snob-factor 2026.3.17

Mixture modelling by Minimum Message Length (MML) using Factor Snob

About this Project

This repository is an updated version of the open-source Factor Snob program for Mixture modelling by Minimum Message Length (MML) written in C. The original source code was obtained from Lloyd Allison's website at: http://www.allisons.org/ll/MML/Notes/SNOB-factor/. The goal of this project is to make the program more accessible by providing a Python interface to the program. The original program was written in C and compiled to a binary executable.

I have made many changes to the code mainly to improve readability, and to fix some minor bugs I discovered. The current version produces the same output as the original, except for a couple of display bugs present in the original version (see below). Here is a summary of the changes:

• Refactored function declarations into a central "snob.h" include file.
• Previous version included "glob.c" within other files. "glob.c" now contains some generic and exported functions.
• Replaced Old-style C-function prototypes with moder C function prototypes.
• Removed as many global variables as possible and eliminated many side effects. The previous version relied too much on
  side effects and global variables. The current version has made variable passing more explicit. In some cases where global variables are still used, they have been grouped into global structures instead. This should make it easier to eliminate them entirely in the future. There is still plenty of room for improvement.
• Refactored many functions to replace goto statements with more modern control structures.
• Replaced many print statements with a logging facility that can be dialed down using a debug level.
• Added functions for printing a summary of the existing sample data
• Added functions for generating a JSON representation of the classification result equivalent to the data produced by the prclass command.
• Added functions for generating class assignments to the sample items. Also, useful for classification through Python.
• Fixed a bug in which during printing of Class information, the wrong class pointer was used to calculate one of the Factor costs. This is why the total costs for the Factor model did not match the sum of the parameter, data and variable costs. It also affected the display of the boost character on the class header information.
• Added functions to allow loading vset and Sample data through function calls rather than from file. This allows vset and sample information to be supplied through Python ctypes calls.
• Added functions to allow saving and loading of the population model to and from file. This allows the population model to be saved and loaded through Python ctypes calls.
• Added a Scikit-learn style interface for fitting and predicting using the Factor Snob model. This allows the model to be used in a similar way to other Scikit-learn classification models. The features include the ability to save fitted models and to predict on new data based on previously saved models.
• Added a python module that can provide the data, run the Factor Snob routines and extract the results within Python, using ctypes.

The original documentation file is included in the docs folder. It contains a lot of useful information about the original SNOB programs and may be useful for understanding some of the input and output parameters.

Any published work using this software should cite the original author's work as well as this repository. See the

The MML Book

[1] C. S. Wallace, Statistical and Inductive Inference by Minimum Message Length, Springer, isbn-13:978-0-387-23795-4, 2005.

Using the Python Module

The Python module can be installed using pip as follows:

pip install snob-factor

The module can be used in a similar way to other Scikit-learn classification models. The following example shows how to use the module to fit a Factor Snob model to some sample data and then use the model to predict the class of some new data.

import pandas as pd
import snob

# Load the sample data
train = pd.read_csv('train.csv')  
sfc = snob.SNOBClassifier(
    name='sst',
    attrs={                   # these are the features of the data
        'distance': 'real',   # a real-valued attribute
        'theta': 'radians',   # an angle in radians angles are treated specially
        'phi': 'radians',     # another angle in radians 
    },
    cycles=3,                 # maximum number of cycles to run
    steps=40,                 # maximum number of estimation/assignment steps to run
    moves=2,                  # maximum number of failed class shuffles before giving up each cycle
    tol=0.01,                 # minimum percentage improvement in cost to indicate convergence
    seed=1234567              # random number seed
)

# Fit the model to the sample data
sfc.fit(train)
sfc.save_model('sst.mod')     # save the model to a file for use later

classes = sfc.get_classes()   # get the class parameters for the fitted model
snob.show_classes(classes)    # display the classification summary

# get class assignments for training data
train_pred = sfc.predict()    # assignments the training data, note that predict is called without arguments
print(train_pred)             # train_pred is a pandas DataFrame with the class assignments

# Load some new data to predict
test = pd.read_csv('test.csv')

# Predict the class of the new data
test_pred = sfc.predict(test)
print(test_pred)

The SNOBClassifier class has a number of parameters that can be used to control the behaviour of the model. The attrs parameter is a dictionary that defines the features of the data. The keys are the names of the features and the values are the types of the features. The types can be one of the following:

• 'real' - a real-valued feature
• 'radians' - an angle in radians
• 'degrees' - an angle in degrees
• 'multi-state' - a categorical feature with more than two but preferably fewer than 20 states
• 'binary' - a boolean or two-valued feature

Since SNOBClassifier is an unsupervised learning model, the fit() method does not require a target variable. The fit() method takes a pandas DataFrame as input. The DataFrame must contain columns for each of the features defined in the attrs parameter. The fit() method will run the Factor Snob algorithm on the data and produce a classification model for the data. After fitting, the classifier will be fully parameterized and can be used to predict the classes of new data.

The get_classes() method of a fully parameterized classifier can be used to get the class parameters for the fitted model. This returns a list of dictionaries with each dictionary representing one class. The show_classes() function from the snob package can be used to display a summary of the class parameters. The show_classes() function takes the list of class dictionaries as returned by the get_classes() method.

The model can be saved to a file using the save_model(). This method takes a single argument which is the name of the file to save the model to. A previously saved model, can be used by specifying a from_file parameter during initialization of the classifier. The from_file parameter should be set to the name of the file containing the saved model. The attrs parameter is always required even when a from_file parameter is provided. The attrs parameter should be the same as the one used to create the saved model.

The predict() method can then be used to predict the class of new data. The first time a restored model is used to for prediction, the model will be loaded into memory and used to fully parameterize the classifer before prections are performed. Details of class parameters will only be available after the classifer is fully parameterized.

The predict() method takes a pandas DataFrame as input. The DataFrame must contain columns for each of the features.

For example, the model above can be loaded from the saved model file and used as follows:

sfc = snob.SNOBClassifier(
    name='sst',
    attrs={                   # these are the features of the data
        'distance': 'real',   # a real-valued attribute
        'theta': 'radians',   # an angle in radians angles are treated specially
        'phi': 'radians',     # another angle in radians 
    },
    from_file='sst.mod'       # load the model from the file
)
new_data = pd.read_csv('new_data.csv')
new_pred = sfc.predict(new_data)    # No need to run fit again, the model will be loaded from the file

class_info = sfc.get_classes()      # must run predict first to fully parameterize the model
snob.show_classes(class_info)
print(new_pred)

Description:

This program implements an unsupervised classification (or clustering) algorithm based on the Minimum Message Length (MML) principle. The fundamental goal is to find the best model to explain the structure of your data, where "best" means the model that allows for the most compact description of the data.

The total "message length" is the sum of two parts:

1. Part 1: The Model Cost: The length of the message required to describe the classification model itself (i.e., the number of classes and all their parameters).
2. Part 2: The Data Cost: The length of the message required to describe the data, given the model.

The algorithm works as a two-level iterative process to find the model that minimizes this total message length.

1. The Outer Loop: Model Discovery ("Surgery")

This is the high-level search for the optimal number of classes and their relationships. The algorithm starts with an initial set of classes and then iteratively tries to improve the model by performing "surgical" operations:

• Splitting: A single class is split into two.
• Merging: Two classes are merged into one.
• Deleting: An entire class is removed.

After each operation, the algorithm re-evaluates the total message length. The change is only kept if it results in a shorter, more efficient explanation of the data.

2. The Inner Loop: Parameter Estimation (like E-M)

For any given set of classes, the algorithm must find the optimal parameters for them. This inner loop is very similar to the well-known Expectation-Maximization (E-M) algorithm:

• Assignment Step (like "E-Step"): For each data point, the algorithm calculates the probability that it belongs to each of the current classes. This is a "soft" assignment.
• Update Step (like "M-Step"): The parameters of each class (e.g., the mean and standard deviation for a Gaussian attribute, or the probabilities for a categorical attribute) are recalculated based on the weighted collection of data points assigned to them in the previous step.

This inner loop repeats until the class parameters and assignments stabilize.

The "Factor" in Snob Factor

This is not just a simple mixture model. Within each class, it can also model the covariance between attributes using a latent factor. This is a hidden, continuous variable that can influence multiple attributes simultaneously. By using a factor, the model can explain correlations between variables within a class, leading to a more powerful and compact model (i.e., a shorter message length) than if all attributes were assumed to be independent.