References in periodicals archive ?
We will focus on penalizing existing local score metrics with our sentiment-dependent scoring function for the BN classifiers, hence the SDBN proposed in this paper.
The proposed SDBN is created from a sentiment dependent score table (SDST) similar to the conventional CPT which contains network parameters from the data.
We conducted set of experiments using the proposed SDBN algorithm on 8 different product review domains.
We performed a grid search with 10-fold cross-validation for the three algorithms and observed that both SDBN and Baseline-BN gave best accuracies using SimpleEstimator with [alpha] = 0.
We evaluated the performance of the SDBN with reduced attribute sets since attribute selection tends to improve BN's accuracy [16].
As emphasised in Table 1, we observed the proposed SDBN to have improved and sometimes comparable performance with the baseline classifiers.
We also performed experiment using the SDBN for top-10, top-20, top-30, top-50, and top-100 attributes alone as shown in Table 2.
As shown in Table 3, we also performed experiment by using SDBN with other scoring functions reported in Section 3.
The proposed SDBN uses a multi-class approach to compute sentiment dependencies between pairs of variables by using a joint probability from different sentiment evidences.