Naive Bayes and Logistics Regression

Posted by GwanSiu on June 17, 2018

1. Naive Bayes Algorithm

First, we see the whole picture of Naive Bayes. Naive bayes algorithm is supervised learning algorithm, it has two formulations for different inputs, such as discrete-valued inputs and continuous-valued inputs. Maximum likellihood principle is usually used in parameters estimation of naive bayes.

Then, we see the the assumption of Naive Bayes is conditional independence:

\[\begin{equation} (\forall i,j,k)P(X=x_{i}\arrowvert Y=y_{j}, Z=z_{k})=P(X=x_{i}\arrowvert Z=z_{k}) \end{equation}\]

In general, we assume that Y is any discrete-valued variable, and attributes $X_{1},…,X_{n}$ are any discrete or realvalued attributes. The goal of Naive Bayes is to train a classifier that will output the probability distribution over possible values of Y, for each new instance X that we ask it to classify.

The expression for the probability that Y will take on its $k\text{th}$ possible value, according to Bayes rule, is:

\[\begin{equation} P(Y=y_{k}\arrowvert X_{1},...,X_{n})=\frac{P(X_{1},...,X_{n}\arrowvert Y=y_{k})}{\sum_{j}P(Y=y_{j})P(X_{1},...,X_{n}\arrowvert Y=y_{j})} \end{equation}\]

where the sum is taken over all possible values y j of Y, the arributes of $X$ is independent given $Y$. Now, the formulation can be rewrite as:

\[\begin{equation} P(Y=y_{k}\arrowvert X_{1},...,X_{n})=\frac{\prod_{i}^{n}P(X_{i}\arrowvert Y=y_{k})P(Y=y_{k})}{\sum_{j}P(Y=y_{j})\prod_{i}^{n}P(X_{i}\arrowvert Y=y_{j})} \end{equation}\]

Equation(3) is the fundamental formulation of naive bayes algorithm for naives bayes algorithm.

For example, given a new instance $X_{new} =(X_{1},…,X_{n})$, this equation shows how to calculate the probability that $Y$ will take on any given value, given the observed attribute values of $X$ new and given the distributions $P(Y)$ and $P(X_{i}\arrowvert Y)$ estimated from the training data. If we are interested only in the most probable value of Y, then we have the Naive Bayes classification rule(MAP):

\[\begin{equation} Y\leftarrow \arg\max_{y_{k}} \frac{P(Y=y_{k})\prod_{i}^{n}P(X_{i}\arrowvert Y=y_{k})}{\sum_{j}P(Y=y_{j})\prod_{i}^{n}P(X_{i}\arrowvert Y=y_{j})} \end{equation}\]

usually, we consider denominator as a constant term, and equation above can be simplified as, equivalent to MLE:

\[\begin{equation} Y\leftarrow \arg\max_{y_{k}} P(Y=y_{k})\prod_{i}^{n}P(X_{i}\arrowvert Y=y_{k}) \end{equation}\]

1.1 Models of Naive Bayes

1.1.1 Model 1: Bernoulli Naive Bayes

Support: Binary vectors of length $K$: $x\in {0,1}^{K}$.

Generative story:

  1. $Y\sim \text{Bernoulli}(\Phi)$.
  2. $X_{K}\sim \text{Bernoulli}(\theta_{k,Y}),\forall k\in{1,…,K}$.


\[\begin{align} p_{\phi,\theta}(x,y) &= p_{\phi,\theta}(x_{1},...,x_{K},y)\\ &=p_{\phi}(y)\prod_{k=1}^{K}p_{\theta_{k}}(x_{k}\arrowvert y) \\ &=(\Phi)^{y}(1-\Phi)^{(1-y)}\prod_{k=1}^{K}(\theta_{k,y})^{x_{k}}(1-\theta_{k,y})^{(1-x_{k})} \end{align}\]

Classification: Find the class that maximizes the posterior:

\[\begin{equation} \tilde{y}=\arg\max_{y}p(y\arrowvert x) \end{equation}\]

