Data Mining (and machine learning)

David Corne, Heriot-Watt University - [email protected] slides and related resources: http://www.macs.hw.ac.uk/~dwcorne/Teaching/dmml.html

Data Mining(and machine learning)

Lecture 2: Some common data processing tasks


Today

• About classification

• A simple classification method (1-NN)

• Normalisation of instances and fields

• Discretization of fields

• Coursework B


Some notationA dataset has N records {R1,R2,…,RN},

… each of which has F fields {A1,A2,…,AF}.

Field AF is the target field – the one we want to predict.

v(i,j) is the value of field Aj in record Ri

So, for example, what is

N

i

F

jjiv

1

1

1),(

David Corne, , Heriot-Watt University - [email protected] slides and related resources: http://www.macs.hw.ac.uk/~dwcorne/Teaching/dmml.html

Some more terminologySometimes the values of a field are numeric. We call this a

numeric field (!)

Sometimes the values of a field are words or categories. E.g. {red, blue, green}

{true, false}

{male, female}

{low, medium, high}

this type of field is called a categorical field, or a nominal field.


So, what is this mostly about?

You have some data:

3 7 2 9 1 A

5 8 2 9 2 B

4 1 1 1 3 B

6 8 2 1 4 B

David Corne Heriot-Watt University - [email protected] slides and related resources: http://www.macs.hw.ac.uk/~dwcorne/Teaching/dmml.html


You have some data:

3 7 2 9 1 A

5 8 2 9 2 B

4 1 1 1 3 B

6 8 2 1 4 B

You get a new line of data, but it is not classified

3 6 1 8 1 ?



You have some data:

3 7 2 9 1 A

5 8 2 9 2 B

4 1 1 1 3 B

6 8 2 1 4 B

You get a new line of data, but it is not classified

3 6 1 8 1 ?

You need to predict the class value

David Corne, and Nick Taylor, Heriot-Watt University - [email protected] slides and related resources: http://www.macs.hw.ac.uk/~dwcorne/Teaching/dmml.html

Broadly how it’s doneIn order to be able to predict the class value of any new lineof data that comes along, this is what happens:


Broadly how it’s doneIn order to be able to predict the class value of any new lineof data that comes along, this is what happens:

Get your already-classified dataset

Run a machine learning method on these data, which produces a classifier

Use the classifier to make predictionsof the class value for new instances


E.g. here is some data we saw before

3 7 2 9 1 A

5 8 2 9 2 B

4 1 1 1 3 B

6 8 2 1 4 B

If we run a machine learning method like RIPPER, the classiferthat comes out is a rule (or a set of rules), e.g.:

If (A1<5) AND (A5<2) THEN class is AIf (A2>1) THEN class is B


E.g. here is some data we saw before

3 7 2 9 1 A

5 8 2 9 2 B

4 1 1 1 3 B

6 8 2 1 4 B

If instead we do machine learning with a Neural Network, or Naive Bayes, or Regression, etc …, then the output classifier is (more or less) a thresholded mathematical function of the attribute values. The type of function varies a lot between methods, but, it might be e.g.:

764.0)8.6()09.1()02.0()8.2( 4321 AAAA

With the convention that: if this is < 0 the class is A, otherwise B


Or the classifier could be a decision tree, such as:

A1>2 ?

A4<5 ? A4>7 ?

Class is A Class is B

yes

yes

no

no

Class is B


Applications

Just a reminder: you should already be developing an awarenessof this from previous lectures, from browsing data repositories, andfrom using your imagination.

Business, Commerce, Engineering, Medicine, and Science provide endless important applications for this process, in which:

• there are existing classified datasets, that we can learn from• and for which producing a classifier that can predict class values for new data instances is highly valuable.


An Important Issue

Overfitting: It is very possible to learn a classifier that gives accurate performance on the existing classified dataset, but which performs very poorly in its predictions on new, previously unseen data.

I will have more to say on overfitting, and how to combat it, in later lectures.

David Corne Heriot-Watt University - [email protected] slides and related resources: http://www.macs.hw.ac.uk/~dwcorne/Teaching/dmml.html

The main issue for today

Dataset preparation:

There are various things we can and should do to our dataset so that we make it easier, or indeed possible, to apply our chosen machine learning method.

E.g. we may want to apply the ID3 Decision Tree learning algorithm – but this only works with categorical data.

We may want to apply a Neural Network, but this needs the data to be all numeric, and for all values to be relatively small (e.g. between -1 and 1).

Etc..


