Analysis Of Lazy Super Parent Tree Augmented Naive Bayes
In this section, we propose a heuristic, called Lazy Super Parent Tree Augmented Naive Bayes (LSPTAN) that seeks to solve the problems discussed above, enabling the application of a semi-Naive Bayes techniques in large ADC tasks. Thus, we can evaluate whether the premise of independence among attributes, assumed by Naive Bayes, impacts effectiveness in large ADC tasks, an open research problem.\looseness=-1
The Lazy Super Parent TAN (LSPTAN) heuristic is a postergated version of the SP-TAN that constructs a Tree Augmented Naive Bayes for each test example. Attributes dependencies are generated based on information from the example that is being classified. To build a lazy version of SP-TAN we adapted the method of evaluation and the selection of candidates for Super Parent and Favorite Children.\looseness=-1
The SP-TAN algorithm exploits accuracy to select a candidate to Super Parent ($a_{sp}$). In our strategy, we select the candidate $a_{sp}$ whose classification model generates …show more content…
Therefore, LSPTAN builds a simpler network than the SP-TAN. We select only the best Super Parent to a test document. But there is no limitation on the choice of the Favorite Children. Thus, all the children attributes that increment the probability that the document belongs to a class, are included in the classification model.\looseness=-1
The LSPTAN heuristic initially builds the model based on Naive Bayes and initializes a set of orphans $O$, inserting into $O$ all the terms of the vocabulary. Then, for each test document, the technique evaluates each term as a Super Parent ($a_{sp}$) and, at the end, it selects as $a_{sp}$ the term that has the highest probability $P(c_i | d_t, a_{sp})$. Thus, the $P(c_i | d_t, a_{sp})$ for a $a_{sp}$ is defined as by the Equation~\ref{eq::lsptan}, where $f$ is the frequency of a term in the document $d_t$.\looseness=-1
\begin{center}
\begin{equation}
The Lazy Super Parent TAN (LSPTAN) heuristic is a postergated version of the SP-TAN that constructs a Tree Augmented Naive Bayes for each test example. Attributes dependencies are generated based on information from the example that is being classified. To build a lazy version of SP-TAN we adapted the method of evaluation and the selection of candidates for Super Parent and Favorite Children.\looseness=-1
The SP-TAN algorithm exploits accuracy to select a candidate to Super Parent ($a_{sp}$). In our strategy, we select the candidate $a_{sp}$ whose classification model generates …show more content…
Therefore, LSPTAN builds a simpler network than the SP-TAN. We select only the best Super Parent to a test document. But there is no limitation on the choice of the Favorite Children. Thus, all the children attributes that increment the probability that the document belongs to a class, are included in the classification model.\looseness=-1
The LSPTAN heuristic initially builds the model based on Naive Bayes and initializes a set of orphans $O$, inserting into $O$ all the terms of the vocabulary. Then, for each test document, the technique evaluates each term as a Super Parent ($a_{sp}$) and, at the end, it selects as $a_{sp}$ the term that has the highest probability $P(c_i | d_t, a_{sp})$. Thus, the $P(c_i | d_t, a_{sp})$ for a $a_{sp}$ is defined as by the Equation~\ref{eq::lsptan}, where $f$ is the frequency of a term in the document $d_t$.\looseness=-1
\begin{center}
\begin{equation}