Training: Find the class-conditional MLE parameters

$P(Y)$ is independent with the class, and is used in all the data. For each $P(X_{k}\arrowvert Y)$, we condition on the data with the corresponding class.

\[\begin{align} \Phi &= \frac{\sum_{i=1}^{N}\mathbb{I}(y^{(i)}=1)}{N} \\ \theta_{k,0}&=\frac{\sum_{i}^{N}\mathbb{I}(y^{(i)}=0\wedge x_{k}^{(i)}=1)}{\sum_{i=1}^{N}\mathbb{I}(y^{(i)}=0)} \\ \theta_{k,1} &= \frac{\sum_{i=1}^{N}\mathbb{I}(y^{(i)}=1\wedge x^{(i)}_{k}=1)}{\sum_{i=1}^{N}\mathbb{I}(y^{(i)}=1)}\\ &\forall k\in {1,...,K} \end{align}\]

1.1.2 Model 2: Multinomial Naive Bayes

Support: Option 1: integer vector(word Ids),$x=[x_{1},…,x_{M}],\text{where } x_{m}\in{1,…,K}$ a word id.

Generative story:

For $i\in {1,…,N}$: $y^{(i)}\sim \text{Bernoulli}(\Phi)$.

For $j\in {1,…,M_{i}}$: $x_{j}^{(i)}\sim \text{Multinomial}(\theta_{y^{(i)}},1)$


\[\begin{align} p_{\phi,\theta} &= p_{\phi}(y)\prod_{k=1}^{K}p_{\theta_{k}}(x_{k}\arrowvert y) \\ &= (\phi)^{y}(1-\phi)^{(1-y)}\prod_{j=1}^{M_{i}}\theta_{y,x_{j}} \end{align}\]

1.1.3 Model 3: Gaussian Naive Bayes

Support: $X\in \mathbb{R}^{K}$

Model: Product of prior and the event model:

\[\begin{align} p(x,y) &= p(x_{1},...,x_{k},y)\\ &= p(y)\prod_{k=1}^{K}p(x_{k}\arrowvert y) \end{align}\]

Gaussian Naive Bayes assumes that $p(x_{k}\arrowvert y)$ is given by a Normal distribution.

  • When $X$ is multivariate-Gaussian vector:
  • The joint probability of a data and it label is:
\[\begin{align} p(x_{n},y_{n}^{k}=1\arrowvert \tilde{\mu},\Sigma) &= p(y_{n}^{k}=1)\times p(x_{n}\arrowvert y_{n}^{k}=1, \tilde{\mu},\Sigma) \\ &= \pi_{k}\frac{1}{(2\pi\arrowvert \Sigma\arrowvert)^{1/2}}exp(-\frac{1}{2}(x_{n}-\tilde{\mu}_{k})^{T}\Sigma^{-1}(x_{n}-\tilde{\mu}_{k})) \end{align}\]
  • The naive Bayes simplification:
\[\begin{align} p(x_{n},y_{n}^{k}=1\arrowvert \mu,\sigma) &= p(y_{n}^{k}=1)\times \prod_{j}p(x_{n}^{j}\arrowvert y_{n}^{k}=1, \mu_{k}^{j}, \sigma_{k}^{j}) \\ &= \pi_{k}\prod_{j}\frac{1}{\sqrt{2\pi}\sigma_{k}^{j}}exp(-\frac{1}{2}(\frac{x_{n}^{j}-\mu_{k}^{j}}{\sigma_{k}^{j}})^{2}) \end{align}\]
  • More generally:
  • where $p(\bullet \arrowvert \bullet)$ is an arbitrary conditional(discrete or continuous) 1-D density:
\[\begin{equation} p(x_{n},y_{n}\arrowvert \eta, \pi) = p(y_{n}\arrowvert \pi)\times \prod_{j=1}^{m}p(x_{n}^{j}\arrowvert y_{n}, \eta) \end{equation}\]

1.1.4 Model 4: Multiclass Naive Bayes

Model: The only change is that we permit $y$ to range over $C$ classes.

\[\begin{align} p(x,y) &= p(x_{1},...,x_{k}, y) \\ &= p(y)\prod_{k=1}^{K}p(x_{k}\arrowvert y) \end{align}\]

Now, $y\sim \text{Multinomial}(\Phi,1)$ and we have a seperate conditional distribution $p(x_{k}\arrowvert y)$ for each of the $C$ classes.

