Abstract:
Stream-based active learning for data selection in a real world application Abstract. In this paper, we derive active learning algorithms from a difficult real world classification task. Much research in active learning has been done in the pool-based setting (cf [11]). The strong sequential ordering of our data entails us to study the stream-based setting. Our application deals with the classification of vehicle interactions into rough severity classes for road safety purposes, based on occupation measurements supplied by video sensors. As the data are noisy and the classes intricate, this task is not easily handled with classical batch learning. For that reason, we turn to data selection through active learning. We study algorithms for active learning in the stream-based setting, and evaluate them on our classification task. We show experimentally that data selection based on intricate minority classes and misclassified instances improve the classification results in general and independantly for each class with respect to random data selection and classical batch learning with all available data.
Citations
|
1117
|
E.: Data Mining: Practical machine learning tools and techniques. 2nd edn
– Witten, Frank
- 2005
|
|
483
|
Boosting the Margin: A New Explanation for the Effectiveness of Voting Methods
– Schapire, Freund, et al.
- 1997
|
|
168
|
Selective sampling using the query by committee algorithm
– Freund, Seung, et al.
- 1997
|
|
151
|
Heterogeneous Uncertainty Sampling for Supervised Learning
– Lewis, Catlett
- 1994
|
|
102
|
Less is more: active learning with support vector machines
– Schohn, Cohn
- 2000
|
|
88
|
Toward optimal active learning through sampling estimation of error reduction
– Roy, McCallum
- 2001
|
|
48
|
Query learning strategies using boosting and bagging
– Abe, Mamitsuka
- 1998
|
|
32
|
Online choice of active learning algorithms
– BARAM, EL-YANIV, et al.
- 2003
|
|
26
|
Active learning using adaptive resampling
– IYENGAR, APTE, et al.
|
|
22
|
Integrative windowing
– Furnkranz
- 1998
|
|
22
|
Active Learning: Theory and Applications
– Tong
- 2001
|
|
10
|
Automatically labeling video data using multi-class active learning
– Yan, Yang, et al.
- 2003
|
|
9
|
Haim Sompolinsky. Query by committee
– Seung, Opper
- 1992
|
|
3
|
A time-based analysis of road user behavior in normal and critical encounter
– Horst
- 1990
|
|
2
|
Automatic detection of vehicle interactions in a signalized intersection
– Saunier, Midenet, et al.
- 2003
|