Useful interlude: 1-NN classification

• Almost the simplest possible classification method.

• The `classifier’ is just the dataset itself

• Predict the class of a new instance new by finding the instance c in the dataset that is closest to new, and predicting that it will have the same class value as c


How to work out distance between two instances?

Depends on the data, and on your common sense. More on that later; but for now let’s say we are only dealing with numeric data. One way to work out distance is the SSD (sum-squared distance)

21

1)) ,() ,(( ibviav

F

i

If you have two records a and b, then the distance between themcan be given by:

Just add up the squared differences of the fields, omitting the class field.

Why don’t I take the square root of this?


Visualisation of 1-NN

Suppose you have data with two numeric fields and one class field,which is either red or green. We can treat the data as 2D points, coloured by the class. Suppose this is the pre-classified dataset:




A new unclassified instance comes along – what is the class?




The way you answer this visually is very similar to the way 1-NN works


Life is difficult / interesting

Just an aside to point out that, in many interesting and importantclassification problems, the situation is more like this:


The performance of 1-NN classification

Calculate it for any dataset of R records like this:Initialise accuracy to 0. For each record r Find the closest other record, o, in the dataset If the class value of o is the same as the class value of r, than add 1 to accuracyReturn 100 * accuracy/R

Thereby, you obtain a percentage accuracy result


Scaling, Normalisation and Discretization

So we will learn these things:Ways to do normalisation and scaling (which may

be necessary or sensible, depending on how the data were generated, and on what DM/ML methods you want to use)

Ways to do discretization (for turning numeric fields into categorical fields)


NormalisationYou have data about word-counts for specific words on web pages. You want to

be able to predict whether the page is about sport or business.

“winner” “game” “team” “sales” “run” Page content

16 22 81 75 10 business

12 14 44 16 12 business

4 7 20 0 2 sport

2 3 7 6 1 ???

What category would 1-NN predict for the 4th record?

David Corne, r, Heriot-Watt University - [email protected] slides and related resources: http://www.macs.hw.ac.uk/~dwcorne/Teaching/dmml.html

NormalisationThe closest record to the 4th is the 3rd, so it would be

predicted to be sport. But this is probably wrong. Why?


16 22 81 75 10 business

12 14 44 16 12 business

4 7 20 0 2 sport

2 3 7 6 1 ???


NormalisationThe closest record to the 4th is the 3rd, so it would be

predicted to be sport. But this is probably wrong. Why?


16 22 81 75 10 business

12 14 44 16 12 business

4 7 20 0 2 sport

2 3 7 6 1 ???

How could you pre-process the data to help ML methods?


Record-wise Normalisation

In these data, the sensible thing to do is transform each record so that it is normalised by the total number of words. This makes them more comparable, each providing a “fingerprint” in terms of the relative proportions of the probe words.

Sometimes this is necessary, sometimes it is useful – it just takes common sense.


Record-wise normalised version


0.0784 0.1078 0.397 0.368 0.049 business

0.12 0.14 0.45 0.16 0.12 business

0.12 0.21 0.61 0 0.061 sport

0.105 0.158 0.368 0.316 0.053 ???

1-NN now shows record 4 closer to 1 and 2

Each row is simply scaled so that he numeric fields add up to 1


Record-wise Normalisation

Using the notation in slide 3 – given a dataset of N numeric fields, how would you represent record-wise normalization of record r ?


Min-Max NormalisationThis is columnwise normalisation, done separately for each field.

For each numeric attribute, scale it so that each value is in a specific range

[a, b]. E.g. usually [0,1], sometimes [-1, 1], etc. E.g. if we use [0,1], we do

this for each record r.

)(range

)min(),( becomes ),(

i

iirvirv

Where min(i) is the smallest value for field i in the dataset,max(i) is the largest, and range(i) is max(i) – min(i).


When min-max normalisation might be useful

Height (mm)

Weight (kg)

IQ Mark in DMML module

1856 75 95 70

1502 56 101 82

1904 86 112 90

1775 61 110 90

1901 81 92 ???


Height (mm)

Weight (kg)


1856 75 95 70

1502 56 101 82

1904 86 112 90

1775 61 110 90

1901 81 92 ???

What is the closest record to record 5?


Height (mm)

Weight (kg)


1856 75 95 70

1502 56 101 82

1904 86 112 90

1775 61 110 90

1901 81 92 ???

What is the closest record to record 5?Will this be a good predictor for DMML Mark?


Height (mm)

Weight (kg)


1856 75 95 70

1502 56 101 82

1904 86 112 90

1775 61 110 90

1901 81 92 ???

