Ranking Algorithms
Introduction
Section titled “Introduction”JustAI uses machine learning algorithms to optimize content selection in real time. Our platform tackles the exploration–exploitation tradeoff with two complementary approaches: Weighted Thompson Sampling for statistical optimization and Disjoint Linear Thompson Sampling for context-aware decisions. Together, they adaptively balance discovering new high-performing content with maximizing the impact of proven winners—at scale and in production.
The Challenge
Section titled “The Challenge”Content optimization is fundamentally a multi-armed bandit problem:
- Exploitation: Serve the best-known variant based on historical performance.
- Exploration: Test less-proven or new variants to find future winners.
Traditional A/B testing fixes traffic splits and requires long test periods, slowing down learning and missing short-term opportunities. Bandit algorithms continuously reallocate traffic based on live performance data, automatically adapting as conditions change.
The challenge intensifies when:
- Cold Starts: New variants have no history.
- Context Sensitivity: A winning subject line for enterprise customers may underperform for SMBs.
- Temporal Dynamics: User preferences shift over time.
Weighted Thompson Sampling
Section titled “Weighted Thompson Sampling”Our Weighted Thompson Sampling Scorer models each variant’s conversion rate using Bayesian inference.
How it works:
- Beta Distribution Modeling: Successes and failures feed a Beta(α, β) distribution.
- Optional Bias Correction: Median-based adjustments normalize across variants.
- Multi-Metric Weighting: Combine open rates, click rates, and more into a single score.
- Graceful Degradation: Neutral priors handle sparse data without skew.
Enabled by default, but best for campaigns with known segments.
Disjoint Linear Thompson Sampling
Section titled “Disjoint Linear Thompson Sampling”Our Disjoint Linear Thompson Sampling Scorer is a contextual bandit that learns how user and request features affect performance.
How it works:
- Feature Hashing: Compress request attributes into a high dimensional vector.
- Per-Variant Linear Models: Reward ≈ θᵀx + noise.
- Bayesian Updates: Maintain uncertainty with normal-inverse-gamma priors.
- Contextual Scoring: Sample θ from the posterior, score = θᵀx.
Benefits:
- Automatic Segmentation: Learns that different audiences prefer different content.
Best for campaigns with high cardinality ranking features and no upfront segmentation.