Expressive and Interpretable User Engagement Prediction using Multivariate Survival Processes

Author

Akshay Aravamudan, Florida Institute of TechnologyFollow

Date of Award

5-2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical Engineering and Computer Science

First Advisor

Georgios C. Anagnostopoulos

Second Advisor

Anthony O. Smith

Third Advisor

Adrian M. Peter

Fourth Advisor

Nezamoddin Nezamoddini-Kachouie

Abstract

The ability to characterize how information diffuses online is of paramount importance to stakeholders that are interested in tasks such as proposing solutions for mitigating and countering dis/misinformation, predicting user engagement of content in social media, planning marketing campaigns to roll-out products and planning dissemination of political campaign messaging among others. One such facet of learning the dynamics of information diffusion is the ability to predict user engagement or the popularity of a single piece of information as it spreads through an online medium. Existing works in this regard mainly either obfuscate user level information or utilize frameworks that are difficult to interpret underlying user behaviors. Survival analysis is a powerful statistical approach that can be used to model information diffusion by taking advantage of elegant representations of underlying probability distributions. We utilize this probabilistic framework to propose a novel model for user engagement prediction that considers unobserved heterogeneity in user behaviors. In addition to this, we propose a novel deep learning framework to predict user engagement by modeling underlying functions in survival analysis using expressive neural networks while retaining the rich interpretable aspects of survival analysis. We also propose two new variants of a discriminative loss function that is designed to train a single model for arbitrary observation periods and forecast horizons. We show via experiments that these discriminative loss functions perform better than traditional generative loss (log-likelihood) when data is scarce. We also describe how our proposed approaches retain the rich interpretable framework of point processes by exploring how we can infer Granger-causal relationships among diffusion actors. We then illustrate how these estimates can be used to extract meaningful insights from real-world data.

Recommended Citation

Aravamudan, Akshay, "Expressive and Interpretable User Engagement Prediction using Multivariate Survival Processes" (2025). Theses and Dissertations. 1593.
https://repository.fit.edu/etd/1593