What is the closest record to record 5?Will this be a good predictor for DMML Mark?Which field will be most important for predicting Mark?


Height (mm)

Weight (kg)


1856 75 95 70

1502 56 101 82

1904 86 112 90

1775 61 110 90

1901 81 92 ???

What is the closest record to record 5?Will this be a good predictor for DMML Mark?Which field will be most important for predicting Mark?How does min-max normalisation help?


Discretization

Height (mm) Weight (kg) IQ Mark in DMML

module

1856 75 95 70

1502 56 101 82

1904 86 112 90

1775 61 110 90

1901 81 92 65

We can’t run the ID3 decision tree algorithm on this.


Discretization

Height (mm) Weight (kg) IQ Mark in DMML

module

tall heavy medium A

short light medium A+

tall heavy smart A+

medium medium smart A+

tall heavy medium B

We can run the ID3 decision tree algorithm on this.


Discretization

Discretization is simply a process that converts a numerical field into a class field.

How might you do this?


Equal Width Binning (EWB)

EWB(5), for example, means you discretize into 5 values, where each value has equal `width’.

If field f values range from 0 to 100, then, then each bin has width 20. In the converted dataset, we can just label this bins 1,2,…,5, or we can use appropriate linguistic terms, such as:

{very low, low, medium, high, very high}


Equal Frequency Binning (EFB)

EFB(3), for example, means you discretize into 3 values, where each bin covers around the same amount of data.


Contrived Example for EWB(2)

Height (mm)

Weight (kg)

IQ DMML mark

1856 75 95 70

1502 74 101 82

1904 77 112 90

1775 61 110 90

1901 76 92 70

Height (mm)

Weight (kg)

IQ DMML mark

high high low low

low high low high

high high high high

high low high high

high high low low


Other discretization methods

There are others, which almost all differ from EWB and EFB in one respect: they use class value information in attempt to find bins that make good sense with regard to classification.

We may look at those later


Finally: Missing values

Often, a dataset has missing values for some fields in some records. E.g. a certain blood test was not taken for some patients. A questionnaire response for some question was unreadable, etc…


Coursework B

There are several ways to deal with missing values. Find out yourself

Do some research with google, or google scholar (or your favourite other search engine).

Write a report of only ~100 words, entitled “Methods for dealing with missing values in datasets”

Provide at least 2 relevant references (to papers or URLs) at the end of your report.

The whole thing should only take 30mins—1hr


Next week

Basic statistics

Histograms

A little more normalisation and discretization

Coursework 1


Extra slides on Data Quality / Data Cleaning

I am not teaching Data Quality / Data Cleaning as part of F21DL this year. But here are some slides you may be interested in anyway


On Data Quality

Suppose you have a database sitting in front of you, and I ask `Ìs it a good quality database?’’

What is your answer? What does quality depend on?

Note: this is about the data themselves, not the system in use to access it.


A Conventional Definition of Data Quality

Good quality data are:

Accurate, Complete, Unique,

Up-to-date, and Consistent ;

meaning …


A Conventional Definition of Data Quality, continued …

Accurate: This refers to how the data were recorded in the first place. What might be the inaccurately recorded datum in the following table?

Barratt John 22 Maths BSc Male

Burns Robert 24 CS BSc Male

Carter Laura 20 Physics MSc Female

Davies Michael 12 CS BSc Male



Complete: This refers to whether or not the database really contains everything it is supposed to contain.

E.g. a patient’s medical records should contain references to all medication prescribed to date for that patient.

The BBC TV Licensing DB should contain an entry for every address in the country. Does it?



Unique: Every separate datum appears only once. How many `Data Quality errors’ can you find in the following table, and what types are they?

Surname Firstname DoB Driving test passed:

Smith J. 17/12/85 17/12/05

Smith Jack 17/12/85 17/12/2005

Smith Jock 17/12/95 17/12/2005



Up-to-date: The data are kept up to date.

The post office recently changed my postcode from EX6 8RA to EX6 8NU. Why does this make it difficult for me to get a sensible quote for home insurance or car insurance?

Can you think of a DB where it doesn’t matter whether or not the data are kept up to date??



Consistent: The data contains no logical errors or impossibilities. It makes sense in and of itself.

Why is the following mini DB inconsistent?

