max + p_iter.ln() + self.log_prior(&label).unwrap());
        }

        return result;
    }
}

#[cfg(test)]
mod test_attributes {
    use super::*;

    #[test]
    fn attribute_add() {
        let mut model = Attributes::new();
        model.add(&"rust".to_string(), &"naive".to_string());
        assert_eq!(
            *model
                .get_frequency(&"rust".to_string(), &"naive".to_string())
                .0
                .unwrap(),
            1
        );
    }

    #[test]
    fn get_non_existing() {
        let mut model = Attributes::new();
        assert_eq!(
            model
                .get_frequency(&"rust".to_string(), &"naive".to_string())
                .0,
            None
        );
    }

}

#[cfg(test)]
mod test_labels {

    use super::*;

    #[test]
    fn label_add() {
        let mut labels = Labels::new();
        labels.add(&"rust".to_string());
        assert_eq!(*labels.get_count(&"rust".to_string()).unwrap(), 1);
    }

    #[test]
    fn label_get_nonexistent() {
        let mut labels = Labels::new();
        assert_eq!(labels.get_count(&"rust".to_string()), None);
    }

    #[test]
    fn get_labels() {
        let mut labels = Labels::new();
        labels.add(&"rust".to_string());
        assert_eq!(labels.get_labels().len(), 1);
        assert_eq!(labels.get_labels().last().unwrap(), "rust");
    }

    #[test]
    fn get_counts() {
        let mut labels = Labels::new();
        labels.add(&"rust".to_string());
        labels.add(&"rust".to_string());
        assert_eq!(labels.get_labels().len(), 1);
        assert_eq!(*labels.get_count(&"rust".to_string()).unwrap(), 2);
    }

    #[test]
    fn get_nonexistent_counts() {
        let mut labels = Labels::new();
        assert_eq!(labels.get_labels().len(), 0);
        assert_eq!(labels.get_count(&"rust".to_string()), None);
    }

    #[test]
    fn get_nonexistent_total() {
        let mut labels = Labels::new();
        assert_eq!(labels.get_total(), 0);
    }

    #[test]
    fn get_total() {
        let mut labels = Labels::new();
        labels.add(&"rust".to_string());
        labels.add(&"rust".to_string());
        labels.add(&"naive".to_string());
        labels.add(&"bayes".to_string());
        assert_eq!(labels.get_total(), 4);
    }

}

#[cfg(test)]
mod test_naive_bayes {
    use super::*;
    use std::f64::consts::LN_2;

    #[test]
    fn test_prior() {
        let mut nb = NaiveBayes::new();
        let mut data: Vec<String> = Vec::new();
        data.push("rust".to_string());
        data.push("naive".to_string());
        data.push("bayes".to_string());
        nb.model.train(&data, &"👍".to_string());
        let prior = nb.prior(&"👍".to_string());
        assert_eq!(prior, Some(1.0));
    }

    #[test]
    fn test_log_prior() {
        let mut nb = NaiveBayes::new();
        let mut data: Vec<String> = Vec::new();
        data.push("rust".to_string());
        data.push("naive".to_string());
        data.push("bayes".to_string());
        nb.model.train(&data, &"👍".to_string());
        let prior = nb.log_prior(&"👍".to_string());
        assert_eq!(prior, Some(0.0));
    }

    #[test]
    fn test_prior_nonexistent() {
        let mut nb = NaiveBayes::new();
        let mut data: Vec<String> = Vec::new();
        data.push("rust".to_string());
        data.push("naive".to_string());
        data.push("bayes".to_string());
        nb.model.train(&data, &"👍".to_string());
        let prior = nb.prior(&"👎".to_string());
        assert_eq!(prior, None);
    }

    #[test]
    fn test_classification() {
        let mut nb = NaiveBayes::new();
        let mut data: Vec<String> = Vec::new();
        data.push("rust".to_string());
        data.push("naive".to_string());
        data.push("bayes".to_string());
        nb.model.train(&data, &"👍".to_string());
        let mut data2: Vec<String> = Vec::new();
        data2.push("golang".to_string());
        data2.push("java".to_string());
        data2.push("javascript".to_string());
        nb.model.train(&data2, &"👎".to_string());

        let classes = nb.classify(
            &(vec![
                "rust".to_string(),
                "scala".to_string(),
                "c++".to_string(),
            ]),
        );
        assert_eq!(classes.get(&"👍".to_string()).unwrap(), &0.5);
        assert_eq!(classes.get(&"👎".to_string()).unwrap(), &0.0000000005);
        print!("{:?}", classes);

    }

    #[test]
    fn test_log_classification() {
        let mut nb = NaiveBayes::new();
        let mut data: Vec<String> = Vec::new();
        data.push("rust".to_string());
        data.push("naive".to_string());
        data.push("bayes".to_string());
        nb.model.train(&data, &"👍".to_string());
        let mut data2: Vec<String> = Vec::new();
        data2.push("golang".to_string());
        data2.push("java".to_string());
        data2.push("javascript".to_string());
        nb.model.train(&data2, &"👎".to_string());

        let classes = nb.log_classify(
            &(vec![
                "rust".to_string(),
                "scala".to_string(),
                "c++".to_string(),
            ]),
        );
        assert_eq!(classes.get(&"👍".to_string()).unwrap(), &-LN_2);
        assert_eq!(classes.get(&"👎".to_string()).unwrap(), &-100.69314718055995);
        print!("{:?}", classes);

    }

}