Une approche par l’apprentissage approfondi pour l’extraction des connaissances dans les big data

Abstract

In the realm of predictive analytics, two major issues capture attention: mobility prediction and breast cancer diagnosis. Although these problems may seem unre- lated, they are crucial in their respective domains. Mobility prediction is essential for effective urban planning, while accurate diagnosis of breast cancer can save lives. In this thesis, we tackle two problems: (i): Mobility Prediction (MP) problem in big data, i.e, next location prediction of mobile users. (ii): Breast Cancer (BC) clas- sification problem in high dimensionality datasets. We propose four contributions which are: (i): mini survey: we presented some of the well-known solutions used for reduce datasets high dimensionality, and we provided a comparison between the presented solutions. (ii): we introduced a practical comparison using two different classifiers (Multi Layer Perceptron (MLP) and Support Vector Machine (SVM)) combined with five different dimensionality reduction techniques, in order to understand the affect of high dimensionality on classifying Breast Cancer (BC). The results showed that using dimensionality reduction techniques increased the classification accuracy in some cases. The former, also showed that choosing the wrong combination of dimensionality reduction algorithms may lead to a worse results. The following accuracies are some of the results that we got: MLP-PCA, and MLP-ISOMAP outperformed simple MLP model with 0.7% of classification accuracy. SVM-PCA, and SVM-ISOMAP outperformed simple MLP with 0.3%. An example of choosing a bad combination of a dimensionality reduction algo- rithm and an MLP classifier is illustrated in the following example: MLP-AE, and MLP-RFE decreased the classification accuracy by 1.7% compared to simple MLP (iii): WP-BERTA: our proposed solution for next location prediction problem. The former is a combination of Bertwordpiece embedding algorithm and Transformer Roberta algorithm. (iv): WP-CamemBERT: our proposed solution for next loca- tion prediction problem. The former is a combination of Bertwordpiece embedding algorithm and Transformer CamemBERT algorithm. We showed that our solutions (WP-BERTA, and WP-CamemBERT) outperformed state-of-the-art solutions by 3 increasing the next location prediction accuracy by at least 3% compared to the state-of-the-art algorithms.