/**
* Like the TrainAndTest activity, but uses active training in order to
* choose the documents to add to the training set. The method simulates
* the process that an actual user would go through in order to label
* documents for addition to the training set, then tests performance at
* the specified intervals.
*
* @param string $label class label of the new classifier
* @param string $dataset_name name of the dataset to train and test on
*/
function runActiveTrainAndTest($label, $dataset_name)
{
$this->setDefault('max_train', 200);
$this->logOptions();
$classifier = $this->makeFreshClassifier($label);
$data = $this->loadDataset($dataset_name, $label);
$pages = $data['train'];
$classifier->prepareToLabel();
$classifier->initBuffer($pages);
$end = min($this->options['max_train'], $pages->length);
for ($i = 1; $i <= $end; $i++) {
list($new_doc, $disagreement) = $classifier->findNextDocumentToLabel();
if ($new_doc) {
$key = Classifier::makeKey($new_doc);
$doc_label = $new_doc['TRUE_LABEL'];
$classifier->labelDocument($key, $doc_label);
$classifier->refreshBuffer($pages);
$classifier->computeBufferDensities();
$classifier->train();
}
if ($this->isTestPoint($i, $end)) {
Classifier::setClassifier($classifier);
$this->testClassifier($classifier, $data);
$classifier->prepareToLabel();
}
}
}
/**
* Creates a fresh array from an existing page summary array, and augments
* it with extra data relevant to the labeling interface on the client.
*
* @param array $page original page summary array
* @param float $score classification score (estimated by the Naive Bayes
* text classification algorithm) for $page
* @param float $disagreement disagreement score computed for $page
* @param int $crawl_time index the page came from
* @param string $keywords query supplied to the crawl mix used to find
* $page
* @return array reduced page summary structure containing only the
* information that the client needs to display a summary of the page
*/
function prepareUnlabelledDocument($page, $score, $disagreement, $crawl_time, $keywords)
{
$phrase_model = $this->model("phrase");
// Highlight the query keywords, if any.
$disjunct_phrases = explode("|", $keywords);
$words = array();
foreach ($disjunct_phrases as $disjunct_phrase) {
list($word_struct, $format_words) = $phrase_model->parseWordStructConjunctiveQuery($disjunct_phrase);
$words = array_merge($words, $format_words);
}
$title = $phrase_model->boldKeywords($page[self::TITLE], $words);
$description = $phrase_model->getSnippets(strip_tags($page[self::DESCRIPTION]), $words, 400);
$description = $phrase_model->boldKeywords($description, $words);
$cache_link = "?c=search&a=cache" . "&q=" . urlencode($keywords) . "&arg=" . urlencode($page[self::URL]) . "&its=" . $crawl_time;
/*
Note that the confidence is a transformation of the score that
converts it into a value between 0 and 1, where it's 0 if the score
was exactly 0.5, and increases toward 1 as the score either
increases toward 1 or decreases toward 0.
*/
return array('title' => $title, 'url' => $page[self::URL], 'key' => webencode(Classifier::makeKey($page)), 'cache_link' => $cache_link, 'description' => $description, 'score' => $score, 'positive' => $score >= 0.5 ? 1 : 0, 'confidence' => abs($score - 0.5) / 0.5, 'disagreement' => $disagreement);
}
