Citation Context

...ssification benchmark during the late nineties was a Reuters collection called Reuters-21578 (based on Reuters-22173) with 21578 documents, that had to be classified in about 100 different categories =-=[1]-=-. This benchmark is still used currently to compare the performance of different algorithms but the challenges now lie in moving towards larger scale document classification. In 2002, Reuters released...

Citation Context

...e systems is a time consuming and expensive task, making ITS unfeasible except for very well defined domains. One way of improving the usability of these systems is to use machine learning algorithms =-=[7,17]-=-. This releases the designer from having to write explicit rules and provides the system with novel ways of using content. Instead of using rules, these algorithms learn from the examples found in tra...

Citation Context

... been done to find the efficient algorithm which can classify these documents automatically. These algorithms use machine learning concept such as neural networks, naive Bayes and k-nearest neighbors =-=[1]-=-, [2], [3] in order to find the general model of each Web’s category based on a set of labeled documents. Unfortunately, the efficient algorithms still need a high portion of labeled documents to cons...

Citation Context

...ee machine learning methods that have provided some of the best performances in other classification tasks [16]: Support Vector Machines (SVM) [12] [6], Naïve Bayes (NB) [7], and Neural Networks (NN) =-=[5]-=- [4]. It is not possible to describe them thoroughly in this article, so we will only summarize those issues that might be required to reproduce the results. Our Naïve Bayes and SVM implementations ar...

Citation Context

...erable previous works on hierarchical classification of textual documents using many different approaches. Several authors have applied neural networks to supervised text classification. Calvo et al. =-=[2]-=- approach a backpropagation algorithm. In the input layer they use as many units (neurons) as the features considered and there is an output unit for each possible class. Ruiz et al. [12] analyse the ...

Citation Context

...re projected in a k-dimensional (where k < n) subspace. Typically, k takes on values between 50 and 150 with smaller values of k useful for clustering and large values of k better for classification (=-=Calvo & Ceccatto, 2000-=-; Osinski et al., 2004; Efron, 2002; SAS Institute, 2004). Based on previous studies, k for this study was tested between 50 and 250 in order to validate the number of SVD dimensions that proved most ...

Citation Context

...mensionality reduction techniques (a.k.a. feature selection), finding that χ 2 was among the most effective ones. However, they did not specify what averaging of χ 2 they followed. In a previous work =-=[1]-=- we have used the χ 2 statistic with the χ 2 avg(t) averaging scheme.sAfter the process of dimensionality reduction, the documents are represented by vectors and different classification techniques ca...

Citation Context

...cores from χ 2 and MI. Additionally, they used Averaging (Avg) for DF and IG. Calvo and Ceccatto proposed the usage of Weighted Maximum (WMax), where features are weighted by the category probability(=-=Calvo & Ceccatto 2000-=-). Equations 1, 2, 3, and 4 provide the mathematical definitions of Avg, Max, WAvg, and WMax respectively. FAvg(wk) = � M i=1 f(wk, ci) M (1) FMax(wk) = M max i=1 {f(wk, ci)} (2) where f(wk, ci) is th...

Citation Context

...odules. The framework has implementations of kNearest-Neighbor (kNN), Naïve Bayes (NB), Support Vector Machine (SVM), and Decision Tree (DT) classifiers [12, 10]. Other methods such as Neural Network=-= [3, 2]-=- classifiers are under development. The framework architecture allows extensions to be built by subclassing its main classes [11]. Class inheritance contributes to code reuse and quality. In this proj...