Machine Learning: 2 Books in 1: Machine Learning for Beginners, Machine Learning Mathematics. An Introduction Guide to Understand Data Science Through the Business Application

 
Assessing model accuracy
There is no-one-size-fits-all or jack-of-all-trades technique in the field
of statistics, it is impossible for a single technique to dominate across the
vast variety of data sets. The most frequently used metric in the
“regression" environment is the "mean squared error (MSE)", which can be
used for quantification of the extent to which the “predicted answer value”
is near to the true answer value for the target observation. For the predicted
responses that are extremely close to the true responses, the MSE is
computed as small but if the predicted and true responses vary considerably
concerning some of the observations then the MSE is computed as large.
For example, assume we have clinical data for some patients such as their
weight, blood pressure, gender, age, history of family illnesses along with
information stating if they are diabetic or not. This patient-related data set
can be used to train a statistical technique for predicting diabetes risk based
on clinical measures.

---

The most frequently used metric in the "classification" environment is the
"confusion matrix". The core characteristic of statistical learning is that
with continuous learning the model becomes more flexible and the
training errors are reduced, although the test error may not be decreased.
Bias and Variance
In the context of machine learning, bias is defined “the
simplified assumptions made by a model to further simplify the learning
of the target task”. Parametric models are designed with an inherent strong
bias, which makes learning much faster and simpler but significantly
reduces the overall flexibility of the model for a wider variety of
application. “Decision trees", "k-nearest neighbors", and "support
vector machines" are categorized as low-bias algorithms for machine
learning models. The “high-bias” ML algorithms are "linear regression",
"linear discriminant analysis", and "logistic regression".
In the context of machine learning, a variance can be defined “as the
amount by which the estimation of the target function will be altered with
the use of a different training data set”. Non-parametric models" with
high flexibility tend to have a high variance score. “Linear regression",
"linear discriminant analysis", and "logistic regression" are low-variance
ML algorithms. “Decision Trees", "k-Nearest neighbors", and "Support
Vector Machines” are high variance ML algorithms.

---