Project Report: Self-Labeled Techniques For Semi-Supervised Learning

Project Report
Self-labeled techniques for semi-supervised learning
Mark Laane
Abstract. The aim of this project is to reimplement a selection of semi-super-

vised learning techniques surveyed by Isaac Triguero et al. in paper [1] and to
independently reproduce the reported results. For this, two self-labeled tech-
niques were chosen and implemented in Python programming language: Stand-
ard self-training and Tri-Training. Two well-known UCI datasets were chosen
for testing the implementations: abalone and dermatology. The transductive and
inductive classification accuracy of algorithms was measured on the datasets. The
classification results achieved in this project are much lower than the results
achieved by Triguero et al. indicating an inferior implementation of the algo-
rithms.
1 Background
Semi-Supervised learning is a learning paradigm that joins unsupervised and super-

vised learning paradigms. It can be useful on data that has small amount of labeled data
and a large amount of unlabeled data. In supervised learning only labeled data can be
used, but in semi-supervised learning the unlabeled data is leveraged to provide better
models. It is useful in situations where there is large amount of unlabeled data and
labeling all samples is unviable. [1]
Self-labeled techniques are a subset of semi-supervised classification techniques that
grow the labeled dataset by iteratively labeling some samples from the unlabeled da-
taset. [1]
2 Base classifiers
Some Self-labeled techniques build on supervised learning classifiers by using them

as base classifiers. The base classifiers are used on labeled data to predict the labels of
the unlabeled samples. By incorporating those predictions to the labeled dataset, the
size of the labeled dataset is made larger and base classifier is retrained with this larger
dataset. [1]
In the paper, 4 base classifiers were used with self-labeled methods: K-nearest neigh-
bor (KNN), C4.5, Naive Bayes (NB) and Support vector machines (SVM). They all are
classic and well known classifiers. In this project KNN and NB was used and the C4.5
decision tree classifier was substituted with CART as it is offered in Scikit-learn library
and is similar to C4.5. [2]
3 Implemented algorithms
In this project two self-labeled techniques were implemented: Standard self-training

and Tri-Training. Per classification proposed by Isaac Triguero et al. both are single-
view, single learning and incremental learning algorithms. This means that the feature
space does not have to be split into independent views (single-view), the labeled dataset
is only incrementally enlarged from unlabeled dataset (incremental) and identical learn-
ers or base classifiers are used (single-learning). The algorithms differ in the number of
classifiers used: Self training is a Single-classifier algorithm and uses only one instance
of base classifier. Tri-Training is a multi-classifier and uses three instances of a given
base classifier.
3.1 Standard Self-Training
The implementation of Standard Self-Training is based on description of the algo-

rithm in paper [3]. Training a Standard Self-Training classifier is an iterative process
where the base classifier is retrained multiple times. The base classifier is initially
trained with initial labeled samples. Then, it is used for labelling the unlabeled samples
and the most confident predictions are added to the labeled set. Then the classifier is
retrained with this extended labeled set. The training and labeling process is repeated
until all the training samples are labeled or maximum number of iterations is reached.
The algorithm can be summarized as follows:
Notation:
C – base classifier used
D – initial training dataset
L – labeled training samples, L ⊆ D
U – unlabeled training samples, U ⊆ D
1. Train C on L
2. Use C to predict labels on U and select most confident predictions
3. Remove the most confident predictions from U and add them to L with the pre-
dicted labels
4. Repeat steps 1-4 until U is empty or the maximum number of iterations is
reached. (40 in our implementation)
The final C can then be used to classify unseen data.
3.2 Tri-Training
The implementation of Tri-Training is based on description of the algorithm in paper
[4]. In Tri-Training, three base classifiers are trained on randomly subsampled sets of
the labeled data. Then each of them will be iteratively trained by also considering the
labels that two other base classifiers have predicted. The predictions from the other two
base classifiers inherently introduce noise by mislabeling some samples. To set an up-
per bound for classification noise rate during the training process a restriction is set in
place. Chosen base classifier will only be retrained if following equation holds
𝑒𝑡 |𝐿𝑡−1 |
0 < 𝑒 𝑡−1 < |𝐿𝑡 |
<1 (1)
where |𝐿𝑡 | is the number of samples that can be added in the t-th round (agreed by
the other two base classifiers) and 𝑒𝑡 is the upper bound of classification error rate of
those labeled samples 𝐿𝑡 in t-th round. If |𝐿𝑡 | is too large to satisfy equation 1, then the
number of samples can be reduced by randomly subsampling the set so it would contain
𝑒 𝑡−1 |𝐿𝑡−1 |
s samples, where 𝑠 = ⌈ − 1⌉. The algorithm continues as long there is at least
𝑒𝑡
one 𝐶𝑖, where equation 1 can be satisfied.
The algorithm can be summarized as follows:
1. Randomly subsample L to form three sets of labeled data 𝐿1 , 𝐿2 and 𝐿3

2. Train three instances of C, each 𝐶𝑖 on 𝐿𝑖
3. For each 𝐶𝑖:
i. From U take samples where other two classifiers agree 𝐿𝑎𝑔𝑟𝑒𝑒 (two other clas-
sifiers predict the same label)
ii. If restrictions on the added classification noise are met, retrain 𝐶𝑖 with the
𝐿𝑖 ∪ 𝐿𝑎𝑔𝑟𝑒𝑒
4. Repeat step 3 until none of the classifiers are retrained any more.
The three trained classifiers can then be used for classification by using majority
voting.
4 Experiments
10-fold cross-validation procedure was used to validate the classification algorithms.

