The dawn of Internet of Things (IoT) has opened the opportunity for mainstream adoption of machine learning analytics. However, most research in machine learning has focused on discovery of new algorithms or fine-tuning the performance of existing algorithms. Little exists on the process of taking an algorithm from the lab-environment into the real-world, culminating in sustained value. Real-world applications are typically characterized by dynamic non-stationary systems with requirements around feasibility, stability and maintainability. Not much has been done to establish standards around the unique analytics demands of real-world scenarios.

This research explores the problem of the why so few of the published algorithms enter production and furthermore, fewer end up generating sustained value. The dissertation proposes a ‘Design for Deployment’ (DFD) framework to successfully build machine learning analytics so they can be deployed to generate sustained value. The framework emphasizes and elaborates the often neglected but immensely important latter steps of an analytics process: ‘Evaluation’ and ‘Deployment’. A representative evaluation framework is proposed that incorporates the temporal-shifts and dynamism of real-world scenarios. Additionally, the recommended infrastructure allows analytics projects to pivot rapidly when a particular venture does not materialize. Deployment needs and apprehensions of the industry are identified and gaps addressed through a 4-step process for sustainable deployment. Lastly, the need for analytics as a functional area (like finance and IT) is identified to maximize the return on machine-learning deployment.

The framework and process is demonstrated in semiconductor manufacturing – it is highly complex process involving hundreds of optical, electrical, chemical, mechanical, thermal, electrochemical and software processes which makes it a highly dynamic non-stationary system. Due to the 24/7 uptime requirements in manufacturing, high-reliability and fail-safe are a must. Moreover, the ever growing volumes mean that the system must be highly scalable. Lastly, due to the high cost of change, sustained value proposition is a must for any proposed changes. Hence the context is ideal to explore the issues involved. The enterprise use-cases are used to demonstrate the robustness of the framework in addressing challenges encountered in the end-to-end process of productizing machine learning analytics in dynamic read-world scenarios.

Stock market news and investing tips are popular topics in Twitter. In this dissertation, first I utilize a 5-year financial news corpus comprising over 50,000 articles collected from the NASDAQ website matching the 30 stock symbols in Dow Jones Index (DJI) to train a directional stock price prediction system based on news content. Next, I proceed to show that information in articles indicated by breaking Tweet volumes leads to a statistically significant boost in the hourly directional prediction accuracies for the DJI stock prices mentioned in these articles. Secondly, I show that using document-level sentiment extraction does not yield a statistically significant boost in the directional predictive accuracies in the presence of other 1-gram keyword features. Thirdly I test the performance of the system on several time-frames and identify the 4 hour time-frame for both the price charts and for Tweet breakout detection as the best time-frame combination. Finally, I develop a set of price momentum based trade exit rules to cut losing trades early and to allow the winning trades run longer. I show that the Tweet volume breakout based trading system with the price momentum based exit rules not only improves the winning accuracy and the return on investment, but it also lowers the maximum drawdown and achieves the highest overall return over maximum drawdown.

This thesis studies recommendation systems and considers joint sampling and learning. Sampling in recommendation systems is to obtain users' ratings on specific items chosen by the recommendation platform, and learning is to infer the unknown ratings of users to items given the existing data. In this thesis, the problem is formulated as an adaptive matrix completion problem in which sampling is to reveal the unknown entries of a $U\times M$ matrix where $U$ is the number of users, $M$ is the number of items, and each entry of the $U\times M$ matrix represents the rating of a user to an item. In the literature, this matrix completion problem has been studied under a static setting, i.e., recovering the matrix based on a set of partial ratings. This thesis considers both sampling and learning, and proposes an adaptive algorithm. The algorithm adapts its sampling and learning based on the existing data. The idea is to sample items that reveal more information based on the previous sampling results and then learn based on clustering. Performance of the proposed algorithm has been evaluated using simulations.

Advances in data collection technologies have made it cost-effective to obtain heterogeneous data from multiple data sources. Very often, the data are of very high dimension and feature selection is preferred in order to reduce noise, save computational cost and learn interpretable models. Due to the multi-modality nature of heterogeneous data, it is interesting to design efficient machine learning models that are capable of performing variable selection and feature group (data source) selection simultaneously (a.k.a bi-level selection). In this thesis, I carry out research along this direction with a particular focus on designing efficient optimization algorithms. I start with a unified bi-level learning model that contains several existing feature selection models as special cases. Then the proposed model is further extended to tackle the block-wise missing data, one of the major challenges in the diagnosis of Alzheimer's Disease (AD). Moreover, I propose a novel interpretable sparse group feature selection model that greatly facilitates the procedure of parameter tuning and model selection. Last but not least, I show that by solving the sparse group hard thresholding problem directly, the sparse group feature selection model can be further improved in terms of both algorithmic complexity and efficiency. Promising results are demonstrated in the extensive evaluation on multiple real-world data sets.