Extending Semi-supervised Learning Methods for Inductive Transfer Learning

Shi, Yuan; Lan, Zhenzhong; Liu, Wei; Bi, Wei

doi:10.1109/icdm.2009.75

Cited by 31 publications

(47 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The splice data set has 3190 examples with 60 attributes in each example and one binary class label. We adopt the same strategy in [20,25] to divide the data sets. The mushroom data set is split into the source task and the target task based on the attribute stalk-shape, the source task contains examples whose stalk-shape are tapering and the examples in the target task have stalkshape of enlarging.…”

Section: Experiments 51 Experimental Settingsmentioning

confidence: 99%

“…The four real data sets used in the experiments are monk, mushroom, splice and krvskp (http://archive.ics.uci.edu/ml/). Although these data sets are for single-task learning, we adopted a preprocessing method [20,25] on them to fit the transfer learning scenario except the monk data sets. We exclude more complex text datasets because our rulebased classifier is not appropriate for them.…”

Section: Experiments 51 Experimental Settingsmentioning

confidence: 99%

“…The challenge is to find relevant and valuable information from the source domain and transfer it to the target domain, to obtain a satisfactory accuracy without excessive cost. Current learning methods for inductive transfer learning can be mainly divided into instance-based transfer [3,20,25] and feature-based transfer [1,22]. The former approach essentially assumes a distance function for selecting the right instance in the source domain to help classifying an instance in the target domain.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Feature-based Inductive Transfer Learning through Minimum Encoding

Shao

Suzuki²

2011

Proceedings of the 2011 SIAM International Conference on Data Mining

View full text Add to dashboard Cite

This paper proposes an Extended Minimum Description Length Principle (EMDLP) for feature-based inductive transfer learning, in which both the source and the target data sets contain class labels and relevant features are transferred from the source domain to the target one. Despite numerous works on this topic, few of them have a solid theoretical framework and are parameter-free. Our EMDLP overcomes these flaws and allows us to evaluate the inferiority of the results of transfer learning with the add-sum of the code lengths of five components: the corresponding two hypotheses, the two data sets with the help of the hypotheses, and the set of the transferred features. We design a code book to build the connections between the source and the target tasks. Extensive experiments using both real and artificial data sets show that EMDLP is robust against noise and performs better on the classification accuracy than the state-of-the-art methods.

show abstract

Section: Experiments 51 Experimental Settingsmentioning

confidence: 99%

Section: Experiments 51 Experimental Settingsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Feature-based Inductive Transfer Learning through Minimum Encoding

Shao

Suzuki²

2011

Proceedings of the 2011 SIAM International Conference on Data Mining

View full text Add to dashboard Cite

show abstract

“…First, we measure the distance between instances by Transfer Learning [6]. The goal of Transfer Learning is to move the knowledge learned from one environment to deal with the problem in a new environment.…”

Section: B the Weight Of Training Datamentioning

confidence: 99%

“…Therefore, this paper firstly gives the corresponding initial weights to the cross-project training instances through the Transfer Learning [6]. Then give the final weight combines with number of faults information and calculate the prior probabilities of the presence or absence of defects.…”

Section: Introductionmentioning

confidence: 99%

Is the Number of Faults Helpful for Cross-Company Defect Prediction?

Jing¹,

Zhang²,

Jin³

2017

International Conferences on Software Engineering and Knowledge Engineering

View full text Add to dashboard Cite

Abstract-In the field of Cross-company Defect Prediction (CCDP), how to deal with the data to make it more accurately predict the cross-company software defects is the focus problem we need to consider. Now a mainstream idea is to determine the weight of the training data based on the similarity between the training data and the test set, and then build the model on the basis of these weighted data. However, sometimes, when we deal with some problems with imbalance class, directly using the weight calculated above may lead to errors. Because a large number of non-defective instance's weight accumulation will lead to the defective instance's weight has a little impact on the final result. This is why we need to consider the addition of the number of defects. Considering the number of defects will effectively eliminate the impact of a large number of non-defective instance's weight accumulation. Therefore, we propose a Transfer-learning Naï ve Bayes model considering the number of defective information(NTNB). The method consists of two major stages: weight the data and build the prediction model. In the stage of weighting the data, we not only consider the similarity between the data but also consider the number of defects to get the final weights for data. And we conducted a set of comparative experiments on six open crosscompany datasets. The results show that considering the number of defects information can effectively avoid some defective instance is misjudged as non-defective instance, and improve the accuracy of prediction in some unbalanced problems.

show abstract