Discovering Archetypes to Interpret Evolution of Individual Behavior

TL;DR

This paper introduces a Gaussian Hidden Markov Model cluster to identify archetypical patterns of individual evolution in social networks, revealing insightful behavioral archetypes and differences across datasets and demographics.

Contribution

The paper proposes a novel G-HMM cluster method for discovering behavioral archetypes and demonstrates its effectiveness in social network datasets with improved prediction accuracy.

Findings

Identified four archetypes for researchers: Steady, Diverse, Evolving, Diffuse.

Discovered archetypes for StackOverflow: Experts, Seekers, Enthusiasts, Facilitators.

Improved session prediction by 24-32% and achieved lower perplexity than baselines.

Abstract

In this paper, we aim to discover archetypical patterns of individual evolution in large social networks. In our work, an archetype comprises of $\textit{progressive stages}$ of distinct behavior. We introduce a novel Gaussian Hidden Markov Model (G-HMM) Cluster to identify archetypes of evolutionary patterns. G-HMMs allow for: near limitless behavioral variation; imposing constraints on how individuals can evolve; different evolutionary rates; and are parsimonious. Our experiments with Academic and StackExchange dataset discover insightful archetypes. We identify four archetypes for researchers: $\textit{Steady}$, $\textit{Diverse, Evolving and Diffuse}$. We observe clear differences in the evolution of male and female researchers within the same archetype. Specifically, women and men differ within an archetype (e.g. Diverse) in how they start, how they transition and the time spent in mid-career. We also found that the differences in grant income are better explained by the differences in archetype than by differences in gender. For StackOverflow, discovered archetypes could be labeled as $\textit{Experts, Seekers, Enthusiasts, and Facilitators}$. We have strong quantitative results with competing baselines for activity prediction and perplexity. For future session prediction, the proposed G-HMM cluster model improves by an average of $32\%$ for different Stack Exchanges and $24\%$ for Academic dataset. Our model also exhibits lower perplexity than the baselines.