Hierarchical regression model tree for explainable actor segmentation and response prediction on social networks

Li, Jinning

Permalink

https://hdl.handle.net/2142/117823 Copy

Description

Title

Hierarchical regression model tree for explainable actor segmentation and response prediction on social networks

Author(s)

Li, Jinning

Issue Date

2022-12-05

Director of Research (if dissertation) or Advisor (if thesis)

Abdelzaher, Tarek

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Social Network Analysis
• Regression Model Tree
• Response Prediction
• Machine Learning

Abstract

Social network systems have produced large-scale data of social signals. However, the potential mechanism of social signal propagation and how it affects people's beliefs and responses are still not well investigated. In this project, we propose a framework and an explainable Hierarchical Regression Model Tree (HRMT) algorithm to solve the individual-level and segmentation-level response prediction tasks and therefore provide the solution to analyze how people's morality, demographics, and other psychographic characteristics affect their beliefs and response to the social information influence. We develop a text-based actor enrichment prediction module based on the Bidirectional Encoder Representations from Transformers (BERT) language model and predict the message enrichment with a weakly-supervised topic detection model. The Hierarchical Regression Model Tree is constructed with regression-error greedy search and reliability test algorithms and then used to construct the segments of actors based on tree structure and predict future responses. These results can be applied for many downstream researches and tasks, such as sociological analysis, influence campaign detection, advertisement, and recommender systems. We also proposed two novel evaluation metrics, normalized segment Discounted Cumulative Gain (nsDCG) and invariant nsDCG. Experimental evaluations show the proposed HRMT outperforms the state-of-the-art models by 0.12 in the nsDCG metrics. We also introduce the application of HRMT in analyzing the characteristics of actors' beliefs based on the tree structure.

Graduation Semester

2022-12

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Jinning Li

Owning Collections