Machine and Deep Learning models for house price prediction in United States of America and Portugal

Abstract

The present study describes the development process of a system to predict the houses’ prices in Portugal. Two main phases of this process were the data extraction and the comparison among several algorithms. Data Extraction was made through Web Scraping techniques applied the Mais Consultores site [1]. This study used Text Mining methods - Rule-based Matching and Similarity – in order to structure and obtain meaning from the information extracted. Afterwards, this thesis made a comparison among the application of Machine Learning and Deep Learning algorithms: Support Vector Machines (SVM), Decision Tree Regressor (DTR), Random Forest, K-Nearest Neighbour (KNN), Artificial Neural Networks ANN, Convolutional Neural Networks CNN, Recurrent Neural Networks RNN, Multi-layer Perceptron MLP and Long Short-term Memory LSTM Network. Finding this solution was the prime motivation of the present thesis. The results obtained by the used algorithms, both Machine Learning and Deep Learning, demonstrated that the algorithms needed more data for the training set. Additionally, the algorithms with the best results, i.e., with the lesser value of Mean Absolute Error (MAE), Mean Square Error (MSE) and Root Mean Square Error (RSME) and the better score were the Deep Learning algorithms.

A presente estudo utilizou a medotologia CRISP-DM para a caraterização e descrição do processo de desenvolvimento de um sistema para estimação dos preço das casas em Portugal. Duas fases importantes no processo foram: a extração de dados e a comparação entre vários algoritmos. A extração de dados foi realizada através de técnicas de Web Scraping a partir do site Mais Consultores [1]. Utilizaram-se métodos de Text Mining - Rule-based Matching e Similarity – para estruturar e retirar significado da informação que se extraiu do site. De seguida, realizámos a comparação entre a aplicação de algoritmos de Machine Learning e Deep Learning: Support Vector Machines (SVM), Decision Tree Regressor (DTR), Random Forest, K-Nearest Neighbour (KNN), Artificial Neural Networks ANN, Convolutional Neural Networks CNN, Recurrent Neural Networks RNN, Multi-layer Perceptron MLP and Long Short-term Memory LSTM Network. Encontrar esta solução constituiu a principal motivação da presente tese. Os resultados obtidos pelos algoritmos utilizados, tanto os de Machine Learning como os de Deep Learning, demonstram que os algoritmos precisavam de mais dados para treino. Adicionalmente, os algoritmos com melhores resultados, i.e., com menor Mean Absolute Error (MAE), Mean Square Error (MSE) e Root Mean Square Error (RSME) e maior score foram os algorimos de Deep Learning.