Search

Mestrado Bolonha em Métodos Quantitativos para a Decisão Económica e Empresarial Numa empresa, os clientes alvo representam uma parcela de clientes que são tidos como foco em ações de marketing para venda de determinado produto ou serviço. A conversão de clientes alvo a determinado produto ou serviço gera lucros para a empresa, sendo por isso importante direcionar as ações de marketing a clientes que são mais propensos à conversão. Neste sentido, este estudo tem como principal objetivo obter a probabilidade da conversão de clientes alvo à solução de pagamentos Digital Payment Gateway DPG da SIBS, através de técnicas e algoritmos de Machine Learning. O desenvolvimento deste estudo seguiu a metodologia Cross-Industry Standard Process for Data Mining (CRISP-DM). No balanceamento da classe target, foram utilizadas as técnicas SMOTE e SMOTETomek e os algoritmos de classificação implementados foram: XGBoost, Random Forest e a Regressão Logística. O modelo estimado que apresentou melhor desempenho foi obtido através do algoritmo Random Forest com recurso a dados balanceados através da técnica SMOTE. Este modelo reflete um acerto de 60% das observações pertencentes à classe minoritária. In a company, target customers represent a portion of customers who are the focus of marketing actions for the sale of a certain product or service. The conversion of target customers to a certain product or service generates profits for the company, so it is important to direct marketing actions to customers who are more likely to convert. In this sense, this study's main objective is to obtain the probability of conversion of target customers to the SIBS' Digital Payment Gateway DPG payment solution, through Machine Learning techniques and algorithms. The development of this study followed the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology. In the balancing of the target class, the SMOTE and SMOTETomek techniques were used and the classification algorithms implemented were: XGBoost, Random Forest and Logistic Regression. The estimated model that presented the best performance was obtained through the Random Forest algorithm using balanced data through the SMOTE technique. This model reflects a hit of 60% of the observations belonging to the minority class. info:eu-repo/semantics/publishedVersion

Breast Cancer is the most common cancer found in women and the death rate is still in second place among other cancers. The high accuracy of the machine learning approach that has been proposed by related studies is often achieved. However, without efficient pre-processing, the model of Breast Cancer prediction that was proposed is still in question. Therefore, this research objective to improve the accuracy of machine learning methods through pre-processing: Missing Value Replacement, Data Transformation, Smoothing Noisy Data, Feature Selection / Attribute Weighting, Data Validation, and Unbalanced Class Reduction which is more efficient for Breast Cancer prediction. The results of this study propose several approaches: C4.5 - Z-Score - Genetic Algorithm for Breast Cancer Dataset with 77,27% accuracy, 7-Nearest Neighbor - Min-Max Normalization - Particle Swarm Optimization for Wisconsin Breast Cancer Dataset - Original with 97,85% accuracy, Artificial Neural Network - Z-Score - Forward Selection for Wisconsin Breast Cancer Dataset - Diagnostics with 98,24% accuracy, and 11-Nearest Neighbor - Min-Max Normalization - Particle Swarm Optimization for Wisconsin Breast Cancer Dataset - Prognostic with 83,33% accuracy. The performance of these approaches is better than standard/normal machine learning methods and the proposed methods by the best of previous related studies. &nbsp, Kanker Payudara merupakan kanker yang paling sering ditemukan pada wanita dan tingkat kematiannya masih berada pada posisi dua di antara penyakit kanker laiinya. Akurasi yang tinggi dari pendekatan machine learning yang diusulkan oleh penelitian-penelitian terkait sering dicapai. Namun tanpa pra-pengolahan yang efisien, maka model prediksi Kanker Payudara yang diusulkan masih diragukan. Oleh karena itu, penelitian ini bertujuan untuk meningkatkan kinerja akurasi metode-metode machine learning melalui pra-pengolahan: Missing Value Replacement, Data Transformation, Smoothing Noisy Data, Feature Selection / Attribute Weighting, Data Validation, dan Unbalanced Class Reduction yang lebih efisien untuk prediksi Kanker Payudara. Hasil penelitian ini mengusulkan pendekatan: C4.5 – Z-Score – Genetic Algorithm untuk Breast Cancer Dataset dengan akurasi 77,27%, 7-Nearest Neighbor – Min-Max Normalization – Particle Swarm Optimization untuk Wisconsin Breast Cancer Dataset - Original dengan akurasi 97,85%, Artificial Neural Network – Z-Score – Forward Selection untuk Wisconsin Breast Cancer Dataset – Diagnostic dengan akurasi 98,24%, dan 11-Nearest Neighbor – Min-Max Normalization – Particle Swarm Optimization untuk Wisconsin Breast Cancer Dataset – Prognostic dengan akurasi 83,33%. Kinerja pendekatan-pendekatan tersebut lebih baik dari metode-metode machine learning standar/normal dan metode yang diusulkan penelitian-penelitian terkait sebelumnya yang terbaik. &nbsp

