Data Collection Lab Project: LinkedIn Profile Picture Evaluator

LinkedIn Profile Picture Evaluator

Yuval Margolin | Ravid Gersh | Daniel Maor
Technion - Israel Institute of Technology
Data and Decision Science Faculty

User Journey Video | Technical Overview Video

---

Contents

• Overview
• Scraping Pictures
• Preprocessing Images
• Running the Train and Test Code
• Download Model Weights
• Running The App

---

Overview

In today's professional world, first impressions matter—and a LinkedIn profile picture plays a key role! As part of the Data Collection Lab at the Technion, we built a LinkedIn Profile Image Evaluator 🏆

🔹 What does it do?

Our app analyzes a profile image and provides:

✅ A rating (1-100) based on key visual and professional factors
✅ Three personalized suggestions to enhance the picture

The goal? Helping users make data-driven choices about their professional presence on LinkedIn!

🔍 Technologies Used:

🖥️ MediaPipe & dlib – Face detection & analysis
🤖 OpenAI CLIP – AI-powered image understanding
🔥 PyTorch TabNet – Machine learning model training
⚡ PySpark – Efficient large-scale data processing
📷 OpenCV – Image preprocessing

---

Scraping Pictures

The scraping notebook uses the Google Search API (SerpAPI) to extract profile images from corporate websites. It automates retrieving image URLs by searching for professional profile pictures from company "Our Team" pages.

---

Preprocessing Images

The image analysis pipeline was executed on Azure Databricks, using a notebook named "PreProcess images and features", with data stored in dbfs/FileStore.

Preprocessing Steps:

• Unzipped multiple image archives and merged them into a single directory (/dbfs/FileStore/all_images_combined).
• Ensured no duplicates by renaming files with identical names.
• Labeled images with a score (1-100) and extracted 29 features using CLIP.
• Saved the extracted features into a CSV file (Profile_Pictures_extracted.csv).

Feature Extraction Techniques:

• CLIP: Analyzed images for professional attributes like smile, attire, background clarity, and trustworthiness.
• MediaPipe: Detected facial landmarks and body proportions, ensuring proper face positioning.
• Dlib: Analyzed nose and eye alignment for direct eye contact assessment.
• OpenCV: Evaluated image resolution, lighting quality, and clarity.

---

Running the Train and Test Code

The Train and Test Notebook loads the extracted dataset (Profile_pictures_extracted.csv) and performs:

• Exploratory Data Analysis (EDA): Correlation matrix to identify key feature relationships.
• Model Training: A machine learning model predicts LinkedIn scores from extracted image attributes.
• Model Evaluation: Performance assessed using MSE, R², and feature importance analysis.

This transforms raw image data into actionable insights and predictive analytics for profile evaluation.

---

Download LinkedIn Score Predictor Weights

You can download the pre-trained model weights from:

📥 linkedin_score_predictor.pth

To use it in your notebook, save it in the models folder in dbfs, ensuring:

input_dim = 28

✅ Model loading is already included in the train & test notebook and the app notebook.

---

Running The App

The LinkedIn Image Evaluator App allows real-time evaluation of LinkedIn profile pictures. You can run the evaluation using urls as input, or Alternatively add your own ngrok Authentication in the marked place to run the site.

App Features:

• Users can input image URLs or upload images for scoring.
• The trained model assesses key professional attributes like expression, attire, background, and overall appeal.
• Offers a web-based interactive experience for improving profile images.

🚀 Run the app and optimize your LinkedIn presence today! 🔥