Final Project-

1
Final Project- Mining Mushroom World
2
Agenda

• Motivation and Background
• Determine the Data Set (2)
• 10 DM Methodology steps (19)
• Conclusion

3
Motivation and Background

• To distinguish between edible mushrooms and
  poisonous ones by how they look
• To know whether we can eat the mushroom, to
  survive in the wild
• To survive outside the computer world

4
Determine the Data Set (1/2)

• Source of dataUCI Machine Learning Repository
• Mushrooms Database
• From Audobon Society Field Guide
• Documentationcomplete, but missing statistical
  
• information
• Described in terms of physical characteristics
• Classificationpoisonous or edible
• All attributes are nominal-valued
• Large database 8124 instances (2480 missing
  values for attribute
• 12)

5
Determine the Data Set (2/2)

• 1. Past Usage
• Schlimmer,J.S. (1987). Concept Acquisition
  Through Representational Adjustment (Technical
  Report 87-19).
• Iba,W., Wogulis,J., Langley,P. (1988). ICML,
  73-79
• 2. No other mushrooms data

6
10 DM Methodology steps

• Step 1. Translate the Business Problem
• into a Data Mining Problem
• Data Mining Goalseparate edible mushrooms from
  poisonous ones
• How will the Results be Used- increase the
  survival rate
• How will the Results be Delivered- Decision Tree,
  Naïve Bayes, Ripper, NeuralNet

7
10 DM Methodology steps

• Step 2. Select Appropriate Data
• Data Source
• The Audubon Society Field guide to North American
  Mushrooms (1981). G. H. Lincoff (Pres.), New
  York Alfred A. Knopf
• Jeff Schlimmer donated these data on April 27th,
  1987
• Volumes of Data
• Total 8124 instances
• 4208(51.8) edible 3916(48.2) poisonous
• 2480(30.5) missing in attribute stalk-root

8
10 DM Methodology steps

• Step 2. Select Appropriate Data
• How Many Variables- 22 attributes
• cap-shape, cap-color, odor, population, habitat
  and so on
• How Much History is Required- no seasonality
• As long as we can eat them when we see them

9
10 DM Methodology steps

• Step 3. Get to Know the Data
• Examine DistributionsUse Weka to visualize all
  the 22 attributes with histograms
• Classediblee, poisonousp

10
Step 3. Get to Know the Data

1. Examine Distributions there are 2 types of
   historgrams
2. First- all kinds of values appear
3. (Attribute 21) population abundanta,
   clusteredc, numerousn, scattereds, severalv,
   solitaryy

11
Step 3. Get to Know the Data

• 1. Examine Distributionsthere are 2 types of
  historgrams
• Second- only some kinds of value appear
• (Attribute 7) gill-spacingclosec, crowdedw,
  distantd

12
Step 3. Get to Know the Data

• 1. Examine Distributionsthere are exceptions
• Exception 1- missing values in the attribute
• (Attribute 11) stalk-rootbulbousb, clubc,
  cupu, equale, rhizomorphsz, rootedr,
  missing?
• 2480 of this attribute have
  missing values (Total 8124)

13
Step 3. Get to Know the Data

• 1. Examine Distributionsthere are exceptions
• Exception 2- undistinguishable attribute
• (Attribute 16) veil-typepartialp, universalu

14
Step3. Get to Know the Data

• 2. Compare Values with Descriptions
• no unexpected values except for missing values

15
10 DM Methodology steps

• Step 4. Create a Model Set
• Creating a Balanced Sample- 75(6093) as training
  data, 25(2031) as test data
• Rapid Miners cross-validation function k-1 as
  training, 1 as test

16
10 DM Methodology steps

• Step 5. Fix Problems with the Data
• Dealing with Missing Values- the attribute
  stalk-root has 2480 missing values
• replace all missing values with the average of
  stalk-root value
• We replaced ? with the average value b

17
10 DM Methodology steps

• Step 6. Transform Data to Bring Information
• to the Surface
• all nominal attribute, no numerical analysis in
  this step

18
10 DM Methodology steps

• Step 7. Build Model
• 1. Decision Tree
• Performance
• Accuracy99.11
• Lift189.81

True p True e Class precision
Pred. p 961 0 100
Pred. e 18 1052 98.32
Class recall 98.16 100.00
True p True e Class precision
Pred. p 961 0 100
Pred. e 18 1052 98.32
Class recall 98.16 100.00
19
10 DM Methodology steps

• Step 7. Build Model
• 2. Naïve Bayes
• Performance
• Accuracy95.77
• Lift179.79

True p True e Class precision
Pred. p 902 9 99.01
Pred. e 77 1043 93.12
Class recall 92.13 99.14
True p True e Class precision
Pred. p 902 9 99.01
Pred. e 77 1043 93.12
Class recall 92.13 99.14
20
10 DM Methodology steps

• Step 7. Build Model
• 3. Ripper
• Performance
• Accuracy100
• Lift193.06

True p True e Class precision
Pred. p 979 0 100.00
Pred. e 0 1052 100.00
Class recall 100.00 100.00
True p True e Class precision
Pred. p 979 0 100.00
Pred. e 0 1052 100.00
Class recall 100.00 100.00
21
10 DM Methodology steps

• Step 7. Build Model
• 4. NeuralNet
• Performance
• Accuracy91.04
• Lift179.35

True p True e Class precision
Pred. p 907 110 89.18
Pred. e 72 942 92.90
Class recall 92.65 89.54
True p True e Class precision
Pred. p 907 110 89.18
Pred. e 72 942 92.90
Class recall 92.65 89.54
22
10 DM Methodology steps

• Step 8. Assess Models
• AccuracyRipper and Decision Tree have better
  performances

23
10 DM Methodology steps

• Step 8. Assess Models
• Lift (to compare the performances of different
  classification models)Ripper and Decision Tree
  have higher lifts

24
10 DM Methodology steps

• Step 9. Deploy Models
• We havent go out and find real mushrooms
• Step 10. Assess Results
• Conclusion and questions
• Maybe ripper and decision tree are better models
  for nominal data
• How Rapid Miner separates training data from test
  data

25
(No Transcript)