|
Classifier | Underlying methodology | Classifier applicability | Nature of prediction/label class | Advantage(s) | Disadvantage(s) |
|
|
Naïve Bayes [35] | Bayes theorem | Classification | Categorical | Less parameter tuning, less data learning requirements, computationally fast | Conditional independence between attributes |
Decision trees [36] | Iterative Dichotomiser 3 (ID3) | Classification, regression | Categorical, continuous | Simple to interpret, shows higher accuracy | Target attribute must have discrete values; dataset must not have complex and many attributes (i.e., imbalance); uses greedy approach for generating DTs; prone to overfitting |
Random forest [12] | Aggregation of (decision) trees using bagging with C4.5 algorithm | Classification, regression | Categorical, continuous | Not susceptible to overfitting, reduces error rate while generating DTs | Generates parallel DTs, computationally slow on large and complex datasets |
Gradient boosted trees [37, 38] | Adaptive boosting using C4.5 algorithm | Classification, regression | Categorical, continuous | Boosting reduces error by reducing bias and to some extent variance sequential tree generation with improved learning in each iteration | Uses shallow weak learner trees, computationally faster than RF, harder parameter tuning |
Deep learning [13] | Convolutional neural networks | Classification, regression | Categorical, continuous | Higher accuracy sometimes exceeds human-level performance; DL algorithms scale with data; CNNs require relatively little preprocessing | Requires large amounts of labeled data and substantial computing power |
|