Nowadays, knowledge discovery on social media is attracting growing interest. Social media has become more than a communication tool, effectively functioning as a social sensor for our society. This dissertation focuses on the development of methods for social media-based spatiotemporal event detection and forecasting for a variety of event topics and assumptions. Five methods are proposed, namely dynamic query expansion for event detection, a generative framework for event forecasting, multi-task learning for spatiotemporal event forecasting, multi-source spatiotemporal event forecasting, and deep learning based epidemic modeling for forecasting influenza outbreaks. For the first of these methods, existing solutions for spatiotemporal event detection are mostly supervised and lack the flexibility to handle the dynamic keywords used in social media. The contributions of this work are: (1) Develop an unsupervised framework; (2) Design a novel dynamic query expansion method; and (3) Propose an innovative local modularity spatial scan algorithm. For the second of these methods, traditional solutions are unable to capture the spatiotemporal context, model mixed-type observations, or utilize prior geographical knowledge. The contributions of this work include: (1) Propose a novel generative model for spatial event forecasting; (2) Design an effective algorithm for model parameter inference; and (3) Develop a new sequence likelihood calculation method. For the third method, traditional solutions cannot deal with spatial heterogeneity or handle the dynamics of social media data effectively. This work's contributions include: (1) Formulate a multi-task learning framework for event forecasting; (2) simultaneously model static and dynamic terms; and (3) Develop efficient parameter optimization algorithms. For the fourth method, traditional multi-source solutions typically fail to consider the geographical hierarchy or cope with incomplete data blocks among different sources. The contributions here are: (1) Design a framework for event forecasting based on hierarchical multi-source indicators; (2) Propose a robust model for geo-hierarchical feature selection; and (3) Develop an efficient algorithm for model parameter optimization. For the last method, existing work on epidemic modeling either cannot ensure timeliness, or cannot characterize the underlying epidemic propagation mechanisms. The contributions of this work include: (1) Propose a novel integrated framework for computational epidemiology and social media mining; (2) Develop a semi-supervised multilayer perceptron for mining epidemic features; and (3) Design an online training algorithm. / Ph. D.

The rising awareness of existing problems created by rapid population expansion has resulted in systematic investigations of the characteristics of population growth. These investigations have produced methods for projection of future populations. Attempts have been made to project world population, but the situation is too heterogeneous to provide useful results. Population problems, although of world-wide importance, are problems of particular peoples and particular areas. Some of the earliest methods of projection used in the United States were based on the Malthusian Law and geometric progression.Pritchett and Pearl, in the late 1800's and early 1900's, devised parabolic methods of projection. These early projections were good for short term projection but generally unrealistic for long range use. In 1920 Pearl and Reed devised an empirical curve, later known as the logistic curve of' population growth.. This method received considerable attention. The logistic was supported by many later demographers and the resulting projections satisfied all but a few critics. Whelpton's “analytical method,” and other similar methods, have been widely accepted. They give emphasis to birth-, death-, and net-reproduction-rates and not to mathematical growth curves. Many of the above methods are used to make projections based on census counts to date. These projections are compared and tables used to show the results. / Master of Science

Intelligence analysts today are faced with many challenges, chief among them being the need to fuse disparate streams of data, as well as rapidly arrive at analytical decisions and quantitative predictions for use by policy makers. These problems are further exacerbated by the sheer volume of data that is available to intelligence analysts. Machine learning methods enable the automated transduction of such large datasets from raw feeds to actionable knowledge but successful use of such methods require integrated frameworks for contextualizing them within the work processes of the analyst. Intelligence analysts typically distinguish between three classes of problems: collections, analysis, and operations. This dissertation specifically focuses on two problems in analysis: i) the reconstruction of shredded documents using a visual analytic framework combining computer vision techniques and user input, and ii) the design and implementation of a system for event forecasting which allows an analyst to not just consume forecasts of significant societal events but also understand the rationale behind these alerts and the use of data ablation techniques to determine the strength of conclusions. This work does not attempt to replace the role of the analyst with machine learning but instead outlines several methods to augment the analyst with machine learning. In doing so this dissertation also explores the responsibilities of an analyst in evaluating complex models and decisions made by these models. Finally, this dissertation defines a list of responsibilities for models designed to aid the analyst's work in evaluating and verifying the models. / Ph. D.

