Implementation:Online ml River NaiveBayes ComplementNB

Knowledge Sources	Online_ml_River
Domains	Online_Learning, Naive_Bayes, Imbalanced_Learning
Last Updated	2026-02-08 16:00 GMT

Overview

Complement Naive Bayes is designed for imbalanced datasets by estimating parameters from complement classes rather than the class itself.

Description

Unlike standard multinomial NB, Complement NB computes probabilities using all classes except the target class (the "complement"). For each class c, it calculates feature probabilities from the complement set: P(f|~c) = (total_f - count_f_in_c) / sum(complement). This approach reduces bias toward majority classes in imbalanced data. The model maintains feature counts per class, class totals, and overall feature totals. During prediction, it computes negative log-likelihoods, where lower values indicate better fit.

Usage

Use Complement NB specifically for imbalanced text classification problems where standard Multinomial NB may be biased toward frequent classes. It works with count or TF-IDF features and requires positive input values. Particularly effective when class sizes vary significantly. Supports both online and mini-batch learning modes.

Code Reference

Source Location

Repository: Online_ml_River
File: river/naive_bayes/complement.py

Signature

class ComplementNB(base.BaseNB):
    def __init__(self, alpha=1.0):
        self.alpha = alpha
        self.class_counts = collections.Counter()
        self.feature_counts = collections.defaultdict(collections.Counter)
        self.feature_totals = collections.Counter()
        self.class_totals = collections.Counter()

Import

from river import naive_bayes

I/O Contract

Parameters

Parameter	Type	Default	Description
alpha	float	1.0	Additive smoothing parameter

Attributes

Attribute	Type	Description
class_counts	Counter	Number of instances per class
feature_counts	defaultdict	Feature counts per class
feature_totals	Counter	Total counts per feature across all classes
class_totals	Counter	Total feature counts per class

Input/Output

Method	Input	Output
learn_one	x: dict, y: Any	None
learn_many	X: DataFrame, y: Series	None
predict_proba_one	x: dict	dict
predict_proba_many	X: DataFrame	DataFrame

Usage Examples

import pandas as pd
from river import compose
from river import feature_extraction
from river import naive_bayes

docs = [
    ("Chinese Beijing Chinese", "yes"),
    ("Chinese Chinese Shanghai", "yes"),
    ("Chinese Macao", "maybe"),
    ("Tokyo Japan Chinese", "no")
]

# Single instance learning
model = compose.Pipeline(
    ("tokenize", feature_extraction.BagOfWords(lowercase=False)),
    ("nb", naive_bayes.ComplementNB(alpha=1))
)

for sentence, label in docs:
    model.learn_one(sentence, label)

model["nb"].p_class("yes")
# 0.5

model["nb"].p_class("no")
# 0.25

model["nb"].p_class("maybe")
# 0.25

model.predict_proba_one("test")
# {'yes': 0.275, 'maybe': 0.375, 'no': 0.35}

model.predict_one("test")
# 'maybe'

# Mini-batch learning
X = pd.Series([
   "Chinese Beijing Chinese",
   "Chinese Chinese Shanghai",
   "Chinese Macao",
   "Tokyo Japan Chinese"
])

y = pd.Series(["yes", "yes", "maybe", "no"])

model = compose.Pipeline(
    ("tokenize", feature_extraction.BagOfWords(lowercase=False)),
    ("nb", naive_bayes.ComplementNB(alpha=1))
)

model.learn_many(X, y)

unseen = pd.Series(["Taiwanese Taipei", "Chinese Shanghai"])

model.predict_proba_many(unseen)
#       maybe        no       yes
# 0  0.415129  0.361624  0.223247
# 1  0.248619  0.216575  0.534807

model.predict_many(unseen)
# 0    maybe
# 1      yes
# dtype: object

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment