Project done in the context of a machine learning course INFOF422 : Statistical Foundations of Machine Learning
(Link to course URL : https://www.ulb.be/en/programme/info-f422)
Classifier accuracy : 72.25%
• To participate to the "Pump it Up: Data Mining the Water Table" DrivenData competition by implementing and assessing different supervised learning algorithms and different methods of feature selection in the related classification task.
• to select among the learning and feature selection techniques the ones which appear to be the most accurate and use them for submitting to the DrivenData competition.
• to report your analyses and results as a Jupyter notebook.
DrivenData competition The goal of the competition is to improve maintenance operations of water pumps and ensure that clean, potable water is available to communities across Tanzania. In order to achieve that, a smart understanding of which waterpoints will fail is required. Your objective is to build a predictive model which is able to correctly predict which pumps are functional, which need some repairs, and which don’t work at all, using data from Taarifa and the Tanzanian Ministry of Water. The model has to be trained using the Training set values, Training set labels files available on the DrivenData platform (see Figure 1), it includes roughly 60000 labeled samples and 40 features. The students should then predict the labels for the samples included in the Test set values, and submit them to the platform following the provided Submission format. The students should register using its ULB/VUB netid as username and accept the rules of the competition (notably no hidden additional accounts).
The team of 3 students was asked to :
- implement in the R language a pipeline for data preprocessing, including missing value imputation, normalization (if required), feature engineering and feature selection.
- implement in the R language a model selection procedure. Project 2020-2021 Specifications (and at least three) models (among those presented during the course) which have been taken into consideration and the procedure used for model assessment and selection. The use of figures, formulas, tables and pseudo-code to describe the model selection procedure is strongly encouraged. Note one third of the score will be attributed on the basis of the quality of the documentation.(3 points)
- implement a learning procedure using other R packages
The team has to choose a learning method and a feature selection method among at least three alternatives. For the learning method (at point 2.), the only packages that may be used are those included in this list : • stats/ridge (linear/ridge models) • nnet (neural networks) • tree/rpart (decision trees) • randomForest (random forest) • RSNNS (radial basis functions) • lazy (nearest neighbours) • e1071 (SVM) • glmnet (LASSO/ElasticNet models)
For the point 3. the team is free to employ other learning methods, either already available online, or coded. The quality of the classification models during the selection process should be assessed by using classification accuracy. The report must be an R Jupyter notebook which has to specify and justify (with tables, figures) the selection procedures which led to the final choice. The team has to return, together with the report, the datasets employed in the notebook, the set of predictions submitted to the DrivenData competition and a video summarizing the whole predictive procedure.