The possibility of profitable multiple year pricing using rollover strategies for com and soybeans is identified. Historical futures price distributions are generated for both commodities to determine the probability of prices reaching certain levels. The upper 5%, 10%, and 15% of the distributions are determined. Price forecasting models are developed to help producers anticipate high price levels before they occur. Seven different multiple year strategies containing various combinations of cash, futures, and options contracts are established and six different strategy rules are tested. A total of fifty strategies are then evaluated for each commodity over the 1980-1992 time period. Mean net prices and standard deviations are calculated and the highest return strategies are identified. The strategies are then analyzed based on the two largest risks associated with long-term rollovers: margin calls and spread risk. The tradeoffs between risk and return for the various combinations of cash, futures, and options contracts is discussed. The highest return strategy for both com and soybeans involves selling three years of production when prices reach the upper 5% of the historical distribution, using cash contracts to price the first year's production and futures contracts to price the final two. Substituting options contracts for futures in the final two years results in a strategy void of margin call risk, but subject to increased spread risk. For com, a strategy that does not carry the risk of margin calls receives 93.7% of the high return strategy, while for soybeans this percentage is 98.3. / Master of Science

There is a recent surge in the development of spatio-temporal forecasting models in many applications, and traffic forecasting is one of the most important ones. Long-range traffic forecasting, however, remains a challenging task due to the intricate and extensive spatio-temporal correlations observed in traffic networks. Current works primarily rely on road networks with graph structures and learn representations using graph neural networks (GNNs), but this approach suffers from over-smoothing problem in deep architectures. To tackle this problem, recent methods introduced the combination of GNNs with residual connections or neural ordinary differential equations (NODEs). The existing graph ODE models are still limited in feature extraction due to (1) having bias towards global temporal patterns and ignoring local patterns which are crucial in case of unexpected events; (2) missing dynamic semantic edges in the model architecture; and (3) using simple aggregation layers that disregard the high-dimensional feature correlations. In this thesis, we propose a novel architecture called Graph-based Multi-ODE Neural Networks (GRAM-ODE) which is designed with multiple connective ODE-GNN modules to learn better representations by capturing different views of complex local and global dynamic spatio-temporal dependencies. We also add some techniques to further improve the communication between different ODE-GNN modules towards the forecasting task. Extensive experiments conducted on four real-world datasets demonstrate the outperformance of GRAM-ODE compared with state-of-the-art baselines as well as the contribution of different GRAM-ODE components to the performance. / Master of Science / There is a recent surge in the development of spatio-temporal forecasting models in many applications, and traffic forecasting is one of the most important ones. In traffic forecasting, current works limited in correctly capturing the key correlation of spatial and temporal patterns. In this thesis, we propose a novel architecture called Graph-based Multi-ODE Neural Networks (GRAM-ODE) to tackle the problem by using the separate ODE modules to deal with spatial and temporal patterns and further improve the communication between different modules. Extensive experiments conducted on four real-world datasets demonstrate the outperformance of GRAM-ODE compared with state-of-the-art baselines.

Intelligence analysts today are faced with many challenges, chief among them being the need to fuse disparate streams of data and rapidly arrive at analytical decisions and quantitative predictions for use by policy makers. A forecasting tool to anticipate key events of interest is an invaluable aid in helping analysts cut through the chatter. We present the design of user interfaces for the EMBERS system, an anticipatory intelligence system that ingests myriad open source data streams (e.g., news, blogs, tweets, economic and financial indicators, search trends) to generate forecasts of significant societal-level events such as disease outbreaks, protests, and elections. A key research issue in EMBERS is not just to generate high-quality forecasts but provide interfaces for analysts so they can understand the rationale behind these forecasts and pose why, what-if, and other exploratory questions. This thesis presents the design and implementation of three visualization interfaces for EMBERS. First, we illustrate how the rationale behind forecasts can be presented to users through the use of an audit trail and its associated visualization. The audit trail enables an analyst to drill-down from a final forecast down to the raw (and processed) data sources that contributed to the forecast. Second, we present a forensics tool called Reverse OSI that enables analysts to investigate if there was additional information either in existing or new data sources that can be used to improve forecasting. Unlike the audit trail which captures the transduction of data from raw feeds into alerts, Reverse OSI enables us to posit connections from (missed) forecasts back to raw feeds. Finally, we present an interactive machine learning approach for analysts to steer the construction of machine learning mod-els. This provides fine-grained control into tuning tradeoffs underlying EMBERS. Together, these three interfaces support a range of functionality in EMBERS, from visualization of algorithm output to a complete framework for user feedback via a tight human-algorithm loop. They are currently being utilized by a range of user groups in EMBERS: analysts, social scientists, and machine learning developers, respectively. / Master of Science