Mestrado em Métodos Quantitativos para a Decisão Económica e Empresarial Este trabalho resulta de um estágio desenvolvido na Empresa Quidgest, S.A. O trabalho final de mestrado versa sobre uma aplicação de Machine Learning na resolução da problemática de combate ao branqueamento de capitais e ao financiamento do terrorismo. Tal problema é conhecido como um caso de dados desbalanceados. Por conseguinte, a questão é abordada no decorrer do trabalho, apresentando várias formas de resolução. São ainda tratados os conceitos Machine Learning, Data Mining e Knowledge-Discovery in Databases. No âmbito do Machine Learning, o presente trabalho apenas se debruça sobre algoritmos supervisionados. Mais especificamente, os classificadores Random Forest, Adaboost e Boosting C5.0. Tais métodos foram aplicados sobre um repositório de dados que se encontravam alojados no sistema de gestão de base de dados Microsoft SQL Server. A investigação seguiu a metodologia CRISP-DM e teve a sua implementação no software R. This work results from an internship developed at Quidgest, S.A. This Master Final Work deals with an application of the Machine Learning in order to solve the problem of money laundering and the financing of terrorism. This problem is known as a case of unbalanced data. Therefore, the issue is addressed in the course of the work, presenting various forms of resolution. The concepts of Machine Learning, Data Mining and Knowledge-Discovery in Databases are also discussed. In Machine Learning, this paper only focuses on supervised algorithms. More specifically, the classifiers: Random Forest, Adaboost, and Boosting C5.0. These methods were applied to a data repository that was hosted in Microsoft SQL Server database management system. The research followed the CRISP-DM methodology and was implemented in the R software. info:eu-repo/semantics/publishedVersion

Kematian yang disebabkan penyakit jantung masih sangat tinggi, sehingga perlu peningkatan upaya-upaya pencegahannya, misalnya dengan meningkatkan capaian model prediksinya. Penerapan metode-metode machine learning pada dataset publik (Cleveland, Hungary, Switzerland, VA Long Beach, & Statlog) yang umumnya digunakan oleh para peneliti untuk prediksi penyakit jantung, termasuk pengembangan alat bantunya, masih belum menangani missing value, noisy data, unbalanced class, dan bahkan data validation secara efisien. Oleh karena itu, pendekatan imputasi mean/mode diusulkan untuk menangani missing value replacement, Min-Max Normalization untuk menangani smoothing noisy data, K-Fold Cross Validation untuk menangani data validation, dan pendekatan ensemble menggunakan metode Weighted Vote (WV) yang dapat menyatukan kinerja tiap-tiap metode machine learning untuk mengambil keputusan klasifikasi sekaligus untuk mereduksi unbalanced class. Hasil penelitian ini menunjukkan bahwa metode yang diusulkan tersebut memberikan akurasi sebesar 85,21%, sehingga mampu meningkatkan kinerja akurasi metode-metode machine learning, selisih 7,14% dengan Artificial Neural Network, 2,77% dengan Support Vector Machine, 0,34% dengan C4.5, 2,94% dengan Naïve Bayes, dan 3,95% dengan k-Nearest Neighbor.