Entity Pairing¶

After selecting the first entity using UCB, Compere uses a multi-factor approach to select the second entity for comparison.

Pairing Strategy¶

The second entity is chosen using a composite score:

Factor	Default Weight	Purpose
UCB Score	30%	Exploration value
Rating Similarity	40%	Informative comparisons
Random	30%	Variety

How It Works¶

for entity in remaining_entities:
    score = 0

    # Factor 1: UCB score (exploration value)
    score += ucb_scores.get(entity.id, 0) * 0.3

    # Factor 2: Rating similarity (more informative comparisons)
    rating_diff = abs(entity1.rating - entity.rating)
    if rating_diff < 200:
        score += (200 - rating_diff) / 200 * 0.4

    # Factor 3: Randomness for variety
    score += random.random() * 0.3

Why Rating Similarity Matters¶

Comparing a 1600-rated entity vs 1580-rated entity is more informative than 1600 vs 1200:

Close ratings: Outcome is uncertain, more information gained
Distant ratings: Outcome is predictable, less information gained

The default threshold is 200 rating points.

Recent Comparison Exclusion¶

To prevent repetitive pairings, Compere excludes recent opponents:

recent_comparisons = db.query(Comparison).filter(
    (Comparison.entity1_id == entity1.id) |
    (Comparison.entity2_id == entity1.id)
).order_by(Comparison.created_at.desc()).limit(5).all()

# Exclude recently compared opponents
non_recent = [e for e in remaining if e.id not in recent_opponent_ids]

The last 5 opponents are excluded (when alternatives exist).

Configuration¶

Variable	Default	Description
`PAIRING_UCB_WEIGHT`	`0.3`	Weight for UCB score
`PAIRING_SIMILARITY_WEIGHT`	`0.4`	Weight for rating similarity
`PAIRING_RANDOM_WEIGHT`	`0.3`	Weight for random factor
`PAIRING_RATING_THRESHOLD`	`200.0`	Rating difference threshold
`RECENT_COMPARISON_LIMIT`	`5`	Recent comparisons to exclude

Tip

Weights should sum to 1.0 for consistent behavior.

Tuning Guidelines¶

More Exploration¶

Increase PAIRING_UCB_WEIGHT:

PAIRING_UCB_WEIGHT=0.5
PAIRING_SIMILARITY_WEIGHT=0.3
PAIRING_RANDOM_WEIGHT=0.2

More Informative Comparisons¶

Increase PAIRING_SIMILARITY_WEIGHT:

PAIRING_UCB_WEIGHT=0.2
PAIRING_SIMILARITY_WEIGHT=0.6
PAIRING_RANDOM_WEIGHT=0.2

More Variety¶

Increase PAIRING_RANDOM_WEIGHT:

PAIRING_UCB_WEIGHT=0.3
PAIRING_SIMILARITY_WEIGHT=0.3
PAIRING_RANDOM_WEIGHT=0.4

Similarity Matching Alternative¶

The /dissimilar_entities endpoint uses cosine similarity instead of the multi-factor approach:

Finds entities that are most dissimilar (different categories, characteristics)
Useful for ensuring diverse coverage
Currently uses placeholder embeddings (random vectors)

For production, replace with real embeddings from:

Text embeddings (OpenAI, Sentence Transformers)
Image embeddings (CLIP, ResNet)
Combined multimodal embeddings