1.2 Regularization

1.2.1 Added- 1 Smooth

1.2.2 Added- $\lambda$ Smooth

1.3 Naive Bayes for Discrete-Valued Inputs

Assume the $n$ input attributes $X_{i}$ each take on $J$ possible discrete values, and $Y$ is a discrete variable taking on $K$ possible discrete values, then our learning task is to estimage two sets of parameters. The first is:

\[\begin{equation} \theta_{ijk} \equiv P(X_{i}=x_{ij}\arrowvert Y=y_{k}) \end{equation}\]

for each input attribute $X_{i}$, each of its possible values $x_{ij}$, and each of the possible values $y_{k}$ of $Y$. Therefore, there will be nJK such parameters, and note only $n(J-1)K$ of these are independent, given that they must satisfy $\sum_{j}\theta_{ijk}=1$ for each pair of $i,k$ values.

The second parameter set is prior probability over $Y$:

\(\begin{equation} \pi_{k} \equiv P(Y=y_{k}) \end{equation}\) Note there are $K$ parameters, $(K-1)$ of which are independent.

Estimation of $\theta$

Givening a set of training examples D, the MLE for $\theta_{ijk}$ are given by:

\[\begin{equation} \tilde{\theta}_{ijk} = \tilde{P}(X_{i}=x_{ij}\arrowvert Y=y_{k}) = \frac{\text{#}D\{X_{i}=x_{ij}\wedge Y=y_{k}}\}{\text{#}D\{Y=y_{k}\}} \end{equation}\]

where the $\text{#}D{x}$ operator returns the number of elements in the set $D$ that satisfy property $x$.

One danger of this maximum likelihood estimate is that it can sometimes result in θ estimates of zero, if the data does not happen to contain any training examples satisfying the condition in the numerator. To avoid this, it is common to use a “smoothed” estimate which effectively adds in a number of additional “hallucinated” examples, and which assumes these hallucinated examples are spread evenly over the possible values of $X_{i}$. This smoothed estimate is given by:

\[\begin{equation} \tilde{\theta}_{ijk} = \tilde{P}(X_{i}=x_{ij}\arrowvert Y=y_{k}) = \frac{\text{#}D\{X_{i}=x_{ij}\wedge Y=y_{k}\}+l}{\text{#}D\{Y=y_{k}}+lJ\} \end{equation}\]

where J is the number of distinct values $X_{i}$ can take on, and $l$ determines the strength of this smoothing(i.e., the number of hallucinated examples is $lJ$). This distribution over the $\theta_{ijk}$ parameters, with equal-valued parameters. If $l$ is set to 1, this approach is called Laplace smoothing.

Estimation of $\pi$

Maximum likelihood for $\pi_{k}$ are:

\[\begin{equation} \tilde{\pi}=\tilde{P}(Y=y_{k})=\frac{\text{#} D\{Y=y_{k}\}}{\arrowvert D\arrowvert} \end{equation}\]

where $\arrowvert D\arrowvert$ denotes the number of elements in the training set $D$.

Alternatively, we can obtain a smoothed estimate, or equivalently a MAP estimate based on a Dirichlet prior over the $\pi_{k}$ parameters assuming equal priors on each $\pi_{k}$, by using the following expression:

\[\begin{equation} \tilde{\pi}=\tilde{P}(Y=y_{k})=\frac{\text{#}D\{Y=y_{k}\}+l}{\arrowvert D\arrowvert +lK} \end{equation}\]

where $K$ is the number of distinct values $Y$ can take on, and $l$ again determines the strength of the prior assumptions relative to the observed data $D$.

