DataCleaner offers a set of analyzers with the purpose of training or testing Machine Learning models. The idea is that records can be used to build ("train") a mathematical model that describes the value of any given attribute/column based on the other attributes of the record. For example, you could classify or predict house prices based on historical data given price and relevant factors for the price such as size, location, condition and time of year.
There are two use-cases that are supported:
Classification, which is the act of determining a class of record. For example, you may want to classify what product most likely fits a particular customer based on the attributes of the customer.
Regression, which is determines a continuous value on a numeric scale. For example, you might want try to predict the price of a house based on the traits of the house.
DataCleaner has built in support for the following types of Machine Learning models:
Random Forests
Support Vector Machines
Neural Networks
When training a model, you provide data that represents truthful observations. These records have to include values for the attribute that you will be trying to classify or predict using regression.
Attributes that affect the prediction or classification are called features. Feature has to be numeric as to provide meaningful mathematical input to the model. So sometimes a feature has to be extracted from a raw value instead of just applied as-is. For example, in textual analysis where you're trying to determine the language or nature of a piece of text, you will typically want to extract n-grams from the text. DataCleaner offers the following feature extraction strategies (which you select when adding a column as input to the training component):
Direct (0.0 to 1.0), Takes numerical values as-is.
Scaled (Min-Max), Scales numerical values from the minimum to the maximum value observed.
Vector (One-Hot Encoding), Generates a feature for every distinct value encountered. The values of the feature will be either 0 or 1 to indicate whether or not the record has that particular value.
Vector (2-gram), Generates a feature for every 2-gram observed in the text.
Vector (3-gram), Generates a feature for every 3-gram observed in the text.
Vector (4-gram), Generates a feature for every 4-gram observed in the text.
Vector (5-gram), Generates a feature for every 5-gram observed in the text.