A Collaborative Deep and Shallow Semisupervised Learning Framework for Mobile App Classification

Table 1

The summary of the existing semisupervised learning techniques.

Scheme	Principle	Drawbacks	References
Generative semisupervised learning	It assumes that both labeled and unlabeled samples are generated from the same parametric model; then, it treats the labels of the unlabeled samples as missing values of the model parameters	When the parametric model assumption is incorrect, fitting the model using unlabeled samples would result in performance degradation	[21, 27, 28]
Graph-based semisupervised learning	It constructs a graph whose nodes are samples (including both labeled and unlabeled samples) and edges reflect relations between nodes (e.g., feature similarity); then, the labels are propagated by exploiting the connective characteristics	First, it suffers from poor scalability; second, it is difficult to build the relations between samples when the features are sophisticated	[29–31]
Disagreement-based semisupervised learning	Multiple base learners are initially trained on labeled samples, and then they learn from each other by exploiting the disagreement among them on unlabeled samples	How to guarantee the diversity between base learners is an open problem	[25, 32, 33]