Feature engineering

Leon Bottou

COS 424 – 4/22/2010

Summary

Summary

I. The importance of featuresII. Feature relevance

III. Selecting featuresIV. Learning features

Leon Bottou 2/29 COS 424 – 4/22/2010

I. The importance of features

Leon Bottou 3/29 COS 424 – 4/22/2010

Simple linear models

People like simple linear models with convex loss functions

– Training has a unique solution.

– Easy to analyze and easy to debug.

Which basis functions Φ?

– Also called the features.

Many basis functions

– Poor testing performance.

Few basis functions

– Poor training performance, in general.

– Good training performance if we pick the right ones.

– The testing performance is then good as well.

Leon Bottou 4/29 COS 424 – 4/22/2010

Explainable models

Modelling for prediction

– Sometimes one builds a model for its predictions.

– The model is the operational system.

– Better prediction =⇒ $$$.

Modelling for explanations

– Sometimes one builds a model for interpreting its structure.

– The human acquires knowledge from the model.

– The human then design the operational system.

(we need humans because our modelling technology is insufficient.)

Selecting the important features

– More compact models are usually easier to interpret.

– A model optimized for explanability is not optimized for accuracy.

– Identification problem vs. emulation problem.

Leon Bottou 5/29 COS 424 – 4/22/2010

Feature explosion

Initial features

– The initial pick of feature is always an expression of prior knowledge.

images −→ pixels, contours, textures, etc.signal −→ samples, spectrograms, etc.

time series −→ ticks, trends, reversals, etc.biological data −→ dna, marker sequences, genes, etc.

text data −→ words, grammatical classes and relations, etc.

Combining features

– Combinations that linear system cannot represent:

polynomial combinations, logical conjunctions, decision trees.

– Total number of features then grows very quickly.

Solutions– Kernels (with caveats, see later)– Feature selection (but why should it work at all?)

Leon Bottou 6/29 COS 424 – 4/22/2010

II. Relevant features

Assume we know distribution p (X, Y ).

Y : outputX : input, all featuresXi : one feature

Ri = X \Xi : all features but Xi,

Leon Bottou 7/29 COS 424 – 4/22/2010

Probabilistic feature relevance

Strongly relevant feature

– Definition: Xi ⊥6⊥ Y |RiFeature Xi brings information that no other feature contains.

Weakly relevant feature

– Definition: Xi ⊥6⊥ Y | S for some strict subset S of Ri.

Feature Xi brings information that also exists in other features.

Feature Xi brings information in conjunction with other features.

Irrelevant feature

– Definition: neither strongly relevant nor weakly relevant.

Stronger than Xi ⊥⊥ Y . See the XOR example.

Relevant feature

– Definition: not irrelevant.

Leon Bottou 8/29 COS 424 – 4/22/2010

Interesting example

��

��

Two variables can be useless by themselves but informative together.

Leon Bottou 9/29 COS 424 – 4/22/2010

Interesting example

��

��

Correlated variables may be useless by themselves.

Leon Bottou 10/29 COS 424 – 4/22/2010

Interesting example

��

��

�������

Strongly relevant variables may be useless for classification.

Leon Bottou 11/29 COS 424 – 4/22/2010

Bad news

Forward selection

– Start with empty set of features S0 = ∅.– Incrementally add features Xt such that Xt ⊥6⊥ Y | St−1.

Will find all strongly relevant features.

May not find some weakly relevant features (e.g. xor).

Backward selection

– Start with full set of features S0 = X.

– Incrementally remove features Xi such that Xt ⊥⊥ Y | St−1 \Xt.Will keep all strongly relevant features.

May eliminate some weakly relevant features (e.g. redundant).

Finding all relevant features is NP-hard.

– Possible to construct a distribution that demands

an exhaustive search through all the subsets of features.

Leon Bottou 12/29 COS 424 – 4/22/2010

III. Selecting features

How to select relevant features

when p(x, y) is unknown

but data is available?

Leon Bottou 13/29 COS 424 – 4/22/2010

Selecting features from data

Training data is limited

– Restricting the number of features is a capactity control mechanism.

– We may want to use only a subset of the relevant features.

Notable approaches

– Feature selection using regularization.

– Feature selection using wrappers.

– Feature selection using greedy algorithms.

Leon Bottou 14/29 COS 424 – 4/22/2010

L0L0L0 structural risk minimization

������������������� ���

���������������

��

��

����

Algorithm

1. For r = 1 . . . d, find system fr ∈ Sr that minimize training error.

2. Evaluate fr on a validation set.

3. Pick f? = arg minrEvalid(fr)

Note

– The NP-hardness remains hidden in step (1).

