7.2. Naive Bayes Text Classification

Classifiers must determine the most likely class \(C_i\) for a given input \(\underline{x}\). Using probability theory, this task can be described as follows: For all possible classes \(C_i\) determine the conditional probability

\[ P(C_i | \underline{x}), \]

i.e. the probability that the given input \(\underline{x}\) falls into class \(C_i\). Then select the class for which this probability is maximum.

In the training phase the Naive Bayes Algorithm estimates the so-called likelihood function

\[ p(\underline{x} \mid C_i) \]

and the a-priori class probabilities

\[ P(C_i) \]

from the given training data. In the inference phase it then applies the Bayes-Rule for determining the a-posteriori-probability

(7.1)\[ P(C_i \mid \underline{x}) = \frac{p(\underline{x} \mid C_i) \cdot P(C_i)}{p(\underline{x})}. \]

Then the class \(C_i\), for which \(P(C_i \mid \underline{x})\) is maximum is determined. Note that for determining the class, for which \(P(C_i \mid \underline{x})\) is maximal, doesn’t require the so-called evidence \(p(\underline{x})\), since this value is independent of the class, i.e.

(7.2)\[ argmax_i(P(C_i \mid \underline{x})) = argmax_i (\frac{p(\underline{x} \mid C_i) \cdot P(C_i)}{p(\underline{x})}) = argmax_i (p(\underline{x} \mid C_i) \cdot P(C_i)) \]

The calculation of the likelihood function \(p(\underline{x} \mid C_i)\) is quite difficult if \(\underline{x}=(x_1,x_2,\ldots,x_Z)\) is a vector of interdependent variables \(x_i\). However, the Naive Bayes Classification is based on the simplifying assumption that the input variables \(x_i\) are independent of each other. In this case, the conditional compound probability function \(p(x_1,x_2,\ldots,x_Z \mid C_i)\) is simplified to

(7.3)\[ p(x_1,x_2,\ldots,x_Z \mid C_i)=\prod\limits_{j=1}^Z p(x_j | C_i). \]

Hence, in the inference face the classifier must determine

(7.4)\[ argmax_i \left( \prod\limits_{j=1}^Z p(x_j | C_i) \cdot P(C_i) \right) \]

The conditional probability \(p(x_j | C_i)\) is the probability, that in documents of class \(C_i\) the value of the \(j.th\) component in the feature-vector is \(x_j\).

If the Naive Bayes Classifier is applied for text-classification, then \(p(x_j | C_i)\) describes the probability, that word \(x_j\) appears at least once in a document of class \(C_i\). This probability can be estimated by

(7.5)\[ p(x_j | C_i) = \frac{\#(x_j,C_i)}{\#(C_i)}, \]

where \(\#(x_j,C_i)\) is the number of class \(C_i\)-training-documents, which contain word \(x_j\) and \(\#(C_i)\) is the number of class \(C_i\)-training-documents.

The a-priori-possiblities are estimated by

(7.6)\[ P(C_i)=\frac{\#(C_i)}{N}, \]

where \(N\) is the total number of training-documents.

Moreover, inference, in the case that the Naive Bayes algorithm is applied for text-classification applies the following variant of equation (7.4):

(7.7)\[ argmax_i \left( \prod\limits_{x_j \in D} p(x_j | C_i) \cdot P(C_i) \right), \]

where \(D\) is the set of all words, which are contained in the document that shall be classified. In contrast to equation (7.4), in equation (7.7) the product is calculated only over the words, contained in the current document.

Example: Naive Bayes Spam Filter

The following labeled Emails are available for training a Naive Bayes Classifier:

  • 4 Training documents labeled with class Good:

      - nobody owns the water
  - the quick rabbit jumps fences
  - the quick brown fox jumps
  - next meeting is at night

  • 4 Training documents labeled with class Bad:

      - buy pharmaceuticals now
  - make quick money at the online casino
  - meeting with your superstar
  - money like water

Task: Determine the class (Bad or Good) which is assigned to the new Email

	- the money jumps

by a Naive Bayes Classifier.

Solution:

  1. Determine for the relevant words the, money and jumps and both classes the conditional probabilites according to equation (7.5):

    \[\begin{split} P(the|Bad) & = & 0.25 \\ P(the|Good) & = & 0.75 \\ P(money|Bad) & = & 0.5 \\ P(money|Good) & = & 0.0 \\ P(jumps|Bad) & = & 0.0 \\ P(jumps|Good) & = & 0.5 \\ \end{split}\]

  2. Determine for both classes the a-priori probability according to equation (7.6):

    \[\begin{split} P(Bad) & = & 0.5 \\ P(Good) & = & 0.5 \\ \end{split}\]

  3. Determine the class, for which the argument of equation (7.7) is maximal:

    \[\begin{split} \mbox{Class Bad: } & 0.25 \cdot 0.5 \cdot 0.0 \cdot 0.5 = 0 \\ \mbox{Class Good: } & 0.75 \cdot 0.0 \cdot 0.5 \cdot 0.5 = 0 \\ \end{split}\]

For both classes the same a-posteriori probability value has been calculated. In this case no classification is possible!

