Explicit or Implicit? On feature engineering for ML-based Variability-Intensive Systems

a short presentation about the use of feature in Machine Learning and configurable systems and perspectives to bring together these 2 domains.

1. Explicit or Implicit? On Feature Engineering for
ML-based Variability-intensive Systems
Paul Temple
1
and Gilles Perrouin
2
1 Univ Rennes, CNRS, Inria, IRISA
2 UNamur, NaDI, PReCISE, FNRS
January 26, 2023
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

2. Machine Learning to infer constraints
Temple et al.,Using machine learning to infer constraints for product lines,
SPLC'16https://doi.org/10.1145/2934466.2934472
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

3. Machine Learning in SE
Sampling!selectcongurations
Measuring!measuretarget value
Learning!traina model
Whatfeaturesto train on?
Pereira et al.,Learning software conguration spaces, JSS'21 (182)
https://doi.org/10.1016/j.jss.2021.111044
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

4. Machine Learning in SE
Sampling!selectcongurations
Measuring!measuretarget value
Learning!traina model
Whatfeaturesto train on?
Pereira et al.,Learning software conguration spaces, JSS'21 (182)
https://doi.org/10.1016/j.jss.2021.111044
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

5. Machine Learning in SE
Sampling!selectcongurations
Measuring!measuretarget value
Learning!traina model
Whatfeaturesto train on?
Pereira et al.,Learning software conguration spaces, JSS'21 (182)
https://doi.org/10.1016/j.jss.2021.111044
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

6. What features to train on?
The ones from the FM
But take care of...
Homogeneous dimensions
Categorical features!set of Boolean features
Constraints (manual check? Solvers?)
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

7. What features to train on?
The ones from the FM
But take care of...
Homogeneous dimensions
Categorical features!set of Boolean features
Constraints (manual check? Solvers?)
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

8. Features and Congurable systems
Multiple denitions in SPL
Requirements, design element, artifact!distinguishproducts
Explicit!direct linkbetween user selection and concrete features
Halin et al.,Test them all, is it worth it?, EMSE'19 (24)
https://doi.org/10.1007/s10664-018-9635-4
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

9. Features and Congurable systems
Multiple denitions in SPL
Requirements, design element, artifact!distinguishproducts
Explicit!direct linkbetween user selection and concrete features
Halin et al.,Test them all, is it worth it?, EMSE'19 (24)
https://doi.org/10.1007/s10664-018-9635-4
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

10. Features in ML
Tabular data
Dened by experts
Observable
Distinguish categories
)similar to software congurations
Other domains (image, sound, ...)
Dened by experts
Combination of "primitives"
Distinguish categories
)less explicit but still manageable
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

11. Features in ML
Tabular data
Dened by experts
Observable
Distinguish categories
)similar to software congurations
Other domains (image, sound, ...)
Dened by experts
Combination of "primitives"
Distinguish categories
)less explicit but still manageable
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

12. Features in DL
Embedding
Input elements!tensors
New structured representations!similarity distance
Automatically inferred features)implicit
)Performance%interpretation&control&
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

13. Features in DL
Embedding
Input elements!tensors
New structured representations!similarity distance
Automatically inferred features)implicit
)Performance%interpretation&control&
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

14. Features in DL
Embedding
Input elements!tensors
New structured representations!similarity distance
Automatically inferred features)implicit
)Performance%interpretation&control&
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

15. What do we want for the future?
Here we are today
SE and Software Variability use ML/DL models
10-15 years late compared to ML practitioners
Choose: explicit ...
Stay with ML models!are we done yet?
Control and interpretability!required in some domains
or implicit?
Use DL models and automate!Less
Probably
or maybe a trade-o?
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

16. What do we want for the future?
Here we are today
SE and Software Variability use ML/DL models
10-15 years late compared to ML practitioners
Choose: explicit ...
Stay with ML models!are we done yet?
Control and interpretability!required in some domains
or implicit?
Use DL models and automate!Less
Probably
or maybe a trade-o?
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

17. What do we want for the future?
Here we are today
SE and Software Variability use ML/DL models
10-15 years late compared to ML practitioners
Choose: explicit ...
Stay with ML models!are we done yet?
Control and interpretability!required in some domains
or implicit?
Use DL models and automate!Less
Probably
or maybe a trade-o?
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023

18. What do we want for the future?
Here we are today
SE and Software Variability use ML/DL models
10-15 years late compared to ML practitioners
Choose: explicit ...
Stay with ML models!are we done yet?
Control and interpretability!required in some domains
or implicit?
Use DL models and automate!Less
Probably
or maybe a trade-o?
Temple and Perrouin (1 Univ Rennes, CNRS, Inria, IRISA 2 UNamur, NaDI, PReCISE, FNRS)SPLs and ML Features January 26, 2023