In economic studies, the input-output (I-O) tables are of great importance. The regional version of these tables are not usually constructed from survey data but are estimated using non survey regionalization methods by the individual analysts. While the LQ-based methods are widely used for this purpose, the current paper presents a new approach based on estimating the regional input coefficients directly by developing machine learning (ML) regression models. ML algorithms are capable to efficiently capture the complicated relationship between the regional economic variables and the corresponding input coefficients. The performance of the Multi layered perceptron (MLP), support vector regression (SVR), and random forest (RF) algorithms are examined in estimating input coefficients on Japan regional I-O tables of 2005. A comparison is made between the results of the ML algorithms and the LQ-based methods. It is found that the best-performing ML algorithm yields results far superior in terms of accuracy and variance to those from LQ-based methods.

The accurate prediction of nonstationary construction costs can contribute to the enhancement of the understanding about sources and patterns of construction costs fluctuations. This understanding can facilitate informed decision making about investment in construction projects. It can help investors better manage the risks associated with construction cost fluctuations and achieve maximum profit. This paper puts forward a novel prediction model for the construction costs of residential buildings. The proposed model comprises two sub-models. A set of variables that determine the building characteristics and the market conditions are the inputs to the first sub-model. This sub-model uses unsupervised deep Boltzmann machine (DBM) learning approach to learn the complex relationships among the explained and explanatory variables. The results are then used in order to build a regression model using support vector regression (SVR) and multi-layer Perceptron (MLP). The first sub-model estimates the current construction cost of a given residential building. The second sub-model, which is based on the adaptive multiscale ensemble-learning paradigm, incorporates ensemble empirical mode decomposition (EEMD), autoregressive integrated moving average (ARIMA), and MLP. This sub-model generates a construction cost time series based on estimated costs of the first sub-model and predicts the construction cost of the residential building under study in the following time steps. In order to evaluate the prediction performance of the proposed model, it is applied to a dataset on the construction costs of 360 residential buildings. The results show that the model is successfully able to predict construction costs of residential buildings to the accuracy performance of 96%.

INSURER website has illustrated that "Moving towards a resilient and sustainable society requires multidisciplinary research on probabilistic modeling and simulation of hazards, the performance of infrastructure, the ensuing impacts, and the process of decision making for recovery, on quantification of sustainability and resilience measures for the society, and on developing software tools for evaluation of policies on optimal enhancement of sustainability and resilience given limited resources." As a small project of this super project, my thesis aims to facilitate the post-disaster decision-making process in construction sector. The strength of the thesis is that the nature of it can be implemented not only in the post-disaster situations but also in different cases and areas such as real estate, stock market, and energy price.