The example above uncovers a drawback of Naive Bayes classification. If the document, that shall be classified, contains a word \(x_j\), which does not appear in the class \(C_i\)-training data, the corresponding conditional probability \(P(x_j|C_i)\) is zero and as soon as one of the factors of (7.7) is zero, the entire product is zero. In order to avoid this problem of zero-factors, Naive Bayes classifiers are usually modified by applying smoothing .

Smoothing in the context of Naive Bayes classification means, that the conditional probabilities are estimated not as defined in equation (7.5), but in a slightly modified form, which guarantees that the values of the smoothed conditional probabilities are always non-zero. There exist different smoothing techniques. An approach, which is frequently applied for smoothing in the context of Naive Bayes document classification is defined by replacing the conditional probabilities \(p(x_j|C_i)\) in equation (7.7) by the following weighted conditional probabilities:

(7.8)\[ p_{weight}(x_j \mid C_i)=\frac{w \cdot P_{ass,i,j} + |x_j| \cdot P(x_j \mid C_i)}{w+|x_j|}, \]

where

  • \(P_{ass,i,j}\) is an assumed probability that a word \(x_j\) belongs to class \(C_i\). By default this probability can be set to \(1/K\), where \(K\) is the number of classes, that must be distinguished. However, this probability can also be set individually, e.g. for word viagra the probabilities may be \(P_{ass,viagra,bad}=0.95\) and \(P_{ass,viagra,good}=0.05\). I.e. this term provides the possibility to integrate prior-knowledge.

  • \(w\) is a weight-factor, which can be set in order to control how strong the assumed probability \(P_{ass,i,j}\) contributes to the smoothed probability \(p_{weight}(x_j \mid C_i)\). Default: \(w=1\).

For the classification of a new document containing words \(D\), instead of equation (7.7), the following equation is applied:

(7.9)\[ argmax_i \left( \prod\limits_{x_j \in D} p_{weight}(x_j | C_i) \cdot P(C_i) \right), \]

Example: Naive Bayes Spam Filter with Smooting

The given training data is the same as in example exampleNB. The smoothed conditional probabilities according to (7.8) are then:

\[\begin{split} P_{weight}(the|Bad) & = & \frac{0.5 + 4 \cdot 0.25}{1+4}=0.3 \\ P_{weight}(the|Good) & = & \frac{0.5 + 4 \cdot 0.75}{1+4}=0.7 \\ P_{weight}(money|Bad) & = & \frac{0.5 + 2 \cdot 0.5}{1+2}=0.5 \\ P_{weight}(money|Good) & = & \frac{0.5 + 2 \cdot 0.0}{1+2}=0.167 \\ P_{weight}(jumps|Bad) & = & \frac{0.5 + 2 \cdot 0.0}{1+2}=0.167 \\ P_{weight}(jumps|Good) & = & \frac{0.5 + 2 \cdot 0.5}{1+2}=0.5 \\ \end{split}\]

The a-priori class probabilities are unchanged:

\[\begin{split} \mbox{Class Bad: } & 0.25 \cdot 0.5 \cdot 0.0 \cdot 0.5 = 0 \\ \mbox{Class Good: } & 0.75 \cdot 0.0 \cdot 0.5 \cdot 0.5 = 0 \\ \end{split}\]

Applying the smoothed probabilities, the class, for which the argument of equation (7.9) is maximal for the new email \((the, money, jumps)\):

\[\begin{split} \mbox{Class Bad: } & 0.3 \cdot 0.5 \cdot 0.167 \cdot 0.5 = 0.0125 \\ \mbox{Class Good: } & 0.7 \cdot 0.167 \cdot 0.5 \cdot 0.5 = 0.029 \\ \end{split}\]

Hence this email is classified to be Good.

Note, that for classification it was sufficient to calculate

\[ p(\underline{x} \mid C_i) \cdot P(C_i), \]

for all classes \(C_i\) and decide on the class, for which this value is maximal. However, the calculated values are not the a-posteriori-probabilities \(P(C_i | \underline{x})\), since we ignored the evidence \(p(\underline{x})\) (see equation (7.2)). However, we can easily obtain the a-posteriori-probabilities \(P(C_i | \underline{x})\), by applying the marginalisation-rule for the calculation of the evidence in the denominator of the Bayes Rule (equation (7.1))

\[ p(\underline{x}) = \sum\limits_{i=1}^K p(\underline{x},C_i) = \sum\limits_{i=1}^K p(\underline{x} | C_i) \cdot P(C_i). \]

In the example above, the a-posteriori probabilities are then:

\[\begin{split} P(Bad | (the,money,jumps)) & = & \frac{0.0125}{0.0125+0.029} & = & 0.30 \\ P(Good | (the,money,jumps)) & = & \frac{0.029}{0.0125+0.029} & = & 0.70 \\ \end{split}\]
7.1. Validation of Classifiers 7.3. Text Classification Application: Fake News detection