Quantification and semi-supervised classification methods for handling changes in class distribution

Question

1. What have the authors contributed in "Quantification and semi-supervised classification methods for handling changes in class distribution" ?

2. What future works have the authors mentioned in the paper "Quantification and semi-supervised classification methods for handling changes in class distribution" ?

3. What is the primary reason why ROC analysis has gained such prominence in the data mining community?

4. How many examples can be generated without duplicating them?

Accepted Answer

In this paper the authors address the problem where the class distribution changes and only unlabeled examples are available from the new distribution.. The authors design and evaluate a number of methods for coping with this problem and compare the performance of these methods.. The authors also introduce a hybrid method that utilizes both quantification and semi-supervised learning.

Accepted Answer

The authors discuss improvements to the SSL methods in their discussion of future work toward the end of this section.. The failure of the SSL-based methods is also notable and should help guide future research.. One possibility the authors are investigating is to use a classifier that accepts class membership probabilities in the training phase, so that the uncertain predictions generated by semi-supervised learning can be factored into the learning process.. A more ambitious setting that the authors intend to study is where the concept drift is of a much less restrictive form, such that the actual concept can change over time, rather than just its class distribution.

Accepted Answer

In fact, this desire for classifiers that exhibit robust behavior over wide ranges of class distributions and misclassification costs is the primary reason that ROC analysis [10] has gained such prominence in the data mining community.

Accepted Answer

In order to generate the desired class distributions without duplicating any examples, the size of the partitions must be limited.

Accepted Answer

Since the original data set contains 11,673 (23.9% of 48,482) positive and 37,168 negative examples, these partitions can be generated without duplicating examples.

Accepted Answer

A third way of compensating for differing class distributions is to sample from the original distribution so that its class distribution matches the estimated class distribution of the new data.

Accepted Answer

A more ambitious setting that the authors intend to study is where the concept drift is of a much less restrictive form, such that the actual concept can change over time, rather than just its class distribution.

Accepted Answer

The maximum number of positive or negative examples required is 11,656 (2,350 + 2 • 4,653), which for the positives examples occurs when Pr= 99% and for the negative examples when Pr=1%.

Accepted Answer

CDE-based methods outperform SSL-based methods because, as the authors shall discuss in Section 5, class distribution estimation (i.e., quantification) is fundamentally an easier task than classification and thus the CDE-based methods introduce less uncertainty than the SSL-based methods, which require us to classify examples from the new distribution before a classifier can be retrained.

Accepted Answer

The class distribution estimation based methods generally performed the best and when accuracy must be maximized the CDEAC method is the best choice based on its performance, computational requirements (i.e., only a single iteration) and the fact that there are no parameters, such as the number of iterations, to set.

Accepted Answer

The performance curves for the other methods, including the Oracle method (not shown) exhibit a ―U‖ shape with a minimum near a positive class rate of 50%.

Accepted Answer

The Oracle method provides a potential upper bound for achievablePermission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page.

Accepted Answer

Note that it still requires pr, the unadjusted estimate of NEWCD, which is calculated in the same manner as it was calculated in the first iteration of CDE-Iterate—a classifier isbuilt from ORIGlabel, used to classify NEWunlabel, and pr is calculated from the predicted class labels.

Accepted Answer

The results thus far for the SSL-based methods have not been very promising, probably due to the fact that, as just discussed, it is easier to accurately estimate the class distribution of the unlabeled examples than to classify them.

Quantification and semi-supervised classification methods for handling changes in class distribution

Chat with Paper

AI Agents for this Paper

Most frequently asked questions

1. What have the authors contributed in "Quantification and semi-supervised classification methods for handling changes in class distribution" ?

2. What future works have the authors mentioned in the paper "Quantification and semi-supervised classification methods for handling changes in class distribution" ?

3. What is the primary reason why ROC analysis has gained such prominence in the data mining community?

4. How many examples can be generated without duplicating them?

5. How many examples can be generated without duplicating examples?

6. What is the method for calculating class distributions?

7. What is the definition of a more ambitious setting?

8. How many examples are required for the positives and negatives data sets?

9. Why do CDE-based methods outperform SSL-based methods?

10. What is the method for estimating class distributions?

11. What is the performance curve for the CDE-AC method?

12. What is the upper bound for the Oracle method?

13. What is the cost ratio for the CDE-Iterate algorithm?

14. What is the reason why the results for the SSL-based methods have not been very promising?

Figures

Citations

A Review on Quantification Learning

Class distribution estimation based on the Hellinger distance

From classification to quantification in tweet sentiment analysis

Quantification-oriented learning based on reliable classifiers

Cross project defect prediction using class distribution estimation and oversampling

References

SMOTE: synthetic minority over-sampling technique

C4.5: Programs for Machine Learning

Data Mining: Practical Machine Learning Tools and Techniques

UCI Machine Learning Repository

SMOTE: Synthetic Minority Over-sampling Technique

Related Papers (5)

Quantifying counts and costs via classification

Quantification via Probability Estimators

Counting positives accurately despite inaccurate classification

Adjusting the outputs of a classifier to new a priori probabilities: a simple procedure

Estimating Class Priors in Domain Adaptation for Word Sense Disambiguation