1.4 Naive Bayes for Continuous-Valued Inputs

In the case of continuous inputs $X$, we use Gaussian distribution to represent $P(X_{i}\arrowvert Y)$. Specifically, for each possible discrete value $y_{k}$ of $Y$, the distribution of each continuous $X_{i}$ is Gassuain, and is defined by a mean and standard deviation specific to $X_{i}$ and $y_{k}$. In order to train such a Naive Bayes classifier we must therefore estimate the mean and standard deviation of each of these Gaussians:

\[\begin{align} \mu_{ik} &=\mathbb{E}[X_{i}\arrowvert Y=y_{k}] \\ \sigma_{ik}^{2} &= \mathbb{E}[(X_{i}-\mu_{ik})^{2}\arrowvert Y=y_{k}] \end{align}\]

for each attibute $X_{i}$ and each possible value $y_{k}$ of $Y$. These are totally $2nK$ of these parameters, and all of them must be estimated independently.

Also, we must estimate the prior distribution over Y:

\[\begin{equation} \pi_{k} = P(Y=y_{k}) \end{equation}\]

The above formulation assume that e data X is generated by a mixture of class-conditional (i.e., dependent on the value of the class variable Y) Gaussians.

Furthermore, the Naive Bayes assumption introduces the additional constraint that the attribute values $X_{i}$ are independent of one another within each of these mixture components. In particular problem settings where we have additional information, we might introduce additional assumptions to further restrict the number of parameters or the complexity of estimating them.

For example, if we have reason to believe that noise in the observed $X_{i}$ comes from a common source, then we might further assume that all of the $σ_{ik}$ are identical, regardless of the attribute $i$ or class $k$.

Maximum Likelihood Estimation of $\mu_{ik}$:

\[\begin{equation} \tilde{\mu}_{ik} = \frac{1}{\sum_{j}\delta(Y^{j}=y_{k})}\sum_{j}X_{i}^{j}\delta(Y^{j}=y_{k}) \end{equation}\]

where the superscript $j$ refers to the jth training example, and where $\delta(Y=y_{k})$ is 1 if $Y = y_{k}$ and 0 otherwise. Note the role of δ here is to select only those training examples for which $Y = y_{k}$.

Maximum Likelihood Estimation of $\delta^{2}_{ik}$

\[\begin{equation} \tilde{\sigma}_{ik}^{2} = \frac{1}{\sum_{j}\delta(Y^{j}=y_{k)}}\sum_{j}(X_{i}^{j}-\tilde{\mu}_{ik})^{2}\delta(Y^{j}=y_{k}) \end{equation}\]

This maximum likelihood estimator is biased, so the minimum variance unbiased estimator (MVUE) is sometimes used instead. It is

\[\begin{equation} \tilde{\sigma}_{ik}^{2} = \frac{1}{\sum_{j}\delta(Y^{j}=y_{k)}-1}\sum_{j}(X_{i}^{j}-\tilde{\mu}_{ik})^{2}\delta(Y^{j}=y_{k}) \end{equation}\]

2. Logistic Regression and Naive Gaussian Bayes

2.1 Discriminative Model and Generative Model

Generative classifiers learn a model of the joint probability, $p(x,y)$, of the inputs $x$ and the label $y$, and make their predictive by using Bayes rules to calculate $p(y\arrowvert x)$, and then picking the most likely label $y$.

Discriminative classifiers model the posterior $p(y\arrowvert x)$ directly, or the a mapping function from input data $x$ to the class labels.

we can see more deep insight that generative model requires more parameters than discriminative model, because it model generated process of data. Therefore, generative model can handle missing data, and detect abnormal data. However, leaving aside computational issues and matters such as handling missing data, discreiminative model are almost always to be preferred to generative ones.

2.2 Logistic Regression and Gaussian Naive Bayes

2.2.1 Logistic Regression

Logistic regression is discriminative algorithm, which learns functions of the form $f:X\rightarrow Y$, or $P(Y\arrowvert X)$ in the case where $Y$ is discrete-valued, and $X=(X_{1},…,X_{n})$ is any vector containing dicrete or continuous variables.