Leon Bottou 15/29 COS 424 – 4/22/2010

L0L0L0 structural risk minimization

������������������� ���

���������������

��

��

����

Let Er = minf∈Sr

Etest(f ). The following result holds (Ng 1998):

Etest(f?) ≤ min

r=1...d

Er + O

√ hrntrain

+ O

√r log d

ntrain

+O

(√log d

nvalid

)

Assume Er is quite good for a low number of features r.Meaning that few features are relevant.

Then we can still find a good classifier if hr and log d are reasonable.We can filter an exponential number of irrelevant features.

Leon Bottou 16/29 COS 424 – 4/22/2010

L0L0L0 regularisation

minw

1

n

n∑i=1

`(y, fw(x)) + λ count{wj 6= 0}

This would be the same as L0-SRM.

But how can we optimize that?

Leon Bottou 17/29 COS 424 – 4/22/2010

L1L1L1 regularisation

The L1 norm is the first convex Lp norm.

minw

1

n

n∑i=1

`(y, fw(x)) + λ|w|1

Same logarithmic property

(Tsybakov 2006).

L1 regulatization can weed an

exponential number of irrelevant

features.

See also “compressed sensing”.

Leon Bottou 18/29 COS 424 – 4/22/2010

L2L2L2 regularisation

The L2 norm is the same as the maximum margin idea.

minw

1

n

n∑i=1

`(y, fw(x)) + λ‖w‖2

Logarithmic property is lost.

Rotationally invariant regularizer!

SVMs do not have magic properties

for filtering out irrelevant features.

They perform best when dealing

with lots of relevant features.

Leon Bottou 19/29 COS 424 – 4/22/2010

L1/2L1/2L1/2 regularization ?

minw

1

n

n∑i=1

`(y, fw(x)) + λ‖w‖12

This is non convex.

Therefore hard to optimize.

Initialize with L1 norm solution

then perform gradient steps.This is surely not optimal,but gives sparser solutionsthan L1 regularization !

Works better than L1 in practice.

But this is a secret!

Leon Bottou 20/29 COS 424 – 4/22/2010

Wrapper approaches

Wrappers

– Assume we have chosen a learning system and algorithm.

– Navigate feature subsets by adding/removing features.

– Evaluate on the validation set.

Backward selection wrapper

– Start with all features.

– Try removing each feature and measure validation set impact.

– Remove the feature that causes the least harm.

– Repeat.

Notes

– There are many variants (forward, backtracking, etc.)

– Risk of overfitting the validation set.

– Computationally expensive.

– Quite effective in practice.

Leon Bottou 21/29 COS 424 – 4/22/2010

Greedy methods

Algorithms that incorporate features one by one.

Decision trees

– Each decision can be seen as a feature.

– Pruning the decision tree prunes the features

Ensembles

– Ensembles of classifiers involving few features.

– Random forests.

– Boosting.

Leon Bottou 22/29 COS 424 – 4/22/2010

Greedy method example

The Viola-Jones face recognizer

Lots of very simple features.∑R∈Rects

αr∑

(i,j)∈Rx[i, j]

Quickly evaluated by first precomputing

Xi0 j0 =∑i≤i0

∑j≤j0

x[i, j]

Run AdaBoost with weak classifiers bases on these features.

Leon Bottou 23/29 COS 424 – 4/22/2010

IV. Feature learning

Leon Bottou 24/29 COS 424 – 4/22/2010

Feature learning in one slide

Suppose we have weight on a feature X.

Suppose we prefer a closely related feature X + ε.

����������������

���������������� ���

����������������������

���������������������

Leon Bottou 25/29 COS 424 – 4/22/2010

Feature learning and multilayer models

����������������

� ��������� �����

����������������

� ����������������

Leon Bottou 26/29 COS 424 – 4/22/2010

Feature learning for image analysis

2D Convolutional Neural Networks

– 1989: isolated handwritten digit recognition

– 1991: face recognition, sonar image analysis

– 1993: vehicle recognition

– 1994: zip code recognition

– 1996: check reading

INPUT 32x32

Convolutions SubsamplingConvolutions

C1: feature maps 6@28x28

Subsampling

S2: f. maps6@14x14

S4: f. maps 16@5x5

C5: layer120

C3: f. maps 16@10x10

F6: layer 84

Full connectionFull connection

Gaussian connections

OUTPUT 10

Leon Bottou 27/29 COS 424 – 4/22/2010

Feature learning for face recognition

Note: more powerful but slower than Viola-Jones

Leon Bottou 28/29 COS 424 – 4/22/2010

Feature learning revisited

Handcrafted features

– Result from knowledge acquired by the feature designer.

– This knowledge was acquired on multiple datasets

associated with related tasks.

Multilayer features

– Trained on a single dataset (e.g. CNNs).

– Requires lots of training data.

– Interesting training data is expensive

Multitask/multilayer features

– In the vicinity of an interesting task with costly labels

there are related tasks with abundant labels.

– Example: face recognition ↔ face comparison.

– More during the next lecture!

Leon Bottou 29/29 COS 424 – 4/22/2010