supervised learning, categorization and signal-detection ... · categorization and signal-detection...

25
Supervised learning, categorization and signal- detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning Decision trees Signal detection theory Supervised learning, categorization and signal-detection theory Pantelis P. Analytis February 28, 2018 1 / 25

Upload: others

Post on 05-Oct-2020

8 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Supervised learning, categorization andsignal-detection theory

Pantelis P. Analytis

February 28, 2018

1 / 25

Page 2: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

1 Introduction

2 The linear world

3 Instance-based learning

4 Decision trees

5 Signal detection theory

2 / 25

Page 3: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

What is supervised learning?

Classification—categorization problems

Regression—estimation problems

3 / 25

Page 4: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

The lens model: a generic framework

developed by Egon Brunswik in the 30s

Has been adapted and interpreted in different ways bypsychologists.

4 / 25

Page 5: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

The linear lens

Hammond (1955), Todd (1954)

Linear model: y = x0 + x1 · b1 + x2 · b2.... + xn · bn

5 / 25

Page 6: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Actuarial vs. clinical judgement

First comparison of human and machine judgement byPaul Meehl in 1954.

Comparison on the same prediction problem, but possiblyusing different past training data. Use of cross-validationto evaluate the models.

Linear models or even simpler algorithms consistentlyoutperform humans.

6 / 25

Page 7: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Paramorphic representation of judgement

Hoffman (1960) used the linear model to develop the firstmethod of preference learning, precursor to conjointanalysis.

7 / 25

Page 8: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Paramorphic representation of judgement

Hoffman (1960) used the linear model to develop the firstmethod of preference learning, precursor to conjointanalysis.

8 / 25

Page 9: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Logistic regression: simple but robust

b0 shifts the position of the curve and b1 (or b1 + ...bn)defines its slope.

Logistic regression can deal with any number of variables

Computationally very cheap. It is used heavily in theindustry although there are many better algorithms.

9 / 25

Page 10: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Instance-based learning—exemplars and prototypes

First appearance in Statistics: Fix and Hodges (1951) andCover and Hart (1967).First appearance of prototypes in Psychology in 1954 in apaper by Fred Attneave. 10 / 25

Page 11: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

K-nearest neighbors

Define a similarity vector:√∑k

i=1 (xi − yi )2

Pros and cons: intuitive and easy to implement butcomputationally very expensive.

Variations of the basic approach: assign weightsdepending on distance, discard some instances.

11 / 25

Page 12: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

K-nearest neighbors

Define a similarity vector:√∑k

i=1 (xi − yi )2

Pros and cons: intuitive and easy to implement butcomputationally very expensive.

Variations of the basic approach: assign weightsdepending on distance, discard some instances.

12 / 25

Page 13: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Shepard’s Law

Humans and other animals generalize on the basis offeature distance from previously seen items.

Shepard suggested that if there is any law in psychology,that should be it (Science, 1987).

13 / 25

Page 14: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Discussion: do we have access to our ownalgorithms?

Most of the decisions we make are intuitive. What doesexactly intuition mean?

Hogarth (2001) has dedicated a book on the topic.

14 / 25

Page 15: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Medical decision trees

Super (1984), Super triage and rapid treatment (STARTtree).

15 / 25

Page 16: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Bailing out decisions

A model of how British judges decide whether to make apunitive bail (Dahmi, 2003).

16 / 25

Page 17: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

First decision trees

Morgan and Sonquist, 1963, formulated the firstregression tree in Statistics.

Hunt et al. 1966, concept learning system.

17 / 25

Page 18: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

CART, ID3 and C.4.5 algorithm

Friedman (1977), Breiman et al. (1984).

Quinlan (1979,1983,1986,1993).

18 / 25

Page 19: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Shannon’s entropy and information gain

Entropy =∑n

i=1−pi · log2 pi

G (S ,A) = Entropy(S) −∑

v∈Values(A)|Sv ||S | · Entropy(Sv )

19 / 25

Page 20: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Pros and cons of decision trees

Pros: Comprehensible and easy to explain, inexpensiveonce trained, work with continuous and categoricalvariables, can capture non-linearities.

Cons: Weaker in estimation tasks, prone to overfittingespecially with small samples (pruning can somewhatcounteract that), can be costly to train.

20 / 25

Page 21: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Random forests to the rescue

Trained on different samples of data and then aggregatedwith majoritarian voting.

Random forests curb the variance of the algorithm,reducing the overall error.

21 / 25

Page 22: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Signal detection theory

Origins in Neyman-Pearson hypothesis testing, firstapplications in the use of radars.

Introduced in psychology to study perception andsensation (1954).

22 / 25

Page 23: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory Criterion: the internal response level beyond which a

decision-maker responds yes (radar operator, doctor orsearch engine)

d’ = separation/spread

23 / 25

Page 24: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Broadly used paradigm in psychophysics.

Also widely employed to define measure of performance ininformation retrieval and recommender systems.

24 / 25

Page 25: Supervised learning, categorization and signal-detection ... · categorization and signal-detection theory Pantelis P. Analytis Introduction The linear world Instance-based learning

Supervisedlearning,

categorizationand signal-detectiontheory

Pantelis P.Analytis

Introduction

The linearworld

Instance-basedlearning

Decision trees

Signaldetectiontheory

Signal detection theory

Luan et al. (2011) draw connections between the fast andfrugal tree framework and SDT.

25 / 25