Data Mining_ Concepts and Techniques - Jiawei Han [265]
Figure 9.18 The pool-based active learning cycle. Source: From Settles [Set10], Burr Settles Computer Sciences Technical Report 1648, University of Wisconsin–Madison; used with permission.
Most of the active learning research focuses on how to choose the data tuples to be queried. Several frameworks have been proposed. Uncertainty sampling is the most common, where the active learner chooses to query the tuples which it is the least certain how to label. Other strategies work to reduce the version space, that is, the subset of all hypotheses that are consistent with the observed training tuples. Alternatively, we may follow a decision-theoretic approach that estimates expected error reduction. This selects tuples that would result in the greatest reduction in the total number of incorrect predictions such as by reducing the expected entropy over U. This latter approach tends to be more computationally expensive.
9.7.4. Transfer Learning
Suppose that AllElectronics has collected a number of customer reviews on a product such as a brand of camera. The classification task is to automatically label the reviews as either positive or negative. This task is known as sentiment classification. We could examine each review and annotate it by adding a positive or negative class label. The labeled reviews can then be used to train and test a classifier to label future reviews of the product as either positive or negative. The manual effort involved in annotating the review data can be expensive and time consuming.
Suppose that AllElectronics has customer reviews for other products as well such as TVs. The distribution of review data for different types of products can vary greatly. We cannot assume that the TV-review data will have the same distribution as the camera-review data; thus we must build a separate classification model for the TV-review data. Examining and labeling the TV-review data to form a training set will require a lot of effort. In fact, we would need to label a large amount of the data to train the review-classification models for each product. It would be nice if we could adapt an existing classification model (e.g., the one we built for cameras) to help learn a classification model for TVs. Such knowledge transfer would reduce the need to annotate a large amount of data, resulting in cost and time savings. This is the essence behind transfer learning.
Transfer learning aims to extract the knowledge from one or more source tasks and apply the knowledge to a target task. In our example, the source task is the classification of camera reviews, and the target task is the classification of TV reviews. Figure 9.19 illustrates a comparison between traditional learning methods and transfer learning. Traditional learning methods build a new classifier for each new classification task, based on available class-labeled training and test data. Transfer learning algorithms apply knowledge about source tasks when building a classifier for a new (target) task. Construction of the resulting classifier requires fewer training data and less training time. Traditional learning algorithms assume that the training data and test data are drawn from the same distribution and the same feature space. Thus, if the distribution changes, such methods need to rebuild the models from scratch.
Figure 9.19 Transfer learning versus traditional learning. (a) Traditional learning methods build a new classifier from scratch for each classification task. (b) Transfer learning applies knowledge from a source classifier to simplify the construction of a classifier for a new, target task. Source: From Pan and Yang [PY10]; used with permission.
Transfer learning allows the