Slide 1

Can We Generate More Value from Data?

Learning from Data

Probabilistic Approach

Learning from Data

From Data to Probability

Probabilistic Data Mining

What Makes Up ‘Problem Dimensionality’?

Addressing Dimensionality

Macroscopic Prediction

Macroscopic Prediction

Boltzmann’s Solution (1877)

Boltzmann’s Solution (cont.)

Boltzmann’s Solution (cont.)

Why Does It Work?

Why Does It  Work? (cont.)

General Maximum Entropy

General Maximum Entropy (cont.)

Addressing Dimensionality

Parametric Approximation

Probability Approximation

Maximum Likelihood

Maximum Likelihood (cont.)

Addressing Dimensionality

Information Geometry

Dual Projections

Pythagorean Geometry

Dual Geometry

Bayesian Estimation

Addressing Dimensionality

Relevance-Based Weighting

What If the Model Is Too Complex?

Relevance-Based Weighting of Data

Local Empirical Distributions

Local Modeling

Multiple Forecasting Applications

Data-Centric Technology

Increasingly Popular Approach

How Do Humans Solve Problems?

Corresponding Technologies

Pros and Cons

Addressing Dimensionality

No Locality in High Dimension?

Limits of Local Modeling

No “Local” Data in High Dimensions

Local Modeling Revisited

Local Modeling Revisited

Cube Encoding

Cube Encoding

Cube Encoding

Cube Encoding

General “Linear” Case

Symbolic Forecasting

Symbolic Forecasting

Decision-Making Process

Lessons Learnt

Hypothesis Formulation …

Feature Selection …

Training Data Selection …

Decision Support Rather Than Automation

Humans To Stay in Control