In the first chapter of this dissertation, I explore how forecaster attention, or the degree to which new information is incorporated into forecasts, is reflected at the lower-dimensional factor representation of multivariate forecast data. When information is costly to acquire, forecasters may pay more attention to some sources of information and ignore others. How much attention they pay will determine the strength of the forecast correlation (factor) structure. Using a factor model representation, I show that a forecast made by a rationally inattentive agent will include an extra shrinkage and thresholding "attention matrix" relative to a full information benchmark, and propose an econometric procedure to estimate it. Differences in the degree of forecaster attentiveness can explain observed differences in empirical shrinkage in professional macroeconomic forecasts relative to a consensus benchmark. Forecasters share the same reduced-form model, but differ in their measured attention. Better-performing forecasters have higher measured attention (lower shrinkage) than their poorly-performing peers. Measured forecaster attention to multiple dimensions of the information space can largely be captured by a single scalar cost parameter. I propose a new class of information cost functions for the classic multivariate linear-quadratic Gaussian tracking problem called separable spectral cost functions. The proposed measure of attention and mapping from theoretical model of attention allocation to factor structure in the first chapter is valid for this set of cost functions. These functions are defined over the eigenvalues of prior and posterior variance matrices. Separable spectral cost functions both nest known cost functions and are consistent with the definition of Uniformly Posterior Separable cost functions, which have desirable theoretical properties. The third chapter is coauthored work with Professor Serena Ng. We estimate higher frequency values of monthly macroeconomic data using different factor based imputation methods. Monthly and weekly economic indicators are often taken to be the largest common factor estimated from high and low frequency data, either separately or jointly. To incorporate mixed frequency information without directly modeling them, we target a low frequency diffusion index that is already available, and treat high frequency values as missing. We impute these values using multiple factors estimated from the high frequency data. In the empirical examples considered, static matrix completion that does not account for serial correlation in the idiosyncratic errors yields imprecise estimates of the missing values irrespective of how the factors are estimated. Single equation and systems-based dynamic procedures that account for serial correlation yield imputed values that are closer to the observed low frequency ones. This is the case in the counterfactual exercise that imputes the monthly values of consumer sentiment series before 1978 when the data was released only on a quarterly basis. This is also the case for a weekly version of the CFNAI index of economic activity that is imputed using seasonally unadjusted data. The imputed series reveals episodes of increased variability of weekly economic information that are masked by the monthly data, notably around the 2014-15 collapse in oil prices.

Accurate forecasting of infectious disease outbreaks is vital for safeguarding global health and the well-being of individuals. Model-based forecasts enable public health officials to test what-if scenarios, evaluate control strategies, and develop informed policies to allocate resources effectively. Model selection is a pivotal aspect of creating dependable forecasts for infectious diseases. This thesis delves into validating forecasts of simple epidemic models. We use incidence data from the 2015-2016 Zika virus outbreak in Antioquia, Colombia, to assess what model features result in accurate forecasts. We employed the Parametric Bootstrapping and Ensemble Kalman Filter methods to assimilate data and then generated 14-day-ahead forecasts throughout the epidemic across five case studies. We visualized each forecast to show the training/testing split in data and associated prediction intervals. Fore- casting accuracy was evaluated using five statistical performance metrics. Early into the epidemic, phenomenological models - like the generalized logistic model - resulted in more accurate forecasts. However, as the epidemic progressed, the mechanistic model incorporating disease latency outperformed its counterparts. While modeling disease transmission mechanisms is crucial for accurate Zika incidence forecasting, additional data is needed to make these models more reliable and precise. / Master of Science / Accurate forecasting of infectious disease outbreaks is vital for safeguarding global health and the well-being of individuals. Model-based forecasts enable public health officials to test what-if scenarios, evaluate control strategies, and develop informed policies to allocate resources effectively. Model selection is a pivotal aspect of creating dependable forecasts for infectious diseases. This thesis delves into validating forecasts of simple epidemic models. We use data from the 2015-2016 Zika virus outbreak in Antioquia, Colombia, to assess what model features result in accurate forecasts. We considered two techniques to generate 14-day-ahead forecasts throughout the epidemic across five case studies. We visualized each forecast and evaluated model accuracy. Early into the epidemic, simple growth models resulted in more accurate forecasts. However, as the epidemic progressed, the model incorporating disease-specific characteristics outperformed its counterparts. While modeling disease transmission is crucial for accurate epidemic forecasting, additional data is needed to make these models more reliable and precise.

The market for natural gas in the United States follows a yearly price pattern of high prices during the winter heating season and lows during the summer months. During the winter heating season the daily and weekly price fluctuations for natural gas are normally related to ambient air temperature and other weather related phenomenon. This paper examines a natural gas price forecasting model developed by the U.S. Department of Energy, Energy Information Agency (EIA). This paper proposes that a more accurate forecasting model can be created from the EIA model by focusing on forecasting price during only the winter heating season and by adding other variables to the EIA model. The forecasting results of the core EIA model are compared to the results of other linear regression models. / Master of Arts

Travel booking behavior has changed substantially over the past two decades. The emergence of new technology and online intermediaries has provided travelers with the flexibility to book up until the date of stay. This has created a fast-paced, dynamic booking environment that disrupts traditional revenue management strategies focused on pricing and allocating rooms based on the time of purchase. The study explores the joint effects of technology and the economy on booking window lead times. It also evaluates a range of forecasting techniques and the importance of utilizing the booking curve for forecasting in dynamic booking environments. / PHD / Travel booking behavior has changed substantially over the past two decades. The emergence of new technology and online intermediaries has provided travelers with the flexibility to book up until the date of stay. This has created a fast-paced, dynamic booking environment that disrupts traditional revenue management strategies focused on pricing and allocating rooms based on the time of purchase. The study explores the joint effects of technology and the economy on booking window lead times. It also evaluates a range of forecasting techniques and the importance of utilizing the booking curve for forecasting in dynamic booking environment.