Paper Reading: A Brief Introduction to Weakly Supervised Learning

incomplete, 想利用未标注数据帮助训练

inexact, 笼统的数据标注，如垃圾邮件分类

inaccurate supervision， 带噪声的数据，如众包

Incomplete supervision

training data set (D={(x_1,y_1),cdots,(x_l,y_l),x_{l+1},cdots,x_m})

active learning (with human intervention)

​ the labeling cost only depends on the number of queries

1. informativeness: an unlabeled instance helps reduce the uncertainty of a statistical model.

   1.1 Uncertainty sampling a single learner, with the least confidence

   1.2 query-by-committee multiple learners, disagree to most

2. representativeness : an instance helps represent the structure of input patterns

   2.1 aim to exploit the cluster structure of unlabeled data

semi-supervised learning (no human intervention is assumed)

​ Here, although the unlabeled data points are not explicitly with label information, they implicitly convey some information about data distribution which can be helpful for predictive modelling.

​ two basic assumptions: the cluster assumption (data have inherent cluster structure) and the manifold assumption (data lie on a manifold).

1. generative methods

   ​ labels of unlabeled instances can be treated as missing values of model parameters, and estimated by approaches such as the EM .

   ​ To get good performance, one usually needs domain knowledge to determine adequate generative model.

2. graph based methods

   ​ the performance will heavily depends on how the graphis constructed.

3. low-density seperation methods

   ​ It is evident that S3VMs try to identify a classification boundary which goes across the less dense region while keeping the labeled data correctly classified.

4. disagreement-based methods

   ​ generate multiple learners and let them collaborate to exploit unlabeled data.

Inexact Supervision

​ Multi-instance learning: predict the labels for unseen bags((X_i) is a positive bag, if there exists (x_{ip}) which is positive, while p is unknown).

Inaccurate Supervision

​ For machine learning, crowdsourcing is commonly used as a cost-saving way to collect labels for training data.