Logistic Regression assumes a parametric form for the distribution $P(Y\arrowvert X)$, then directly estimates its parameters from the training data. The parametric model assumed by Logistic Regression in the case where $Y$ is boolean is:

\[\begin{align} P(Y=1\arrowvert X) &=\frac{1}{1+exp(\omega_{0}+\sum_{i=1}^{n}\omega_{i}X_{i})} \\ P(Y=0\arrowvert X) &=\frac{exp(\omega_{0}+\sum_{i=1}^{n}\omega_{i}X_{i})}{1+exp(\omega_{0}+\sum_{i=1}^{n}\omega_{i}X_{i})} \end{align}\]

One highly convenient property of this form for $P(Y\arrowvert X)$ is that it leads to a simple linear expression for classification. To classify any given $X$ we generally want to assign the value yk that maximizes $P(Y = y_{k}\arrowvert X)$. Put another way, we assign the label $Y = 0$ if the following condition holds:

\[\begin{equation} 1<\frac{P(Y=0\arrowvert X)}{P(Y=1\arrowvert X)} \end{equation}\]

substituting from equation(17) and equation(18), we can get:

\[\begin{equation} 1<exp(\omega_{0}+\sum_{i=1}^{n}\omega_{i}X_{i}) \end{equation}\]

and taking the natural log of both sides we have a linear classification rule that assigns label $Y = 0$ if $X$ satisfies:

\[\begin{equation} 1 < \omega_{0}+\sum_{i=1}^{n}\omega_{i}X_{i} \end{equation}\]

and assigns Y = 1 otherwise.

Interestingly, the parametric form of $P(Y\arrowvert X)$ used by Logistic Regression is precisely the form implied by the assumptions of a Gaussian Naive Bayes classifier. Therefore, we can view Logistic Regression as a closely related alternative to GNB, though the two can produce different results in many cases.

2.2.2 Gaussian Naive Bayes

Here we can derive the form of $P(Y\arrowvert X)$ entailed by the assumptions of a Gaussian Naive Bayes (GNB) classifier, showing that it is precisely the form used by Logistic Regression. In particular, consider a GNB based on the following modeling assumptions:

  1. $Y$ is boolean, governed by a Bernoulli distribution, with parameter $\pi=P(Y = 1)$.
  2. $X = (X_{1},…,X_{n})$, where each $X_{i}$ is a continuous random variable.
  3. For each $X_{i}$, $P(X_{i}\arrowvert Y = y_{k})$ is a Gaussian distribution of the form $N(\mu_{ik},\sigma_{i})$.
  4. For all $i$ and $j\neq i$, $X_{i}$ and $X_{j}$ are conditionally independent given $Y$.

Note here we are assuming the standard deviations σi vary from attribute to attribute, but do not depend on $Y$.

In general, Bayes rules allow us to write:

\[\begin{equation} P(Y=1\arrowvert X)=\frac{P(Y=1)P(X\arrowvert Y=1)}{P(Y=1)P(X\arrowvert Y=1)+P(Y=0)P(X\arrowvert Y=0)} \end{equation}\]

Dividing both the numerator and denominator by the numerator yields:

\[\begin{equation} P(Y=1\arrowvert X) = \frac{1}{1+\frac{P(Y=0)P(X\arrowvert Y=0)}{P(Y=1)P(X\arrowvert Y=1}} \end{equation}\]

or equivalently,

\[\begin{equation} P(Y=1\arrowvert X) = \frac{1}{1+exp(\ln\frac{P(Y=0)P(X\arrowvert Y=0)}{P(Y=1)P(X\arrowvert Y=1})} \end{equation}\]

Due to our conditional independence assumption we can write this:

\[\begin{align} P(Y=1\arrowvert X) &= \frac{1}{1+exp(\ln\frac{P(Y=0)}{P(Y=1)}+\sum_{i}\ln\frac{P(X_{i}\arrowvert Y=0)}{P(X_{i}\arrowvert Y=1)})} \\ &= \frac{1}{1+exp(\ln\frac{1-\pi}{\pi}+\sum_{i}\ln\frac{P(X_{i}\arrowvert Y=0)}{P(X_{i}\arrowvert Y=1)})} \end{align}\]

Note the final step express $P(Y=0)$ and $P(Y=1)$ in terms of the binomial parameter $\pi$.

