Download (3) 2013 test

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Recursive InterNetwork Architecture (RINA) wikipedia , lookup

IEEE 802.1aq wikipedia , lookup

Airborne Networking wikipedia , lookup

List of wireless community networks by region wikipedia , lookup

CAN bus wikipedia , lookup

Kademlia wikipedia , lookup

Transcript
Assessment 2, Data Mining
IAA, Jan. 2013
1. (8 pts.) Terminology: How many “nodes” __5__ and how many “leaves” (terminal nodes) __3___ are
in this decision tree?
___
root node|___|
_________|__________
_|_
_|_
node |___|
node |___| leaf
________|_______
_|_
_|_
node |___|leaf
node|___|leaf
2. (10) In our neural network example with the veterans data, our first attempt at fitting a model using
all of the features (inputs) failed to converge. What did we do to remedy that situation? Explain briefly.
We preceded the neural network with a logistic regression using stepwise to select a model. Inputs not
selected are automatically eliminated from the neural network node.
3. (10 pts.) What Chi-Square p-value _0.001_gives logworth 3?
-3 =log10(p)  10**(-3)= p  0.001=p
4. (12 pts.) I have a neural network for predicting the probability p that someone will be a good
saleswoman. For a particular applicant for a sales job, my neural network gives estimated probability
p=0.3 that she will be a good saleswoman. If she is not good, she will cost me $4,000 per month in the
long run because her sales are less than her salary. On the other hand, if she is good, she will generate
$20,000 in profit (after adjusting for her salary) per month. Based on this information, should I hire the
applicant? Explain using numbers.
Expected profit = -4000(.7)+20000(.3) = -2800+6000 >0 so I would hire her
5. (10 pts.) A fire alarm company runs a logistic regression to estimate the probability p of someone
waking from sleep as a function of the alarm noise level X. Their regression suggests using p = eL/(1+eL)
to estimate the probability where L = -10 +0.5X. For what value of X is there a 50% chance of
awakening?
L=0 is the 50-50 cutoff (because p=1/(1+1) at L=0). Setting L=-10+.5X = 0 implies X=20.
6. (20 pts.) In a group of 10,000 people, there are 5,000 that bought a smart phone and 2,000 that
bought a Kindle reader. Both of these counts include 1,500 people that bought both. The other 4,500
people bought neither. Assume no priors or decisions (i.e. profits & losses) were specified.
For the “rule” Phone=>Kindle, compute the
Support: 1500/10000=15%
Confidence: 1500/5000=30%_
Expected Confidence: 2000/10000=20% and
Lift: 30/20=1.5
7. (30 pts.) A tree for a binary response (0 or 1) has just two leaves. No decisions (i.e. profits & losses) or
priors were specified. Here are the counts of 0s and 1s in the two leaves.
Leaf 1
400 1s
100 0s
Leaf 2
200 1s
300 0s
I can call everything a one which gives the point __(1,1)__ (I misclassify every 0) on the ROC curve
or
I can call everything in leaf 1 a one and everything in leaf 2 a zero which gives the point ___ (1/4, 2/3) __
or
I can call everything a zero which gives the point ____(0,0)____ (I misclassify every 1).
For the second of the three decision rules above what, are the sensitivity 2/3 and specificity 3/4 of the
decision rule.