Fake Internship Detector

Detecting fraudulent job postings using NLP and Machine Learning

app.demo.mp4

---

📖 Overview

The Fake Internship Detector is an NLP-based machine learning project designed to identify fraudulent job and internship postings. With the rise of online job scams targeting students, this project aims to make the hiring space safer by analyzing the language and writing patterns of job descriptions.

The model learns to differentiate between genuine and fake postings using both textual and behavioral signals extracted from real job data.

---

🎯 Objectives

• Detect fraudulent job/internship postings automatically
• Use NLP to analyze textual patterns in job descriptions
• Combine linguistic and behavioral features for improved accuracy
• Build a robust model capable of identifying scams with high reliability

---

🧩 Dataset

• Source: Open-source job posting dataset (such as Kaggle – Real or Fake Job Posting Prediction)
• Total records: ~18,000
• Target variable: fraudulent → 0 = Genuine, 1 = Fake
• Notable challenge: Class imbalance (~95% genuine, 5% fake)

---

⚙️ Project Workflow

🪜 Step 1: Data Cleaning & Preprocessing

• Combined multiple text columns (title, description, requirements, etc.) into one field
• Removed URLs, punctuation, numbers, and special characters using regex
• Converted all text to lowercase
• Applied stopword removal and lemmatization (using NLTK)
• Handled missing values and irrelevant columns

📊 Step 2: Exploratory Data Analysis (EDA)

• Visualized class imbalance (95% genuine vs. 5% fake)
• Generated missing value heatmaps, word clouds, and text length distributions
• Observed that fake posts were shorter, repetitive, and used flashy terms like “money”, “urgent”, “work from home”
• 

⚙️ Step 3: Feature Engineering

Added numerical indicators to strengthen the model:

• num_words
• num_unique_words
• num_chars
• avg_word_len
• num_exclamations, num_question_marks, num_uppercase
• 

🧠 These features complement NLP features by capturing stylistic and behavioral patterns.

🧠 Step 4: Text Vectorization

• Used TF-IDF Vectorizer to transform text into numerical features
• Parameters:
  • max_features = 5000
  • ngram_range = (1, 2) → captures both single words and bigrams
  • stop_words = 'english'

🔀 Step 5: Data Balancing

• Handled severe class imbalance using SMOTE (Synthetic Minority Oversampling Technique)

🤖 Step 6: Model Training & Evaluation

Trained and compared multiple models:

• Logistic Regression
• Random Forest
• SVM

Evaluation Metrics:

• Accuracy
• Precision
• Recall
• F1-Score (primary metric)

🎯 Final Model Performance:
F1-Score ≈ 80% (balanced performance between precision & recall)

📈 Step 7: Visualization

• Used Matplotlib, Seaborn, and Plotly for interactive and comparative insights
• Plotted feature distributions, confusion matrices, and ROC curves
• 

---

🧰 Tech Stack

Category	Tools & Libraries
Language	Python
Data Handling	Pandas, NumPy
Visualization	Matplotlib, Seaborn, Plotly
NLP	NLTK, Scikit-learn (TF-IDF)
Modeling	Logistic Regression, Random Forest, SVM
Balancing	SMOTE
Environment	Jupyter Notebook / Google Colab

---

🚀 Results

• Achieved ~80% F1-Score on test data
• Significantly improved recall (caught more fake postings)
• Successfully identified linguistic and behavioral patterns unique to fraudulent ads

---

📦 Project Structure

app/ dashboard.py model/ random_forest_model.joblib tfidf_vectorizer.joblib notebooks/ data_preprocessing_eda_baseline.ipynb src/ pycache/ preprocessing.cpython-312.pyc model.py preprocessing.py

---

🔍 Future Work

• Integrate a REST API for job posting platforms
• Experiment with transformer-based NLP models (e.g., BERT, RoBERTa)
• Expand dataset for multilingual job postings

---

🏆 Key Takeaways

• Combined NLP + Feature Engineering for hybrid modeling
• Learned to handle imbalanced datasets effectively
• Improved skills in text preprocessing, model tuning, and data visualization
• Created a system that can genuinely help users avoid online fraud

---

📫 Connect With Me

💼 LinkedIn
💻 GitHub
📧 [email protected]

---

⭐ If you liked this project, consider giving it a star on GitHub!