Updated A.Y. 2021-2022
The course provides an introduction to Statistical Learning and Data Mining.
The advances in information technology have made available very rich information data sets, often generated automatically as a by-product of the main institutional activity of a firm or business unit. Most organizations today produce an electronic record of essentially every transaction in which they are involved. Firms collect terabytes data over operating periods (transactions data, e.g. credit cards). Most often these data are collected as secondary data, with no specific sampling design or research question on top.
Data Mining deals with inferring and validating patterns, structures and relationships in data, as a tool to support decisions in the business environment.
The course offers an insight into the main statistical methodologies for the visualisation and the analysis of business and market data, providing the information requirements for specific tasks such as credit scoring, prediction and classification, market segmentation and product positioning. Emphasis will be given to empirical applications using modern software tools (Rstudio, Matlab).
The course has the following intended learning outcomes:
- - to provide a thorough knowledge of data mining methods and statistical learning techniques;
- - to provide the expertise to manage complexity in information and to be able to distill the stylized facts that are relevant for interpretation;
- - to be able to predict business outcomes;
- - to be able to select a predictive method among those available;
- - to be able to communicate the statistical findings to a non expert audience;
- - to be able to perform sophisticated statistical analyses with the appropriate software.
- - to critically appraise the potential and the limitation of the available methodologies.
Introduction to data mining. Tools for data analysis, visualisation and description.The linear regression model.
Model selection and evaluation: bias-variance trade-off, model complexity and goodness of t. Cross-validation. Selection using information criteria.
Regularization and shrinkage methods: rigde regression, lasso, forward stagewise regression. Principal components regression.
Linear methods for classication: Bayes Classication Rule. Discriminant analysis. Canonical variates.Logistic regression.
Semiparametric regression: Regression splines and smoothing splines.
Kernel smoothing methods: Local polynomial regression.
Density estimation. Nearest neighbor classication.
Additive Models, tree-based methods. GAM, Regression andclassication trees. Boosting.
Knowledge and Understanding:
The course covers the modern statistical methodologies for the visualisation and the analysis of business and market data, that are relevant for making decisions in a complex and rapidly changing business environment.
The fundamental theme is supervised statistical learning, which deals with the prediction of quantitative and qualitative outcomes using a potentially large set of inputs. The two problems, regression and classification, constitute the core of the course.
Emphasis is given to the problem of variable and model selection and on the generalizability of a prediction method outside the training sample, via the optimization of the trade-off between model complexity and the in-sample goodness of fit.
G James, D Witten, T Hastie, and R Tibshirani and J Friedman. An Introduction to Statistical Learning with Applications in R. Springer, Springer Series in Statistics, 2009.
Dowloadable at http://www-bcf.usc.edu/~gareth/ISL/