Temporal Aspects in Stream Active Learning
Daniel Kottke, Georg Krempl, Myra Spiliopoulou
Knowledge Management and Discovery Lab
Otto-von-Guericke-University Magdeburg, Germany
daniel.kottke@ovgu.de
www.daniel.kottke.eu/talks/2016_DAGSTAT
Pool vs. Stream Learning
- Spatial components is extended with temporal information
- Classification model might change (Drift)
- Fast, endless instance generation
- Efficient (on-line) algorithms required
- Applications:
- Data from twitter, sensors, bank transactions
Spatial and temporal selection
Spatial and temporal selection
Spatial and temporal selection
Measure spatial usefulness
Spatial and temporal selection
Spatial and temporal selection
Choose the most useful instances based on the spatial value
Spatial Selection Methods
Uncertainty Sampling [1]
Chooses instances with highest uncertainty, i.e. near the decision boundary, based on the posterior probabilities
\[
\mathrm{argmin}_x \left( \mathrm{max}_y \big( P(y \mid x) \big) \right)
\]
[1] "Active learning with drifting streaming data",
by I. Zliobaite, I., A. Bifet, B. Pfahringer, G. Holmes.
IEEE Transactions on Neural Networks and Learning Systems, 25(1), 2014.
Probabilistic Active Learning [2]
Evaluates each labeling candidate in its neighborhood using its label statistics:
- Posterior probability (\(\hat{p}\))
- Number of labels (\(n\))
- Density (\(d\))
\[
\mathrm{argmax} \left( d \cdot \mathtt{E}_{p}\bigg[ \mathtt{E}_{y} \left[ \mathrm{gain}_{p}((n, \hat{p}),y) \right] \bigg] \right)
\]
[2] "Probabilistic Active Learning: Towards Combining Versatility, Optimality and Efficiency",
by G. Krempl, D. Kottke, M. Spiliopoulou. Discovery Science, Bled, 2014. Springer.
Problem Specification
- Input: Single-value stream of spatial usefulness values
- Task: Select the best b % of this value stream instantly
- Additional Policy:
- Difference of current and target budget should not exceed a given tolerance window (BIQF)
- Budget should not be exceeded (Adaptive Threshold)
Incremental Quantile Filter
Balancing
- Tolerance window (\(w_\textrm{tol}\)):
maximal difference between current and the target budget
- If there are label acquisitions left (\(acq_\textrm{left}\) > 0)
\(\rightarrow\) decrease threshold \(\theta\) (and vice versa)
\[
\theta_\textrm{bal} = \theta - \Delta \cdot \frac{acq_\textrm{left}}{w_\textrm{tol}}
\]
- \(\theta_\textrm{bal}\) - Balanced threshold
- \(\theta\) - IQF acquisition threshold
- \(\Delta\) - Data range of IQF window
- \(w_\textrm{tol}\) - Tolerance window size
Drift in spatial usefulness
- Usefulness values might drift as the underlying
distribution drifts
Fast Trend correction (FTCIQF)
- Interpolate trend with linear function
- Curve fitting using incremental mean \(\mu\) and standard deviation \(\sigma\)
- Transforms usefulness values into the standard score
\(\qquad u' = \frac{u-\mu}{\sigma}\)
Experimental Settings
- Algorithms:
- dPAL + BFTCIQF
- dPAL + BIQF [3]
- Variable Uncertainty (VarUncer) [1], Split [1]
- Random
- 10 fold cross-validation
- Different budgets \(b \in \{0.02, 0,05, 0.1, 0.2, 0.3, 0.5\}\)
- BIQF parameter: \(w = 100, w_{\mathrm{tol}} = 50\)
[1] "Active learning with drifting streaming data",
by I. Zliobaite, I., A. Bifet, B. Pfahringer, G. Holmes.
IEEE Transactions on Neural Networks and Learning Systems, 25(1), 2014.
[3] "Probabilistic Active Learning in Data Streams",
by D. Kottke, G. Krempl, M. Spiliopoulou.
Symposium on Intelligent Data Analysis, Saint-Etienne, 2015.
Evaluation of Active Learning
Randomize BFTCIQF with dPAL
Randomize BFTCIQF with Uncertainty Sampling
Conclusion
- IQF strategies successfully select the best instances temporally based on usefulness values
- FTCIQF improves dPAL
- FTCIQF is not beneficial for Uncertainty Sampling
Future Work
- Extend dPAL with Optimised Probabilistic Active Learning [4]
[4] "Optimised Probabilistic Active learning",
by G. Krempl, D. Kottke, V. Lemaire
Machine Learning, 2015.
Thank you for your attention!
Slides, Paper, Bibtex:
www.daniel.kottke.eu/talks/2016_DAGSTAT
Supplemental material:
kmd.cs.ovgu.de/res/pal/
Temporal Aspects of Stream Active Learning
Daniel Kottke, Georg Krempl, Myra Spiliopoulou
Tagung der Deutschen Arbeitsgemeinschaft Statistik (DAGSTAT)
Göttingen, Germany, 2016.
dPAL and high budgets
Uncer and BFTCIQF