Human-Centric AI: Interactive ML Systems

A collection of five interactive Django web applications exploring human-centric machine learning paradigms, developed as coursework at TU Hamburg.

Projects Overview

Project	Topic	Key Technique
1	AutoML Interface	Supervised Learning Pipeline
2	Active Learning	Pool-Based Sampling for Text
3	Explainability	Sparse Models & Counterfactuals
4	Recommender Systems	Matrix Factorization + Cold Start
5	RLHF	Reinforcement Learning with Human Feedback

---

Project 1: Automated Machine Learning Interface

Interactive web interface for end-to-end supervised learning workflows.

Features

• Data Upload & Visualization: CSV upload with scatter plot generation (X vs Y, Feature vs Target)
• Automatic Problem Detection: Infers classification vs regression from target column
• Model Selection: Logistic Regression, SVM, Random Forest, Linear Regression
• Hyperparameter Tuning: Grid search with user-configurable parameter ranges
• Evaluation: Confusion matrix (classification) or True vs Predicted plots (regression)

Technical Implementation

CSV → Preprocessing (StandardScaler, OneHotEncoder) → GridSearchCV → Model → Evaluation

Models: LogisticRegression, SVC, RandomForestClassifier, LinearRegression, RandomForestRegressor

---

Project 2: Active Learning for Text Classification

Pool-based active learning for IMDB sentiment analysis (50k movie reviews).

Features

• Text Representation: TF-IDF vectorization (5000 features)
• Baseline Classifier: Logistic Regression trained on full dataset (~89% accuracy)
• Active Learning Strategies:
  • Least Confident
  • Margin Sampling
  • Entropy-based
  • Random Baseline
  • Mixtures (LC+Entropy, Entropy+Random)
• Termination Conditions: Plateau detection, target accuracy threshold

Active Learning Progress

Technical Implementation

# Utility functions for sample selection
def calculate_utility_scores(probas, strategy):
    if strategy == "least_confident":
        return 1 - np.max(probas, axis=1)
    elif strategy == "entropy":
        return -np.sum(probas * np.log2(probas + 1e-12), axis=1)

---

Project 3: Explainability & Interpretable ML

Interactive exploration of model interpretability using the Palmer Penguins dataset.

Features

• Sparse Decision Trees: GOSDT algorithm for optimal sparse trees
• Sparse Logistic Regression: L1 penalty with adjustable regularization
• Complexity Control: Lambda slider to trade off accuracy vs interpretability
• Counterfactual Explanations: "What-if" analysis using MAD-weighted L1 distance

Technical Implementation

Interpretability Objective: argmin_f (1/n)Σℓ(f(xi), yi) + λΩ(f)

Counterfactual Generation:

1. Sample N points locally around input x
2. Filter to points with desired class prediction
3. Rank by MAD-weighted L1 distance
4. Display top-k counterfactuals

---

Project 4: Cold-Start Recommender with Guided Active Learning

Interactive movie recommendation system addressing the cold-start problem using MovieLens dataset.

Features

• Matrix Factorization: L2-regularized latent factor model
• Active Learning Strategies:
  • Popularity-based (most-rated movies)
  • Uncertainty-based (ratings closest to 3.0)
• Guidance System: Real-time impact preview showing how ratings affect recommendations
• User Study Interface: Complete A/B testing framework with logging

Technical Implementation

Matrix Factorization: min_{U,V} Σ(R_ij - U_i^T V_j)² + λ(||U||²_F + ||V||²_F)

New User Embedding: min_{u} Σ(R_ij - u^T V_j)² + λ||u||²_F  (V fixed)

Key Components

• Impact calculation showing prediction deltas before/after rating
• What-if preview panel for strategic user input
• PDF report generation for methodology documentation

---

Project 5: Reinforcement Learning with Human Feedback (RLHF)

Train a mouse agent to collect cheese using human preference feedback.

Environment

• 5×5 grid world with walls, traps, cheese (regular + organic)
• Rewards: +10 (cheese), -50 (trap), -0.2 (empty/wall bump)
• 4 actions: Up, Down, Left, Right

Features

• REINFORCE Algorithm: Policy gradient with CNN policy network
• Bradley-Terry Preference Model: Learn rewards from human trajectory comparisons
• RLHF Training: Fine-tune policy using learned reward with KL penalty

Architecture

# Policy Network (from project specification)
PolicyNetwork:
    Conv2d(6→16, 3×3) → ReLU → Conv2d(16→32, 3×3) → ReLU
    → Flatten → Linear(800→64) → ReLU → Linear(64→4) → Softmax

# Reward Network
RewardNetwork:
    Conv2d(6→16, 3×3) → ReLU → Flatten → Linear(400→64) → ReLU → Linear(64→1)

RLHF Pipeline

1. Train base policy with REINFORCE
2. Generate trajectory pairs
3. Collect human preferences (which trajectory is better?)
4. Train Bradley-Terry reward model: P(τ₁ ≻ τ₂) = σ(R(τ₁) - R(τ₂))
5. Fine-tune policy with learned reward + KL penalty to original policy

---

Tech Stack

Django Python scikit-learn PyTorch NumPy Pandas Matplotlib TF-IDF NLTK

Installation

# Clone repository
git clone https://github.com/rahulkvr/hcai.git
cd hcai

# Option A: Conda
conda env create -f environment.yml
conda activate hcai

# Option B: pip
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

# Run server
python manage.py runserver

Access at http://127.0.0.1:8000/

---

TU Hamburg — Human-Centric AI Course