Improving machine learning algorithms via efficient data relevance discovery

2018-01-18This is the era of big data, where both challenges and opportunities lie ahead for the machine learning research. The data are created nowadays at an unprecedented pace with a significant cost in collecting, storing, and computing with the current scale of data. As the computational power that we possess gradually plateaus, it is an ever-increasing challenge to fully utilize the wealth of big data, where better data reduction techniques and scalable algorithms are the keys to a solution. We observe that to answer a certain query, the data are not equally important: it is possible to complete the task at hand with fewer but relevant data without compromising accuracy. Based on the models and the query, we provide efficient access to the numerical scores of the data points that represent their relevance to the current task. It enables us to wisely devote the computation resources to the important data, which improves the scalability and the reliability. ❧ We present our work with three applications. We developed efficient numerical algorithms for tensor CP decomposition and random-walk matrix-polynomial sparsification, where we provide an efficient access to the statistical leverage scores for a faster randomized numerical routine. We obverse that the model structures, such as the low-rank models, can be leveraged to provide an efficient approximation to data relevance. We also present our matrix completability analysis framework based on both data and the model, which are both structured. And we show that the underlying completability pattern can be utilized to achieve a more reliable estimation of missing entries