Scribd Logo
Upload

Welcome to Scribd!

  • 33 views

    Solutions To Part 2 of The Mid-Term Examination: I.E. 1062/2062 DATA MINING

    Uploaded by

    Vasanth Raja
    This document provides solutions to questions from a data mining mid-term exam. It includes: 1) Calculations of entropy and information gain for splitting a dataset using attributes like 'Department' and 'Age'. 2) A description of the input and output nodes needed for a neural network model for this dataset. 3) An explanation of the covering algorithm used to generate a rule for the 'Medium' class, which results in the rule "If (Status=Senior) and (Department=Sales) then MEDIUM".

    Copyright:

    Attribution Non-Commercial (BY-NC)
    Download as PDF, TXT or read online from Scribd

    Solutions To Part 2 of The Mid-Term Examination: I.E. 1062/2062 DATA MINING

    Uploaded by

    Vasanth Raja
    33 views4 pages
    This document provides solutions to questions from a data mining mid-term exam. It includes: 1) Calculations of entropy and information gain for splitting a dataset using attributes like 'Department' and 'Age'. 2) A description of the input and output nodes needed for a neural network model for this dataset. 3) An explanation of the covering algorithm used to generate a rule for the 'Medium' class, which results in the rule "If (Status=Senior) and (Department=Sales) then MEDIUM".

    Original Title

    ex1sol

    Copyright

    © Attribution Non-Commercial (BY-NC)

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    This document provides solutions to questions from a data mining mid-term exam. It includes: 1) Calculations of entropy and information gain for splitting a dataset using attributes like 'Department' and 'Age'. 2) A description of the input and output nodes needed for a neural network model for this dataset. 3) An explanation of the covering algorithm used to generate a rule for the 'Medium' class, which results in the rule "If (Status=Senior) and (Department=Sales) then MEDIUM".

    Copyright:

    Attribution Non-Commercial (BY-NC)
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    33 views4 pages

    Solutions To Part 2 of The Mid-Term Examination: I.E. 1062/2062 DATA MINING

    Uploaded by

    Vasanth Raja
    This document provides solutions to questions from a data mining mid-term exam. It includes: 1) Calculations of entropy and information gain for splitting a dataset using attributes like 'Department' and 'Age'. 2) A description of the input and output nodes needed for a neural network model for this dataset. 3) An explanation of the covering algorithm used to generate a rule for the 'Medium' class, which results in the rule "If (Status=Senior) and (Department=Sales) then MEDIUM".

    Copyright:

    Attribution Non-Commercial (BY-NC)
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 4

    I.E.

    1062/2062 DATA MINING


    Solutions to Part 2 of the mid-term examination

    Question 1
    a) (4 points)
    The raw data set of 155 instances has a total of 40+40+20+4=104 in class Low, 30+3+10=43 in
    class Medium, and 5+3=8 in class High. Thus the entropy measure at the root node =
    I(104,43,8) =
    -(104/155)log2(104/155) - (43/155)log2(43/155) - (8/155)log2(8/155) = 1.1202

    b) (8 points)
    Consider "Department" for the splitting attribute: We have
    Sales: 80 Low, 30 Medium, 0 High (total of 110 instances)
    I(80,30,0) = -(80/110)log2(80/110) - (30/110)log2(30/110) - (0/110)log2(0/110) = 0.8454
    Systems: 20 Low, 3 Medium, 8 High (total of 31 instances)
    I(20,3,8) = -(20/31)log2(20/31) - (3/31)log2(3/31) - (8/31)log2(8/31) = 1.2383
    Marketing: 4 Low, 10 Medium, 0 High (total of 14 instances)
    I(4,10,0) = -(4/14)log2(4/14) - (10/14)log2(10/14) - (0/14)log2(0/14) = 0.8631
    Thus E(Department) = (110/155)0.8454 + (31/155)1.2383 + (14/155)0.8631 = 0.9255

    Thus Expected Information Gain from having “Department” at the root node =
    I(104,43,8) - E(Department) = 1.1202 – 0.9255 = 0.1945

    To compute the Gain Ratio we need the intrinsic value of the split, which is = I(110,31,14) =
    -(110/155)log2(110/155) – (31/155)log2(31/155) - (14/100)log2(14/155) = 1.1288
    Thus gain ratio = 0.1945/1.1288 = 0.1724

    c) (2 points)
    There are no Junior, Systems people who are 40+. We therefore look at the instances arriving at
    Junior and Systems: there are 23, with 20 in Low and 3 in Medium. We then assign the “???”
    leaf the majority class out of these 23 instances which is LOW.

    d) (7 points)
    Consider the 48 Senior status records: we have the following data:

    Department Age Salary Count


    Sales 30's Medium 30
    Systems 30's High 5
    Systems 40+ High 3
    Marketing 40+ Medium 10

    Thus we have 48 records with 0 in Low, 40 in Medium, 8 in High. The entropy measure at the
    node being considered is thus = I(0,40,8) =
    -(0/48)log2(0/48) - (40/48)log2(40/48) - (8/48)log2(8/48) = 0.65
    AGE AS SPLITTING ATTRIBUTE:
    20’s: (0 instances)
    I(0,0,0) = 0
    30’s: 0 Low, 30 Medium, 5 High (total of 35 instances)
    I(0,30,5) = -(0/35)log2(0/35) - (30/35)log2(30/35) - (5/35)log2(5/35) = 0.5917
    40+: 0 Low, 10 Medium, 3 High (total of 13 instances)
    I(0,10,3) = -(0/13)log2(0/13) - (10/13)log2(10/13) - (3/13)log2(3/13) = 0.7793

