OnlineStats is an online machine-learning library for creating robust statistical models of streaming data. OnlineStats utilizes continuous learning to adapt in real-time. Unlike alternative methods, the statistical models produced by OnlineStats do not suffer from the effects of overlearning, because the learning algorithm uses self-correction to withstand the obstacles faced in an online environment.

An estimator is used to approximate the average of a given variable in real-time. At each step, an estimator calculates an updated estimate value given its previous estimate and a current observation. Given a sufficient set of parameters, an estimator will settle on an estimate that describes its previous observations sufficiently well.

Estimator Examples

A model uses estimators to approximate a correlation matrix given a set of variables. At each step, a model updates a set of estimators that together predict the correlation coefficient for each pair of variables. There are two types of models, namely correlators and cross-correlators. A cross-correlator does the same as a correlator, that is produces a correlation matrix, however the correlation coefficients are calculated at each time step between a set of variables at the current time and the same set of variables at some previous time delayed by a constant factor.

A statistical model is used to approximate the correlation coefficients between observable variables. At each step, a model calculates a revised correlation matrix by updating a set of estimators associated with each individual variable as well as each variable pair. Given a sufficient set of estimators, a model will converge on a correlation matrix that describes relationships between variables sufficiently well.

Model Examples