Monday, September 29, 2014

Notes on Maximum Likelihood, Maximum A Posteriori and Naive Bayes

Let \(\data\) be a set of data generated from some distribution parameterized by \(\theta\). We want to estimate the unknown parameter \(\theta\). What we can do?

Essentially, we want to find a most likely value of \(\theta\) given \(\data\), that is \(\arg \max P(\theta | \data)\). According to Bayes Rule, we have

\[ P(\theta \given \data) = \frac{P(\data \given \theta)P(\theta)}{P(\data)} \]

and the terms have the following meanings:

  • \(P(\theta \given \data)\): Posterior
  • \(P(\data \given \theta)\): Likelihood
  • \(P(\theta)\): Prior
  • \(P(\data)\): Evidence

Maximum Likelihood Estimation (MLE)

An easy way out is to use the MLE method. We want to find a \(\theta\) the best explains the data. That is, we maximize \(P(\data \given \theta)\). Denote such a value as \(\hat{\theta}_{ML}\). We have

\[ \hat{\theta}_{ML} = \argmax_\theta P(\data \given \theta) = \argmax_\theta P(\mathbf{x}_1, \ldots, \mathbf{x}_N \given \theta ) \]

Note that the above \(P\) is a joint distribution over the data. We usually assume the observations are independent. Thus, we have

\[ P(\mathbf{x}_1, \ldots, \mathbf{x}_N \given \theta ) = \prod_{i=1}^{N} P(\mathbf{x}_i \given \theta ) \]

We usually use logarithm to simplify the computation, as logarithm is monotonically increasing. Thus, we write:

\[ \mathcal{L}(\data \given \theta) = \sum_{i=1}^N \log P(\mathbf{x}_i \given \theta ) \]

Finally, we seek for the ML solution:

\[ \hat{\theta}_{ML} = \argmax_\theta \mathcal{L}(\data \given \theta) \]

If we know the distribution \(P\), we can usually solve the above by setting derivative of \(\theta\) to 0 and solve for \(\theta\), that is,

\[ \frac{\partial L}{\partial \theta} = 0 \]

Maximum A Posteriori (MAP)

In MAP, we maximize \(P(\theta \given \data)\) directly. Denote the MAP hypothesis as \(\hat{\theta}_{MAP}\), we have:

\[\begin{array}{rl} \hat{\theta}_{MAP} = & \argmax_\theta P(\theta \given \data) \\ = & \argmax_\theta \frac{P(\data \given \theta)P(\theta)}{P(\data)} \\ = & \argmax_\theta P(\data \given \theta)P(\theta) \end{array}\]

Note that the last step is due to the evidence (data) \(\data\) is constant, and thus can be omitted in \(\argmax\).

At this step, we notice that the only difference between \(\hat{\theta}_{ML}\) and \(\hat{\theta}_{MAP}\) is the prior term \(P(\theta)\). Another way to interpret is that we consider \(MAP\) is more general than \(MLE\), as if we assume all the possible \(\theta\) are equally probable a priori, e.g., they have the same prior probability, or uniform prior, we can effectively remove \(P(\theta)\) from the MAP formula, and it looks like exactly the same as MLE.

Finally, if the independent observation holds, again we can use logarithm and expand \(\hat{\theta}_{MAP}\) as:

\[ \begin{array}{rl} \hat{\theta}_{MAP} = & \argmax_\theta L(\data \given \theta) \\ = & \argmax_\theta \sum_{i=1}^{N} \log P(\mathbf{x}_i \given \theta ) + \log P(\theta) \end{array} \]

The extra prior term has the effect that we are essentially ‘pulling’ the \(\theta\) distribution towards prior value. This makes sense as we are putting our domain knowledge as prior and intuitively the estimation is biased towards the prior value.

Naive Bayes Classifier

Assume that we are given a set of data \(\data\), where each example \(\mathbf{x_j}=(a_1, a_2, \ldots, a_n)\), which can be viewed as conjunctions of attributes values. \(v_j \in V\) is the corresponding class value. Using MAP, we can classify an example \(\mathbf{x}\) as:

\[v_{MAP}=\argmax_{v_j\in V} P(v_j \given a_1, \ldots, a_n)\]

The problem is that it is hard to find a joint distribution for \(P(\mathbf{x} \given \theta)\). If we use the data to estimate the distribution, we typically don’t have enough data for each attribute. In other words, the data we have is very sparse compared to the whole distribution space.

Naive bayes makes the assumption that each attribute is conditionally independent given the target class \(v_j\), that is,

\[P(a_1, \ldots, a_n \given v_j) = \prod_{i=1}^n P(a_i \given v_j)\]

which can be easily estimated from the data. Thus, we have the following naive bayes classifier:

\[v_{NB} = \argmax_{v_j \in V} P(v_j) \prod_{i=1}^n P(a_i \given v_j)\]

Note that the learning of naive bayes simply involves in estimating \(P(a_i \given v_j)\) and \(P(v_j)\) based on the frequencies in the training data.

Normally the conditional independence assumption does not hold, but naive bayes performs well even if so. More importantly, when conditional independence is satisfied, Naive Bayes corresponds to MAP classification.


MLE, MAP and Naive Bayes are all connected. While MLE and MAP are parameter estimation methods that returns a single value of the paramter being estimated, NB is a classifier that predicts the probability of the class that an example belongs to. We also have the following insightes:

  • Given the data, MLE considers the paramter to be a constant and estimates a value that provide maximum support for the data.
  • MLE does not allow us to ‘inject’ our beliefs about the likely values for the parameter (prior) in the estimation process.
  • MAP allows the fact that the paramter can take values from a prior (non-uniform) distribution that express our prior beliefs regarding the paramters.
  • MAP returns paramter value where the probability is highest given data.
  • Again, both MLE and MAP returns a single and specific value for the paramter. By contrast, bayesian estimation computes the full posterior distribution \(P(\theta \given \data)\).


No comments: