Computational Pipeline: Visual Network Narrative Dissertation

MLOPs and LLMOPs for dissertation analysis workflow

Implementation

1. Download/clone the repository and save to your desired folder
2. Create a new virtual environment

Data Collection and Evaluation

• Web Scraping (Bot)
• Cleaning; Preparing and Monitoring
• Identify security and ethical risks in the data and storage

Resources

Instagram Account

Data Exploration (Clustering)

Model Development

Image Label and Caption Generation

Training and fine-tuning on a subset of data to track performance, identify errors, and optimize models.
Clean post comments i.e. lemmetize, translate emojies, rename hashtags, lowercase sentences, etc
Multlingual sentence transformers

• Huggingface sentence transformer
• LaBASE (Language Agnostic BERT sentence encoder)
• GCN?

Evaluation Metrics:

• Acuracy, Precision, Recall, F1-Score, Hamming Loss
• Confusion Matrix (Visuallization)
• BERTScore, Sentence Transformer Cosine Similarities

Ongoing monitoring of security and ethical risks

Resources

Hugging Face Access, Access to LLM (I used VERDE), GPU Access

Model Deployment

Identify Semantic similarities

Create Structural graph (content-based knowledge representation)

• Viz_weights + generated label_weights
• Viz_weights & generated label_weights + generated captions
• Viz_weights & generated label_weights + generated captions & original post comments

Community Detection (Evaluation of Network Structure):

• Centraility Measures
• ERGM algorithm
• Leidan algorithm

Graph Analysis

• Multipartite, Bipartite
• Multiplex Graphs?