    Thus E(Age) = (0/48)0 + (35/48)0.5917 + (13/48)0.7793= 0.6425

    Thus Expected Information Gain from age =


    I I(0,40,8) - E(Age) = 0.65 - 0.6425 = 0.0075

    DEPARTMENT AS SPLITTING ATTRIBUTE:


    Sales: 0 Low, 30 Medium, 0 High (total of 30 instances)
    I(0,30,0) = -(0/30)log2(0/30) - (30/30)log2(30/30) - (0/30)log2(0/30) = 0
    Systems: 0 Low, 0 Medium, 8 High (total of 8 instances)
    I(0,0,8) = -(0/8)log2(0/8) - (0/8)log2(0/8) - (0/8)log2(0/8) = 0
    Marketing: 0 Low, 10 Medium, 0 High (total of 10 instances)
    I(0,10,0) = -(0/10)log2(0/10) - (10/10)log2(10/10) - (0/10)log2(0/10) = 0

    Thus E(Department) = 0

    Thus Expected Information Gain from Department =


    I I(0,40,8) - E(Department) = 0.65 - 0.0 = 0.65

    So we choose Department and this completes the tree. The final tree is as below:
    104,43,8

    Status
    Senior Junior
    0,40,8 104,3,0

    Department Department
    Marketing Marketing
    Sales 4,3,0 Sales
    0,10,0 0,30,0 80,0,0
    Systems
    0,0,8
    MEDIUM MEDIUM LOW Systems LOW
    HIGH 20,3,0

    Age
    20's 40's
    20,0,0 30's 0,0,0
    0,3,0

    LOW MEDIUM LOW


    e) (2 points)
    Junior Sales employee in her 30's leads to a classification of LOW as shown above by the red
    arrow (note that age is irrelevant to the classification!).
    f) (3 points)
    If we prune the portion shown above we get a single classification leaf of LOW instead since this
    is the majority classification for the 23 instances Status=Junior and Department+Systems (in fact,
    the entire righthand side under Status=Junior could be replaced with a single leaf labeled LOW!).
    As compared to the full tree that is perfect and has 0 error, the error after pruning is 3/155 =
    1.94% (the 3 Medium instances that are Junior an Systems are now wrongly classified as Low).

    Question 2 (5 points)
    A neural net will need seven input nodes and three output nodes defined as below:
    INPUT
    Node 1: 1 if Department=Sales, 0 otherwise
    Node 2: 1 if Department=Systems, 0 otherwise
    Node 3: 1 if Department=Marketing, 0 otherwise
    Node 4: 1 if Age=20’s, 0 otherwise
    Node 5: 1 if Age=30’s, 0 otherwise
    Node 6: 1 if Age=40+, 0 otherwise
    Node 7: 1 if staus is senior, 0 if junior
    OUTPUT
    Node 1: 1 if in class Low, 0 otherwise
    Node 2: 1 if in class Medium, 0 otherwise
    Node 3: 1 if in class High, 0 otherwise

    With the above definition, the instance of Part 5 (Sales, 30’s, Junior) would lead to the input
    vector (1,0,0,0,1,0,0).

    Question 3 (9 points)
    Covering Algorithm for class Medium
    Department
    Sales: MEDIUM=30 out of 110 so p/t=30/110=0.27
    Systems: MEDIUM=3 out of 31 so p/t=3/31=0.10
    Marketing: MEDIUM=10 out of 14 so p/t=10/14=0.71
    Status
    Junior: MEDIUM=3 out of 107 so p/t=3/107 = 0.03
    Senior: MEDIUM=40 out of 48 so p/t=40/48=0.83 (best)
    Age
    20’s: MEDIUM=0 out of 60 so p/t=90/450=0.0
    30’s: MEDIUM=33 out of 82 so p/t=33/82=0.40
    40+: MEDIUM=10 out of 13 so p/t=10/13=0.77

    So the partial rule is “If (Status=Senior) then MEDUM” with an accuracy of 0.83.
    Considering only the instances with Senior Status; there are 48 of these that are covered by the
    above clause we have
    Department
    Sales: MEDIUM=30 out of 30 so p/t=1.0 (best)
    Systems: MEDIUM=0 out of 8 so p/t=0
    Marketing: MEDIUM=10 out of 10 so p/t=1.0

    Age
    20’s: MEDIUM=0 out of 0 so p/t is not meaningful
    30’s: MEDIUM=30 out of 35 so p/t=30/35=0.86
    40+: MEDIUM=10 out of 13 so p/t=10/13=0.77

    Both “Department=Sales” and “Department=Marketing” lead to perfect classification.


    However Sales has a higher coverage (30 as opposed to 10); so we pick Sales.

    The final rule is "If (Status=Senior) and (Department=Sales) then MEDIUM" - this rule
    covers a total of 30 instances (out of 155).

    You might also like