Date Sales Returns Net income

23rd Nov £25,609 £1,003 £24,506

24th Nov £26,202 £1,601 £24,601

25th Nov £28,936 £1,178 £25,758


Note: This definition of data quality is not much use, since there is no way to measure DQ sensibly Completeness: How will we know?? Uniqueness: It is hard to tell whether two entries are

similar, or duplicates! Up-to-date-ness: How do we know? Consistent: consistency errors can be very hard to find,

especially in a very large DB

The database research `world’ is actively engaged in finding waysto measure data quality sensibly. In the meantime, we just use common sense to avoid dirty data at all points of the DQ continuum


The Data Quality ContinuumIt’s rare that a datum is entered once into a DB and

then left alone. Usually, a datum has a long and varied life, into which errors can arise at each and every stage. The continuum is:

– Data gathering

– Data delivery

– Data storage

– Data integration

– Data retrieval

– Data analysis

So, if we want to monitor DQ, we need to monitor it at each of these stages


DQ Continuum: Example

This is an example I am familiar with, helping to illustrate the DQ continuum.

The International Seismological Centre (ISC) is in Thatcham, in Berkshire. It’s a charity funded by various governments. Their role is to be the repository for recording all earthquake events on the planet.


DQ Continuum: ISC example: gathering

ISCData gathering centres


See 2006 earthquake data via my dmml page


DQ Continuum: ISC example: delivery

Raw seismograph data from local collection points to DG centres.

Email or ftp to ISC; some centres provide raw data, some provide interpreted data (e.g. maybe won’t send some data if they believe it in error in the first place)


DQ Continuum: ISC example: integration

The ISC’s role is actually to figure out where and when the Earth tremors were (there are hundreds per month) based on reverse engineering from seismograph readings. They integrate the raw data and attempt to do this, largely by hand and brain, and record their findings in archival CDs


DQ Continuum: ISC example: retrieval/analysis

You can get a CD from ISC anytime, for the earth tremor activity on any particular day.

I’m not sure whether you can get the raw data from them.

Naturally, you can analyse the data and see if you can find inconsistencies or errors.


The ISC DQ Continuum

Where might there occur errors, of: Accuracy? Completeness? Uniqueness? Timeliness? Consisency?

What else is important in this case?


Where DQ problems occur (gathering)

• Manual data entry (how can we improve this?)

• Lack of uniform standards for format and content.

• Duplicates arising from parallel entry• Approximations, alternatives, entries altered

in order to cope with s/w and/or h/w constraints.

• Measurement errors.


Where DQ problems occur (delivery)

• Multiple hops from source to DB – problems can happen anywhere

• Inappropriate pre-processing (e.g. removing some `small’ seismograph readings before sending on to ISC; rounding up or down, when the destination needs more accurate data).

• Transmission problems: buffer overflows, checks (did all files arrive, and all correctly?)


Where DQ problems occur (storage)

• Poor, out of date or inappropriate metadata

• Missing timestamps

• conversion to storage format (e.g. to excel files, to higher/lower precision


Where DQ problems occur (integration)

This is the business of combining datasets – e.g. from different parts of a company, from (previously) different companies following an acquisition; from different government agencies, etc.

• Different keys, different fields, different formats• Different definitions (`customer’, ìncome’, …)• Sociological factors: reluctance to share!


Where DQ problems occur (retrieval/analysis)

The problem here is usually the quality of DBs that store the retrieved data, or the use of the retrieved data in general. Problems arise because:

• The source DB is not properly understood!• Straightforward mistakes in the queries that retrieve the

relevant data. E.g. A database of genes contains entries that indicate

whether or not each gene has a known or suspected link with cancer. A retrieval/analysis task leads to publishing a list of genes that are not relevant to cancer. What is the problem here?


What Keeps DBs Dirty

A good DBMS will have built in tools for:Consistency in data typesConsistency in field valuesConstraints and checks that deal with Null values,

Outliers, Duplication.Automatic timestampsPowerful query language (makes retrieval logic errors

less likely)… so, why are you refused a loan, have mail delivered to the wrong address, and

get charged too much for your mobile calls?


… all this:• Consistency constraints are often not applied, or are

applied!– suppose height is not allowed to go over 2 metres in a

school student DB– My postcode problem

• The data are just too numerous, complex and ill-understood. `Cleaning it’ would cost too much!

• Undetectable problems: incorrect values, missing entries

• Metadata not maintained properly


Single Source vs Multiple Source Schema Level v Instance Level

One useful way to categorize problems, independent of how we did so in the last lecture, is according to whether the problems are the sort we can get if we have just one source of data, or whether the problem arises directly from trying to combine data from multiple sources

Problems can also be schema level or instance level


Single Source / Schema level examples

Scope Problem Unclean Notes

attribute Illegal values DoB=30.02.88 Values out of range

record Violated attribute dependencies

Car-owner = No, make = Toyota

Make should clearly have a Null value here.

