AI Research Advances: Multimodal Modeling of Longitudinal Data and Detection of Synthetic E‑Commerce Reviews

A doctoral dissertation from the University of Nebraska‑Lincoln presents a multimodal research program for building generalizable and reliable models of longitudinal experiential (LE) data. The work progresses from traditional machine‑learning baselines through text‑only language models, vision‑language models, and finally a frozen‑backbone multimodal framework, achieving out‑of‑distribution accuracy improvements from 51 % to nearly 80 % on benchmark tasks. Key design elements include representation engineering, contextual reasoning, missingness‑aware learning, and constraints that reduce shortcut learning, aiming for stable predictions when data are incomplete or distributionally shifted.

A separate thesis introduces the SAFE (Suspicious vs. Authentic Feedback Evaluation) framework to detect large‑language‑model‑generated reviews on Amazon’s e‑commerce platform. By assembling a high‑confidence human‑review subset using metadata such as verified purchase status and helpfulness votes, the study generates synthetic reviews with GPT‑4, LLaMA‑3, and Mistral under various prompting strategies. Experiments with six detectors, including GLTR, DetectGPT, and the proposed SAFE model, show strong detection performance across product categories and even when synthetic reviews are conditioned on real user personas, indicating persistent statistical signatures that differentiate AI‑generated text from genuine human writing.