PersonaAct

PersonaAct is a framework for generating personalized LLM-based agents from real user behavioral data to audit filter bubbles in short-video recommendation systems.

Overview

PersonaAct consists of three components:

1. Multimodal Dataset: Short-video user actions with video frames, audio transcripts, and interaction sequences
2. Interview Agent: Synthesizes user personas through behavioral analysis and targeted questioning
3. Persona-driven Agents: Deployed to audit filter bubble formation in recommendation systems

Installation

git clone https://github.com/Silung/PersonaAct.git
cd PersonaAct

# Install dependencies
pip install -r requeriment.txt
cd interview && pip install -r requirements.txt && cd ..

Quick Start

Data Preparation

python prepare_data.py --input_dir raw_data --output_dir data

Run Interview Agent

cd interview
python app.py

Inference

python infer.py --model_path <path_to_model> --data_path data/

Dataset

The dataset contains:

• 4,485 video interaction samples from 8 personas across 86 sessions
• Video frames (1 FPS), audio transcripts, and user actions
• 25+ content categories with metadata