Predicting from Aggregated Data

Jedh, Maaz

Permalink

https://hdl.handle.net/2142/125115 Copy

Description

Title

Predicting from Aggregated Data

Author(s)

Jedh, Maaz

Issue Date

2020-12-01

Keyword(s)

Mixture of Experts, Data Aggregation

Abstract

Aggregated data, which refers to a collection of data summarized from multiple sources, is a technique commonly used in different fields of research including healthcare, web applications, and sensor networks. Aggregated data is often employed to handle issues such as privacy, scalability, and reliability. However, accurately predicting individual outcomes from grouped datasets can be very difficult. In this thesis, we designed a new learning method, a Mixture of Expert (MoE) model, focused on individual-level prediction when training variables are aggregated. We utilized the MoE model, trained and validated using the eICU Collaborative Research patient datasets, to conduct a series of studies. Our results showed that applying grouping functions to the classification of aggregated data across demographic and behavior metrics could remain effective. This technique was verified by comparing two separately trained MoE models that were evaluated on the same datasets. Finally, we estimated non-aggregated datasets from spatio-temporal aggregated records by expressing the problem in the frequency domain, and trained an autoregressive model for predicting future stock prices. This process can be repeated, offering a potential solution to the issue of learning from aggregated data.

Type of Resource

text

Language

eng

Owning Collections