Record type

Uniqueness violations

Name= Jo Smith, NUS no. = 3067

Name= Ed Brown, NUS no. = 2124NUS no.s should be unique

Source Referential integrity violation

Name= D Corne, Office = EM G.92 Where is G. 92 ?


Single Source / Instance level examples

Scope Problem Unclean Notes

attribute missing values,

mis-spellings,

abbreviations,

Misfields,

Embedded vals

Top speed = 0 mph,Title = Dark Side of the Moan

FullName = J. Smith

Colour = 160mphPhone = “Dan Jones 0207 308653”

Dummy entries – values unavailable at entry time, human error

record Violated attribute dependencies

City = Edinburgh. Postcode = EX6

Record type

Word transposition,

Duplicates, contradictions

Name= Jo Smith, Name = Carr, Jim

Name= J. Smith, Name = Joe Smith

Name = Jo Smith, DoB = 17/12/62

Name = Jo Smith, DoB = 17/11/62

Source Wrong references Name= D Corne, Office = EM G.46 EM G. 46 exists, but is not my office.


Multiple Source Problems/ Instance and Schema level examples

The Smiths buy books and music online from company A:

Customer ID Name Street City Sex

102 Luke Smith 5 Chewie Rd Dawlish, Devon 0

175 Leia Smith Chewie St, 5 Dawlish 1

They also buy books and music online from company B:

Client ID LastName Other names Phone Gender

23 Smith Luke Michael 378988 Male

35 Smith Leia S. +44(0)1626 378988 F


When Companies A and B merge, various problems arise when they merge their DBs

Combining customer fields and client fields – are they really thesame things?

How to ensure that Company A’s customer 37 and Company B’s client 37 get separate entries in the new DB.

Are Luke Smith and Luke Michael Smith the same person??

Do Luke and Leia live at the same address? Etc …

A forced `fast resolution’ to these problems will usually lead toerrors in the new ìntegrated’ DB


A Special but Common type of Problem: Semantic Complexity

Semantic Complexity (SC) is the state of play where different users of a DB have different conceptions of what the data represent.

E.g. Local Police DB keep record of all crimes in an area, where the key is the victim’s name. When someone who was a victim moves to a different area, they remove all records relating to that person.The local council use this DB to produce a report of the total amount of crime every month. Why does it give figures that are too low?


Semantic Complexity: Missing/Default Values

One source of semantic complexity is the different meanings that missing values can have.

E.g. Suppose the histogram of value types in mobile phone no. field is: Number of entries in `mobile phone number' of particular types

0

200

400

600

800

1000

1200

validnumber

NULL All 0s All 9s text anddigits


What does NULL mean?A. This record is of someone who does not have a mobile phone?B. This record is of someone who has a mobile phone, but chose not to supply

the number?C. This record is of someone who has a mobile phone, but who forgot to supply

the number, or it was hard to decipher and recorded as NULL?Maybe some are of type A and some are of type B and some are of type C. For

some applications/analyses, we may wish to know the breakdown into types. What about the All zero and All nine entries? Precisely the same can be

said of them. Or, perhaps the protocols for recording the entries indicated NULL for type A, 0000000 for type B and 9999999 for type C.

The above relate to a quite simple form of semantic complexity – but what if someone uses this DB to estimate the proportion of people who have never had a mobile phone?


Cleaning via basic data analysis

Data Profiling: examine the instances to see how the attributes vary. E.g. Automatically generate a histogram of values for that attribute.

Number of cars sold with this colour

0

100

200

300

400

500

600

black blue bule orange red green white yellow

How does thehistogram help us in finding problemsin this case?


What problems does this analysis alert us to?

Histogram of Online Supermarket customers by age field

0

100

200

300

400

500

600

<0 0-10 11-20 21-30 31-40 41-50 51-60 61-70 71-80 81-90 >90


Acknowledgements

I adapted the data quality/cleaning material from various sources, most notably:

• A ppt presentation called `Data Quality and Data Cleaning: An Overview’ by Tamrapani Dasu and Theodore Johnson, at AT & T Labs

• A paper called `Data Cleaning: Problems and Current Approaches’, by Erhard Rahm and Hong Hai Do, University of Leipzig, Germany.

My thanks to these researchers for making their materials freely available online.

Documents

Data Mining (and machine learning)