From corpus to slate: retrieve, score, re-rank
Diagnose recommendation quality stage by stage: coverage first, ordering second, slate policy third.
Marketplace ranking learns from biased clicks and reinforces its own mistakes.
This chapter moves from retrieve for coverage to construct the slate. We will make the idea concrete in a small code example before placing it back inside a production system with 80 ms ranking budget.
Filter unavailable items, enforce hard policy rules, and greedily balance relevance with diversity under a declared latency budget.
Verify the behavior illustrated in staged_slate.py, then measure decision utility, slice reliability, and system latency.
A random interaction split and click-only labels improve offline NDCG while amplifying already-prominent items. An unsafe release must degrade to a named baseline or the last known-good version.
Build the idea one layer at a time.
A recommender does not leap from millions of items to ten choices in one step. It narrows the problem: retrieve a generous candidate set, score it with richer context, then apply slate-level rules before anything reaches a user.
Retrieval optimizes coverage under a strict latency budget and may combine lexical, embedding, popular, and recent-item sources. Ranking then uses features too expensive or contextual for the index. Re-ranking makes constraints such as availability, policy, diversity, and creator caps explicit across the whole slate.
Each stage has a different contract. Candidate recall upper-bounds ranking quality; ranker metrics cannot diagnose missing inventory; a policy layer may intentionally lower raw relevance to improve coverage or safety. Log source, score, version, and rejection reason so the final slate remains explainable.
- candidate generation
- implicit feedback
- exposure bias
- two-tower embeddings
Follow the reasoning before touching the code.
- 01
Retrieve for coverage
Union several complementary candidate sources and measure whether known relevant items appear before tuning their order.
- 02
Score with shared context
Compute one comparable ranking score from user, item, request, and cross features while respecting an online feature cutoff.
- 03
Construct the slate
Filter unavailable items, enforce hard policy rules, and greedily balance relevance with diversity under a declared latency budget.
candidates = {
'lexical': [('a', 'books'), ('b', 'games')],
'similar': [('b', 'games'), ('c', 'books'), ('d', 'music')],
}
features = {'a': .74, 'b': .91, 'c': .83, 'd': .70}
pool = {item: category for source in candidates.values() for item, category in source}
ranked = sorted(pool, key=features.get, reverse=True)
slate, counts = [], {}
for item in ranked:
category = pool[item]
if counts.get(category, 0) < 1:
slate.append(item); counts[category] = counts.get(category, 0) + 1
print(slate)Union candidates, score them, then enforce a category cap.
What to notice
- Deduplication happens before ranking, so item b does not consume two candidate positions merely because two sources found it.
- The category cap changes the score-sorted order intentionally; this is a policy outcome, not a ranker bug.
YOUR TURN Pause and predict before revealing the answer.
Remove item d from every candidate source. Can a perfect ranker place it in the slate?
Need a hint?
A ranker can only order the pool it receives.
Check your reasoning
No. Retrieval recall is a hard ceiling: a missing candidate cannot be recovered by scoring or re-ranking.
Evaluation, Data & features, Unsupervised & retrieval
Connect the chapter to the course’s quantitative spine.
Diagnose recommendation quality stage by stage: coverage first, ordering second, slate policy third. The discounted cumulative gain below is the shared course-level lens that helps quantify the surrounding system. Name each symbol, check its units or shape, and use only values that were available at the production cutoff.
Discounted cumulative gain
| Symbol | Meaning / shape / unit |
|---|---|
rel_r | relevance at rank r |
r | one-indexed rank |
K | slate cutoff |
Open derivation and numerical substitution
Start from the production quantity being optimized, substitute the observed values with their declared units, then isolate the model-controlled term. Preserve shape annotations at each step so broadcasting or aggregation cannot silently change the result.
- Write the named inputs: rel_r, r, K.
- Substitute one small, hand-checkable batch before vectorizing.
- Calculate an independent reference value and compare within a declared tolerance.
# equation → code contract
inputs = validate_shapes_and_units(batch)
value = compute_c08(inputs)
assert is_finite(value)Change one assumption and make the tradeoff visible.
Use the worked chapter as your prediction, then change one input in the lab and explain why the output moves. The lab runs predefined TypeScript and never executes learner code, so the focus stays on the concept rather than setup.
Exploration feedback loop
Allocate unbiased exploration and observe learning coverage versus short-term utility.
Assumption: Simplified marketplace with position-biased clicks and fixed demand.
Open nonvisual data table
| Item | Computed state | Interpretation |
|---|---|---|
| rank 1 | 35.0% | 6.2% |
| rank 2 | 30.3% | 6.9% |
| rank 3 | 25.6% | 7.5% |
| rank 4 | 20.8% | 8.2% |
| rank 5 | 16.1% | 8.8% |
Trace the complete operating path.
Carry the chapter’s central idea into the system: Diagnose recommendation quality stage by stage: coverage first, ordering second, slate policy third. The path below shows where that reasoning needs an artifact, a test, an owner, or a fallback.
- 01
Turn “Retrieve for coverage” into a written invariant and a failing test.
- 02
Promote staged_slate.py from illustration to a versioned implementation with representative fixtures.
- 03
Use “Construct the slate” as the release checkpoint: Filter unavailable items, enforce hard policy rules, and greedily balance relevance with diversity under a declared latency budget.
- 04
Serve or materialize behind explicit 80 ms ranking budget.
- 05
Join telemetry to mature outcomes and retain a rollback path.
Observability
Join service health, input quality, prediction distributions, slice behavior, and mature outcomes by exact version.
Cost
Measure storage, preprocessing, compute, queueing, and human review under a representative arrival pattern.
Failure modes
A random interaction split and click-only labels improve offline NDCG while amplifying already-prominent items. Add a detector, owner, mitigation, and stop condition for this class of failure.
Alternatives
Compare a rule, a simpler statistical baseline, and a different system boundary before adding model complexity.
Explain the contract, not just the vocabulary.
Hybrid catalog ranker
Specify BM25 plus learned retrieval, ranking, re-ranking, temporal evaluation, and a versioned online path.
- Exposure-aware data contract
- Candidate recall analysis
- Ranker and NDCG report
- Diversity/fallback/latency plan
Read primary material with a purpose.
Return to the opening failure.
Version index, ranker, features, and catalog together; log exposures and measure each stage separately.
For this lesson, the release evidence is the worked code example, a hand-checked formal result, deterministic simulation output, a ≥80% checkpoint, the production rubric, and a named fallback. The course resolves when the system can produce candidate generation, ranking metrics, debiasing, and online experiment design.