Algorithms for the Hard Pre-Image Problem of String Kernels and the General Problem of String Prediction

Sébastien Giguère, Amélie Rolland, Francois Laviolette, Mario Marchand
Proceedings of the 32nd International Conference on Machine Learning, PMLR 37:2021-2029, 2015.

Abstract

We address the pre-image problem encountered in structured output prediction and the one of finding a string maximizing the prediction function of various kernel-based classifiers and regressors. We demonstrate that these problems reduce to a common combinatorial problem valid for many string kernels. For this problem, we propose an upper bound on the prediction function which has low computational complexity and which can be used in a branch and bound search algorithm to obtain optimal solutions. We also show that for many string kernels, the complexity of the problem increases significantly when the kernel is normalized. On the optical word recognition task, the exact solution of the pre-image problem is shown to significantly improve the prediction accuracy in comparison with an approximation found by the best known heuristic. On the task of finding a string maximizing the prediction function of kernel-based classifiers and regressors, we highlight that existing methods can be biased toward long strings that contain many repeated symbols. We demonstrate that this bias is removed when using normalized kernels. Finally, we present results for the discovery of lead compounds in drug discovery. The source code can be found at https://github.com/a-ro/preimage

Cite this Paper


BibTeX
@InProceedings{pmlr-v37-giguere15, title = {Algorithms for the Hard Pre-Image Problem of String Kernels and the General Problem of String Prediction}, author = {Giguère, Sébastien and Rolland, Amélie and Laviolette, Francois and Marchand, Mario}, booktitle = {Proceedings of the 32nd International Conference on Machine Learning}, pages = {2021--2029}, year = {2015}, editor = {Bach, Francis and Blei, David}, volume = {37}, series = {Proceedings of Machine Learning Research}, address = {Lille, France}, month = {07--09 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v37/giguere15.pdf}, url = {https://proceedings.mlr.press/v37/giguere15.html}, abstract = {We address the pre-image problem encountered in structured output prediction and the one of finding a string maximizing the prediction function of various kernel-based classifiers and regressors. We demonstrate that these problems reduce to a common combinatorial problem valid for many string kernels. For this problem, we propose an upper bound on the prediction function which has low computational complexity and which can be used in a branch and bound search algorithm to obtain optimal solutions. We also show that for many string kernels, the complexity of the problem increases significantly when the kernel is normalized. On the optical word recognition task, the exact solution of the pre-image problem is shown to significantly improve the prediction accuracy in comparison with an approximation found by the best known heuristic. On the task of finding a string maximizing the prediction function of kernel-based classifiers and regressors, we highlight that existing methods can be biased toward long strings that contain many repeated symbols. We demonstrate that this bias is removed when using normalized kernels. Finally, we present results for the discovery of lead compounds in drug discovery. The source code can be found at https://github.com/a-ro/preimage} }
Endnote
%0 Conference Paper %T Algorithms for the Hard Pre-Image Problem of String Kernels and the General Problem of String Prediction %A Sébastien Giguère %A Amélie Rolland %A Francois Laviolette %A Mario Marchand %B Proceedings of the 32nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2015 %E Francis Bach %E David Blei %F pmlr-v37-giguere15 %I PMLR %P 2021--2029 %U https://proceedings.mlr.press/v37/giguere15.html %V 37 %X We address the pre-image problem encountered in structured output prediction and the one of finding a string maximizing the prediction function of various kernel-based classifiers and regressors. We demonstrate that these problems reduce to a common combinatorial problem valid for many string kernels. For this problem, we propose an upper bound on the prediction function which has low computational complexity and which can be used in a branch and bound search algorithm to obtain optimal solutions. We also show that for many string kernels, the complexity of the problem increases significantly when the kernel is normalized. On the optical word recognition task, the exact solution of the pre-image problem is shown to significantly improve the prediction accuracy in comparison with an approximation found by the best known heuristic. On the task of finding a string maximizing the prediction function of kernel-based classifiers and regressors, we highlight that existing methods can be biased toward long strings that contain many repeated symbols. We demonstrate that this bias is removed when using normalized kernels. Finally, we present results for the discovery of lead compounds in drug discovery. The source code can be found at https://github.com/a-ro/preimage
RIS
TY - CPAPER TI - Algorithms for the Hard Pre-Image Problem of String Kernels and the General Problem of String Prediction AU - Sébastien Giguère AU - Amélie Rolland AU - Francois Laviolette AU - Mario Marchand BT - Proceedings of the 32nd International Conference on Machine Learning DA - 2015/06/01 ED - Francis Bach ED - David Blei ID - pmlr-v37-giguere15 PB - PMLR DP - Proceedings of Machine Learning Research VL - 37 SP - 2021 EP - 2029 L1 - http://proceedings.mlr.press/v37/giguere15.pdf UR - https://proceedings.mlr.press/v37/giguere15.html AB - We address the pre-image problem encountered in structured output prediction and the one of finding a string maximizing the prediction function of various kernel-based classifiers and regressors. We demonstrate that these problems reduce to a common combinatorial problem valid for many string kernels. For this problem, we propose an upper bound on the prediction function which has low computational complexity and which can be used in a branch and bound search algorithm to obtain optimal solutions. We also show that for many string kernels, the complexity of the problem increases significantly when the kernel is normalized. On the optical word recognition task, the exact solution of the pre-image problem is shown to significantly improve the prediction accuracy in comparison with an approximation found by the best known heuristic. On the task of finding a string maximizing the prediction function of kernel-based classifiers and regressors, we highlight that existing methods can be biased toward long strings that contain many repeated symbols. We demonstrate that this bias is removed when using normalized kernels. Finally, we present results for the discovery of lead compounds in drug discovery. The source code can be found at https://github.com/a-ro/preimage ER -
APA
Giguère, S., Rolland, A., Laviolette, F. & Marchand, M.. (2015). Algorithms for the Hard Pre-Image Problem of String Kernels and the General Problem of String Prediction. Proceedings of the 32nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 37:2021-2029 Available from https://proceedings.mlr.press/v37/giguere15.html.

Related Material