Tri-Training Based Learning from Positive and Unlabeled Data

Zhang, Bangzuo; Zuo, Wangmeng

doi:10.1109/isip.2008.69

Cited by 6 publications

(3 citation statements)

References 10 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The simplest assumption is SCAR (Selected Completely at Random Assumption), according to which the propensity score function, i.e. the probability of a labeling a positive observation, is constant [2,1,9,10,11]. Under the SCAR assumption, a possible approach is to estimate label frequency [12,13,14,15] and then use it to scale the posterior probabilities obtained from the naive method or, alternatively, optimize weighted empirical risk function with weights depending on the label frequency [1,16].…”

Section: Introductionmentioning

confidence: 99%

Beyond the Selected Completely at Random Assumption for Learning from Positive and Unlabeled Data

Bekker

Robberechts

Davis

2020

Lecture Notes in Computer Science

View full text Add to dashboard Cite

Most positive and unlabeled data is subject to selection biases. The labeled examples can, for example, be selected from the positive set because they are easier to obtain or more obviously positive. This paper investigates how learning can be enabled in this setting. We propose and theoretically analyze an empirical-risk-based method for incorporating the labeling mechanism. Additionally, we investigate under which assumptions learning is possible when the labeling mechanism is not fully understood and propose a practical method to enable this. Our empirical analysis supports the theoretical results and shows that taking into account the possibility of a selection bias, even when the labeling mechanism is unknown, improves the trained classifiers.

show abstract

Section: Introductionmentioning

confidence: 99%

Beyond the Selected Completely at Random Assumption for Learning from Positive and Unlabeled Data

Bekker

Robberechts

Davis

2020

Lecture Notes in Computer Science

View full text Add to dashboard Cite

show abstract

“…Finally, select a good classifier from the set. For this step, we trained three classifiers: Support Vector Machine (SVM), J48 Decision Tree, and K-Nearest Neighbors (KNN) simultaneously using a modified version of the Co-Training algorithm [44] called Tri-Training [45]. In Tri-training, the learning process of each classifier is greatly influenced by the other two classifiers, akin to a majority-vote scheme.…”

Section: Pu-learningmentioning

confidence: 99%

“…The Tri-training algorithm [47,45] incorporates the E step from the EM algorithm and a modified version of the idea of Co-Training in order to accommodate a third classifier. In the original algorithm, three SVM classifiers were used.…”

Section: Tri-trainingmentioning

confidence: 99%

Drug-drug interactions: A machine learning approach

Mensah¹

View full text Add to dashboard Cite

Automatic detection of drug-drug interaction (DDI) is a difficult problem in pharmacosurveillance. Recent practice for in vitro and in vivo pharmacokinetic drug-drug interaction studies have been based on carefully selected drug characteristics such as their pharmacological effects, and on drug-target networks, in order to identify and comprehend anomalies in a drug's biochemical function upon co-administration. In this work, we present a novel DDI prediction framework that combines several drugattribute similarity measures to construct a feature space from which we train three machine learning algorithms: Support Vector Machine (SVM), J48 Decision Tree and K-Nearest Neighbor (KNN) using a partially supervised classification algorithm called Positive Unlabeled Learning (PU-Learning) tailored specifically to suit our framework. In summary, we extracted 1,300 U.S. Food and Drug Administration-approved pharmaceutical drugs and paired them to create 1,688,700 feature vectors. Out of 397 drug-pairs known to interact prior to our experiments, our system was able to correctly identify 80% of them and from the remaining 1,688,303 pairs for which no interaction had been determined, we were able to predict 181 potential DDIs with confidence levels greater than 97%. The latter is a set of DDIs unrecognized by our source of ground truth at the time of study. Evaluation of the effectiveness of our system involved querying the U.S. Food and Drug Administration's Adverse Effect Reporting System (AERS) database for cases involving drug-pairs used in this study. The results returned from the query listed incidents reported for a number of patients, some of whom had experienced severe adverse reactions leading to outcomes such as prolonged hospitalization, diminished medicinal effect of one or more drugs, and in some cases, death. ACKNOWLEDGMENTS This thesis would not have been possible without the support and guidance from a number of individuals who have made significant contributions, whether directly or indirectly, towards the preparation and successful completion of this work.

show abstract

Distributed Semi-Supervised Learning with Positive and Unlabeled Data

Liu

Zhang

2022

2022 3rd International Conference on Big Data, Artificial Intelligence and Internet of Things Engineering (ICBAIE)

View full text Add to dashboard Cite

Tri-Training Based Learning from Positive and Unlabeled Data

Cited by 6 publications

References 10 publications

Beyond the Selected Completely at Random Assumption for Learning from Positive and Unlabeled Data

Beyond the Selected Completely at Random Assumption for Learning from Positive and Unlabeled Data

Drug-drug interactions: A machine learning approach

Distributed Semi-Supervised Learning with Positive and Unlabeled Data

Contact Info

Product

Resources

About