The 10% in each split was kept for testing and the rest used for training. To simulate
dataset with unlabeled samples, each training dataset was further split in two: unlabeled
and labeled samples. From unlabeled samples, the labels were removed. Four labelling-
ratios were tested: 10%, 20%, 30% and 40%
Two accuracy scores were measured in the experiments: accuracy of transductive
learning and accuracy of inductive learning. Transductive accuracy describes the clas-
sifier’s ability to correctly predict the labels of unlabeled samples used during the train-
ing process. Inductive accuracy describes the classifiers ability to correctly predict la-
bels of unseen data – test data that is withhold and not been used for training.
5 Results
Each self-labeled algorithm was tested with each base classifier the validation was
performed on each dataset. The results can be seen in the table in Appendix 2 and a
selection of confusion matrixes in Appendix 1.
As expected, raising labelling rate also raises the transductive and test accuracy in
most cases. Abalone dataset proves to be more difficult to classify and extra labels do
not always improve the results. There is also an anomalous result with unknown source,
where of Tri-Training used with Naïve Bayes has the highest accuracy on 10% labelling
ratio. Both algorithms perform worst with Naïve Bayes as a base classifier and best
with CART base classifier. Triguero et al also reported good performance with another
similar decision tree algorithm C4.5. On dermatology dataset, both Self-Training with
CART and Tri-Training with CART achieved best results reaching towards 90% of
accuracy.
Generally the classification accuracy results outputted by the implemented algo-
rithms are lower than the results achieved by Triguero et al. indicating a inferior per-
formance of the implementations. Also, high standard deviation indicates poor perfor-
mance, as the results seems to strongly depend on the fold that the data is being tested
on.
6 Conclusions
In this project two self-labeled techniques - Standard self-training and Tri-Training

were implemented and tested on two datasets: abalone and dermatology. The classifi-
cation accuracy of the implemented algorithms is lower than the results achieved by
Triguero et al. This could be explained by poor implementation. Therefore, further
work should ensue to locate problematic parts of the implementations and fix them to
improve prediction accuracy.
References
[1] I. Triguero, S. García and F. Herrera, "Self-labeled techniques for semi-

supervised learning," Knowledge and Information Systems, vol. 42, no. 2, pp.
245--284, 2015.
[2] "Scikit-learn Decision Trees," [Online]. Available: http://scikit-
learn.org/stable/modules/tree.html#tree-algorithms-id3-c4-5-c5-0-and-cart.
[3] Y. D, "Unsupervised word sense disambiguation rivaling supervised
methods," in Proceedings of the 33rd annual meeting of the association for
computational linguistics, 1995.
[4] L. M. Zhou ZH, "Tri-training: exploiting unlabeled data using three
classifiers," IEEE, 2005).
Appendix 1: Confusion matrixes
Fig. 1. Self-Training (CART), abalone, 40% labeling ratio
Fig. 2. Self-Training (NB), abalone, 40% labeling ratio

Fig. 3. Self-Training (KNN), dermatology, 40% labeling ratio
Fig. 4. TriTraining (NB), dermatology, 40% labeling ratio

Appendix 2: Testing results
Labeling Test Trans Trans

classifier dataset Test std
rate mean mean std
0 TriTraining (KNN) abalone 0.1 0.159610 0.093613 0.186119 0.063314
4 TriTraining (KNN) dermatology 0.1 0.478078 0.061992 0.443366 0.035022
8 TriTraining (NB) abalone 0.1 0.090718 0.035779 0.044423 0.044664
12 TriTraining (NB) dermatology 0.1 0.401126 0.300181 0.406665 0.275658
16 TriTraining (CART) abalone 0.1 0.167789 0.080768 0.181449 0.061071
20 TriTraining (CART) dermatology 0.1 0.715766 0.152987 0.769580 0.041982
24 Self-Training (KNN) abalone 0.1 0.242754 0.084157 0.237461 0.007264
Labeling Test Trans Trans
classifier dataset Test std
rate mean mean std
28 Self-Training (KNN) dermatology 0.1 0.549399 0.072864 0.506809 0.052777
32 Self-Training (NB) abalone 0.1 0.075651 0.023860 0.065025 0.004766
36 Self-Training (NB) dermatology 0.1 0.676877 0.250911 0.645830 0.256829
40 Self-Training (CART) abalone 0.1 0.208044 0.057041 0.210919 0.012029
44 Self-Training (CART) dermatology 0.1 0.778153 0.107618 0.790551 0.068355

Project Report: Self-Labeled Techniques For Semi-Supervised Learning

Uploaded by

Copyright:

Available Formats

Project Report: Self-Labeled Techniques For Semi-Supervised Learning

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Project Report: Self-Labeled Techniques For Semi-Supervised Learning

Uploaded by

Copyright:

Available Formats

Project Report

Self-labeled techniques for semi-supervised learning

Abstract. The aim of this project is to reimplement a selection of semi-super-

Semi-Supervised learning is a learning paradigm that joins unsupervised and super-

Some Self-labeled techniques build on supervised learning classifiers by using them

In this project two self-labeled techniques were implemented: Standard self-training

3.1 Standard Self-Training

The implementation of Standard Self-Training is based on description of the algo-

1. Randomly subsample L to form three sets of labeled data 𝐿1 , 𝐿2 and 𝐿3

10-fold cross-validation procedure was used to validate the classification algorithms.

In this project two self-labeled techniques - Standard self-training and Tri-Training

[1] I. Triguero, S. García and F. Herrera, "Self-labeled techniques for semi-

Fig. 1. Self-Training (CART), abalone, 40% labeling ratio

Fig. 2. Self-Training (NB), abalone, 40% labeling ratio

Fig. 4. TriTraining (NB), dermatology, 40% labeling ratio

Labeling Test Trans Trans

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.