“Training” a Naive Bayes Classifier

Recall the naive Bayes classifier:
- \[P(Y \mid X_1, X_2, ..., X_d) \propto P(Y)\prod_{i=1}^d P(X_i \mid Y)\]
To perform classification we need the:
- Class priors: \(P(Y)\)
- Class-conditional attribute probabilities: \(P(X_i \mid Y)\) for all \(i\).
These were the tallies from our in-class exercise:
- \(Spy\) \(Golfer\) \(Fedora\) Count
  
  T T T 1
  
  T T F 3
  
  T F T 1
  
  T F F 0
  
  F T T 4
  
  F T F 3
  
  F F T 6
  
  F F F 2
- From this we can easily estimate our priors:
  - \(P(Spy=True) = 5/20 = .25\)
  - \(P(Spy=False) = 15/20 = .75\)
- We can also calculate the (full) class-conditional probability distributions:
  - \(Golfer\) \(Fedora\) \(P(Golfer, Fedora | Spy = True)\)
    
    T T 1/5 = .2
    
    T F 3/5 = .6
    
    F T 1/5 = .2
    
    F F 0
    
    \(Golfer\) \(Fedora\) \(P(Golfer, Fedora | Spy = False)\)
    
    T T 4/15 \(\approx\) .27
    
    T F 3/15 = .2
    
    F T 6/15 = .4
    
    F F 2/15 \(\approx\) .13
However, for naive Bayes classification we instead need this:
- \(Golfer\) \(P(Golfer | Spy = True)\)
  
  T 4/5 = .8
  
  F 1/5 = .2
  
  \(Golfer\) \(P(Golfer | Spy = False)\)
  
  T 7/15 \(\approx\) .47
  
  F 8/15 \(\approx\) .53
  
  \(Fedora\)
  \(P(Fedora | Spy = True)\)
  
  T 2/5 = .4
  
  F 3/5 = .6
  
  \(Fedora\)
  \(P(Fedora | Spy = False)\)
  
  T 10/5 \(\approx\) .66
  
  F 5/15 \(\approx\) .33

\(Spy\)	\(Golfer\)	\(Fedora\)	Count
T	T	T	1
T	T	F	3
T	F	T	1
T	F	F	0
F	T	T	4
F	T	F	3
F	F	T	6
F	F	F	2

\(Golfer\)	\(P(Golfer \| Spy = True)\)
T	4/5 = .8
F	1/5 = .2

\(Golfer\)	\(P(Golfer \| Spy = False)\)
T	7/15 \(\approx\) .47
F	8/15 \(\approx\) .53

\(Fedora\)	\(P(Fedora \| Spy = True)\)
T	2/5 = .4
F	3/5 = .6

\(Fedora\)	\(P(Fedora \| Spy = False)\)
T	10/5 \(\approx\) .66
F	5/15 \(\approx\) .33

Performing Classification (Non-Naive):

Assume we have a suspect who is a golfer and who wears a fez, are they a spy?
- We could apply (non-naive) Bayes rule:
- \(P(Spy = T \mid Golf = T, Fedora = F) = \frac{P(Golf = T, Fedora = F \mid Spy=T)P(Spy=T)}{P(Golf = T, Fedora = F)}\)
- \({}=\frac{P(Golf = T, Fedora = F \mid Spy=T)P(Spy=T)}{P(Golf = T, Fedora = F \mid Spy=T)P(Spy=T) + P(Golf = T, Fedora = F\mid Spy=F)P(Spy=F)} = \frac{.6 \times .25}{.6 \times .25 + .2 \times .75} = .5\)

\(P(Spy=T \mid Golf = T, Fedora = F) \propto P(Y)\prod_{i=1}^d P(X_i \mid Y) = .25 \times .8 \times .6 = .12\)
\(P(Spy=F \mid Golf = T, Fedora = F) \propto P(Y)\prod_{i=1}^d P(X_i \mid Y) = .75 \times .47 \times .33 \approx .116\)

Pros:
- Provides a meaningful class probability, not just a class label
- Works in the face of missing attributes (just don’t include them in the calculation)
- Relatively easy to interpret: we can examine the class-conditional probabilities for individual attributes.
Cons:
- Classification performance may be worse than other classifiers: Most real classification tasks will violate the independence assumption to some extent.

Naive Bayes classifier: \[P(Y \mid X_1, X_2, ..., X_d) \propto P(Y)\prod_{i=1}^d P(X_i \mid Y)\]
Each \(P(X_i \mid Y)\) is less then 1.
What is \(.5^{100}\)? \(.5^{1000}\)?
Recall that \(\log(ab) = \log(a) + \log(b)\)
Also, the log function is monotonic: if \(a > b\) then \(\log(a) > \log(b)\)
So, practical implementations generally work with logs: \[\log \left(P(Y)\prod_{i=1}^d P(X_i \mid Y)\right) = \log (P(Y)) + \sum_{i=1}^d \log(P(X_i \mid Y))\]