Now consider just the summation in the denominator. Given our assumption that $P(X_{i}\arrowvert Y=y_{k})$ is Gaussian, we can expand this term as follows:

\[\begin{align} \sum_{i}\ln\frac{P(X_{i}\arrowvert Y=0)}{P(X_{i}\arrowvert Y=1)} &= \sum_{i}\ln\frac{\frac{1}{\sqrt{2\pi\sigma_{i}^{2}}}exp(\frac{-(X_{i}-\mu_{i0})^{2}}{2\sigma_{i}^{2}})}{\frac{1}{\sqrt{2\pi\sigma_{i}^{2}}}exp(\frac{-(X_{i}-\mu_{i1})^{2}}{2\sigma_{i}^{2}})} \\ &=\sum_{i}\ln exp(\frac{(X_{i}-\mu_{i1})^{2}-(X_{i}-\mu_{i0})^{2}}{2\sigma_{i}^{2}}) \\ &=\sum_{i}(\frac{(X_{i}-\mu_{i1})^{2}-(X_{i}-\mu_{i0})^{2}}{2\sigma_{i}^{2}}) \\ &=\sum_{i}(\frac{(X_{i}^{2}-2X_{i}\mu_{i1}+\mu_{i1}^{2})-(X_{i}-2X_{i}\mu_{i0}+\mu_{i0}^{2}}{2\sigma_{i}^{2}}) \\ &=\sum_{i}(\frac{2X_{i}(\mu_{i0}-\mu_{i1})+\mu_{i1}^{2}-\mu_{i0}^{2}}{2\sigma_{i}^{2}}) \\ &=\sum_{i}(\frac{\mu_{i0}-\mu_{i1}}{\sigma_{i}^{2}}X_{i}+\frac{\mu_{i1}^{2}-\mu_{i0}^{2}}{2\sigma_{i}^{2}}) \end{align}\]

we can write the form, which is the same as the form of Logistic Regression:

\[\begin{equation} P(Y=1\arrowvert X) = \frac{1}{1+exp(\omega_{0}\sum_{i=1}^{n}\omega_{i}X_{i})} \end{equation}\]

where $\omega_{0}=\ln\frac{1-\pi}{\pi}+\sum_{i}\frac{\mu_{i1}^{2}-\mu_{i0}^{2}}{2\sigma_{i}^{2}}$ and $\omega_{i}=\frac{\mu_{i0}-\mu_{i1}}{\sigma_{i}^{2}}$ for i=1,…,n.

Hnece, we have:

\[\begin{equation} P(Y=0\arrowvert X) = \frac{exp(\ln\frac{1-\pi}{\pi}+\sum_{i}(\frac{\mu_{i0}-\mu_{i1}}{\sigma_{i}^{2}}X_{i}+\frac{\mu_{i1}^{2}-\mu_{i0}^{2}}{2\sigma_{i}^{2}}))}{1+exp(\ln\frac{1-\pi}{\pi}+\sum_{i}(\frac{\mu_{i0}-\mu_{i1}}{\sigma_{i}^{2}}X_{i}+\frac{\mu_{i1}^{2}-\mu_{i0}^{2}}{2\sigma_{i}^{2}}))} \end{equation}\]

2.3 Learning Process

2.3.1 Two Classes

2.3.2 Multi Classes

3. Summation of Relation Between Naives Classifier and Logistic Regression

Logistic Regression is discriminative learning algorithm, which directly estimates the parameters of $P(Y\arrowvert X)$, whereas Naive Bayes is genertive learning algortihm, which models posterior distribution $P(Y\arrowvert X)$ by the Bayes rule, in other words, it directly estimates parameters for $P(Y)$ and $P(X\arrowvert Y)$.

We can see that Gaussian Naive Bayes algorithm can imply the parametric form $P(Y\arrowvert X)$ of Logistic Regression. Furthermore, it showed that parameters $\omega_{i}$ in Logistic Regression can be expressed in terms of the Gaussian Naive Bayes parameters.

