Skip to main content

​
What is Bayesian Knowledge Tracing?

Bayesian Knowledge Tracing (BKT) is a probabilistic framework that models student knowledge as a hidden state. Every answer provides evidence about the student’s true understanding, allowing us to update our beliefs about their mastery in real-time.

​
The Four BKT Parameters

​
How BKT Works

​
Step 1: Prior Belief

When we first encounter a skill, we initialize mastery with P(L0): 
initial_mastery = 0.25  # We assume 25% chance student knows the skill

​
Step 2: Posterior Update

After each answer, we update our belief using Bayesian inference: 
def bayesian_update(prior, is_correct, p_guess, p_slip):
    if is_correct:
        # Student got it right
        likelihood = (prior * (1 - p_slip)) + ((1 - prior) * p_guess)
        posterior = (prior * (1 - p_slip)) / likelihood
    else:
        # Student got it wrong
        likelihood = (prior * p_slip) + ((1 - prior) * (1 - p_guess))
        posterior = (prior * p_slip) / likelihood
    return posterior

​
Step 3: Learning Transition

We account for potential learning during the attempt: 
def apply_learning_transition(posterior, p_learn):
    # The student might have learned after seeing the problem
    return posterior + (p_learn * (1 - posterior))

​
Implementation Example

from cognition_engine import CognitionEngine

engine = CognitionEngine(supabase_url, supabase_key)

# Track an answer
result = await engine.track_answer(
    user_id="student_123",
    skill_id="algebra_linear",
    is_correct=True,
    time_spent_seconds=45,
    confidence_score=4
)

print(f"Mastery before: {result['mastery_before']:.3f}")
print(f"Mastery after: {result['mastery_after']:.3f}")
print(f"Velocity: {result['velocity']:.4f}")

​
Key Features

​
Adaptive Parameters

BKT parameters can be customized per skill based on domain characteristics: 
# Math problems might have different parameters than reading
math_params = {
    "prior_knowledge": 0.20,      # Lower initial assumption
    "learn_rate": 0.12,           # Faster learning
    "guess_probability": 0.15,    # Lower guessing
    "slip_probability": 0.08       # Lower mistakes
}

​
Plateau Detection

The engine automatically detects learning plateaus: 
if result['plateau_detected']:
    print("Learning plateau detected - intervention recommended")
    # Trigger personalized help or change difficulty

​
Velocity Tracking

Learning velocity measures the rate of mastery improvement: 
velocity = result['velocity']
if velocity > 0.1:
    print("Accelerating - keep pushing!")
elif velocity < -0.05:
    print("Declining - consider review")

​
Cognitive Efficiency Integration

BKT mastery probability is enhanced with cognitive efficiency metrics:
Students who answer quickly with high confidence should show different mastery than those who struggle through guesses.
Self-reported confidence (1-5) provides additional signal about true understanding beyond correctness alone.
Answer changes and time spent provide insights into cognitive effort and uncertainty.

​
BKT vs. Traditional Tracking

​
Best Practices

Avoid overfitting parameters to small datasets. Use default parameters for most skills and only customize when you have substantial data.
Combine BKT mastery probability with velocity analysis and cognitive efficiency for the most complete picture of student understanding.
BKT works best when students attempt at least 5-10 problems per skill. Below this threshold, mastery estimates are less reliable.

​
BKT in Our Multi-Model Architecture

BKT is the foundation of our analytics engine, but we combine it with advanced models:
BKT provides the “what” - interpretable mastery tracking for dashboards and analytics
For deeper insights, we also use:
  • SAKT (Self-Attentive KT): Attention weights show cognitive patterns
  • DKT-Forget: Models the forgetting curve for spaced repetition
  • DTransformer: Dynamic models that account for time gaps and spacing effects
See our Technical Deep Dive guide for details on how we combine BKT with these advanced models.

​
Research Background

BKT is based on research from Carnegie Mellon University’s Human-Computer Interaction Institute. Our implementation extends the classic model with:
  • Cognitive efficiency integration: Time and confidence weighting
  • Plateau detection: Automatic intervention triggers
  • Velocity analysis: Momentum and acceleration tracking
  • Adaptive parameters: Skill-specific tuning

​
Next Steps

Learning Velocity

Understand how we analyze learning momentum

Cognitive Efficiency

Learn about multi-dimensional efficiency metrics