Project on statistical modeling for complex data and big data, for 3rd year BUT students.

Cours description

Objective: Statistical modeling based on a given dataset, following all the steps required to understand its variables, process them and model them for a given objective.

Method:

• Individual work, in groups of 2 or 3 students (chosen at random), using Python.
• Objectives will be progressively set, every 2 sessions each group will have to provide the progress of their work (in the form of a notebook).

Evaluation:

• Continuous assessment (1/5): Work progress.
• Written report (2/5): A written report (clean) in which all the results obtained are brought together, commented on and put into perspective. Due on 13/06 (at the end of the session).
• Oral exam (2/5): 17/06 from 8am. Each group will present its work in the form of a talk (with support). 15-minute presentation + 15-minute Q&A.

Groups (Randomly defined)

• G1 CROMBET, NIASSY
• G2 FALAIS, EZ-ZEROUALI
• G3 VIAUD, CARTON
• G4 VIGLIETTI, HIRSINGER
• G5 GOURAR, DIALLO
• G6 JOUNDI, LOFTI, NOUIRA

Data (training and test sets)

Data G1 (.zip)
Data G2 (.zip)
Data G3 (.zip)
Data G4 (.zip)
Data G5 (.zip)
Data G6 (.zip)

1st Step: data analysis and preparation

• Dataset size and type. Are there any missing data? If so, we will eliminate them (for now).

• How many classes and elements/class?

• Selection of the most significant variables: explore and suggest several methods. Are there any collinearities? Which variables will you retain? How are the selected variables distributed?

• Pay particular attention to the graphic visualization of your results.

2nd Step: choice of classification model, hyperparameters tuning

• Preparation of training dataset for cross-validation (choice of number of folds, split).

• Several models can be adapted to the task at hand: logistic regression, SVM, random forests, xgboost,…
  • Initialize the models
  • Test them separately (k-fold cross-validation)
  • Tune their hyperparameters

• Compare the performance of the different models tested (each with its final choice of hyperparameters).

• Which model/parameters did you choose?

• Could a voting strategy improve predictions?

• With the final model chosen, predict the labels of the test dataset (to be saved on a .csv and sent with the notebook).

• Pay particular attention to the graphical visualization of your results.