In fact, if the GNB assumptions hold, then asymptotically (as the number of training examples grows toward infinity) the GNB and Logistic Regression converge toward identical classifiers.

The two algorithms also differ in two ways:

  1. When the GNB modeling assumptions do not hold, Logistic Regression and GNB typically learn different classifier functions. In this case, the asymptotic (as the number of training examples approach infinity) classification accuracy for Logistic Regression is often better than the asymptotic accuracy of GNB. Although Logistic Regression is consistent with the Naive Bayes assumption that the input features $X_{i}$ are conditionally independent given $Y$, it is not rigidly tied to this assumption as is Naive Bayes. Given data that disobeys this assumption, the conditional likelihood maximization algorithm for Logistic Regression will adjust its parameters to maximize the fit to (the conditional likelihood of) the data, even if the resulting parameters are inconsistent with the Naive Bayes parameter estimates.

  2. GNB and Logistic Regression converge toward their asymptotic accuracies at different rates. As Ng & Jordan (2002) show, GNB parameter estimates converge toward their asymptotic values in order $\log n$ examples, where $n$ is the dimension of $X$. In contrast, Logistic Regression parameter estimates converge more slowly, requiring order $n$ examples. The authors also show that in several data sets Logistic Regression outperforms GNB when many training examples are available, but GNB outperforms Logistic Regression when training data is scarce.

4. Summary

  1. We can use Bayes rule as the basis for designing learning algorithms (function approximators), as follows: Given that we wish to learn some target function $f : X \rightarrow Y$, or equivalently, $P(Y\arrowvert X)$, we use the training data to learn estimates of $P(X\arrowvert Y)$ and $P(Y)$. New X examples can then be classified using these estimated probability distributions, plus Bayes rule. This type of classifier is called a generative classifier, because we can view the distribution $P(X\arrowvert Y)$ as describing how to generate random instances $X$ conditioned on the target attribute $Y$.

  2. Learning Bayes classifiers typically requires an unrealistic number of training examples (i.e., more than $\arrowvert X\arrowvert$ training examples where $X$ is the instance space) unless some form of prior assumption is made about the form of $P(X\arrowvert Y)$. The Naive Bayes classifier assumes all attributes describing $X$ are conditionally independent given Y. This assumption dramatically reduces the number of parameters that must be estimated to learn the classifier. Naive Bayes is a widely used learning algorithm, for both discrete and continuous $X$.

  3. When $X$ is a vector of discrete-valued attributes, Naive Bayes learning algorithms can be viewed as linear classifiers; that is, every such Naive Bayes classifier corresponds to a hyperplane decision surface in $X$. The same statement holds for Gaussian Naive Bayes classifiers if the variance of each feature is assumed to be independent of the class (i.e., if $\sigma_{ik} = \sigma_{i}$).

  4. Logistic Regression is a function approximation algorithm that uses training data to directly estimate $P(Y\arrowvert X)$, in contrast to Naive Bayes. In this sense, Logistic Regression is often referred to as a discriminative classifier because we can view the distribution $P(Y\arrowvert X)$ as directly discriminating the value of the target value $Y$ for any given instance $X$.

  5. Logistic Regression is a linear classifier over $X$. The linear classifiers produced by Logistic Regression and Gaussian Naive Bayes are identical in the limit as the number of training examples approaches infinity, provided the Naive Bayes assumptions hold. However, if these assumptions do not hold, the Naive Bayes bias will cause it to perform less accurately than Logistic Regression, in the limit. Put another way, Naive Bayes is a learning algorithm with greater bias, but lower variance, than Logistic Regression. If this bias is appropriate given the actual data, Naive Bayes will be preferred. Otherwise, Logistic Regression will be preferred.

  6. We can view function approximation learning algorithms as statistical estimators of functions, or of conditional distributions $P(Y\arrowvert X)$. They estimate $P(Y\arrowvert X)$ from a sample of training data. As with other statistical estimators,it can be useful to characterize learning algorithms by their bias and expected variance, taken over different samples of training data.