Skip to Main content Skip to Navigation
Conference papers

Predictive versus Generative Modelling: a Challenge for (Social) Sciences

Gilbert Saporta 1
1 CEDRIC - MSDMA - CEDRIC. Méthodes statistiques de data-mining et apprentissage
CEDRIC - Centre d'études et de recherche en informatique et communications
Abstract : We carry on the reflexion initiated in [1]. Donoho [2], reinterpreting [3], notes that most developments in academic statistics are oriented towards inference: ie fitting, estimating and validating generative models, preferably parcimonious, rather than for prediction where “unfortunately, accuracy and simplicity (interpretability) are in conflict”. Classical inference corresponds to the role of statistics as an auxiliary of sciences. However in most sciences, a good model should also provides accurate predictions, which becomes the sole criterium in decision sciences like pattern recognition, customer behaviour, etc. The most efficient predictive models are rather black-box algorithms like random forests or deep learning. Some consider that “statistics is the least important part of data science” [4] while others claim that even science is obsolete [5]! Meanwhile, renowned scientists [6] are calling practitioners of social sciences to study Machine Learning . The use of black-box models fitted for massive data is probably the main challenge for social sciences due to their lack of interpretability. Getting better predictions, thanks to a better understanding of the real world, needs to combine statistics and machine learning with causal inference. References: [1] G.Saporta (2008). Models for Understanding versus Models for Prediction, In P.Brito, ed., Compstat Proceedings, Physica Verlag, 315-322[2] D.Donoho (2015). 50 years of Data Science, Tukey Centennial workshop,[3] L.Breiman (2001). Statistical Modeling: The Two Cultures, Statistical Science, 16, 3, 199–231 [4] A.Gelman (2013).[5] C.Anderson (2008). The End of Theory: The Data Deluge Makes the Scientific Method Obsolete,[6] H.Varian (2014). Big Data: New Tricks for Econometrics, Journal of Economic Perspectives, 28, 2, 3–28
Document type :
Conference papers
Complete list of metadatas

Cited literature [12 references]  Display  Hide  Download
Contributor : Gilbert Saporta <>
Submitted on : Wednesday, April 22, 2020 - 9:44:53 AM
Last modification on : Thursday, May 7, 2020 - 1:21:57 PM


  • HAL Id : hal-02500634, version 1



Gilbert Saporta. Predictive versus Generative Modelling: a Challenge for (Social) Sciences. Data Science & Social Research Conference, Feb 2016, Naples, Italy. ⟨hal-02500634⟩



Record views


Files downloads