Lesson 021: Calendar as Portfolio Optimization, Not Top-N Ranking

Lesson 021: Calendar as Portfolio Optimization, Not Top-N Ranking

The Lesson

When selecting a fixed-size collection where the items must work together (a calendar, a playlist, a portfolio, a menu), the problem is constrained set optimization — not top-N ranking. The best collection often contains none of the individually top-ranked items, because collection-level properties (diversity, coverage, balance) matter as much as individual quality.

Context

A project needed to select 13 images for a 13-month calendar from a pool of 12,217 Artemis II mission photos. Each image had a preference score from voter data. The naive approach — pick the 13 highest-scoring images — would likely produce a visually redundant calendar (e.g., 13 similar Earth-from-orbit shots) because the most popular image types cluster together in the embedding space.

What Happened

  1. Framed the problem initially as "rank images, take top 13." This was the simplest baseline (Method A) and served as a comparison point.
  2. Realized the calendar has collection-level requirements that top-N ranking ignores: visual diversity (no two images too similar), month fit (bright warm images for summer, dramatic images for December), mission coverage (launch, transit, lunar orbit, return), and a suitable cover image.
  3. Designed a multi-objective utility function that scores the entire 13-image set, not individual images: preference sum + diversity bonus + month-fit sum + cover score - redundancy penalty - uncertainty penalty.
  4. Implemented 5 selection methods: naive top-13 (baseline), cluster-limited top-13, best-per-cluster, month-first greedy, and multi-objective MMR greedy. Each produces a different 13-image slate.
  5. Ran all 5 methods on real data. Method A (naive top-13) and Method E (MMR greedy) shared 0 of 13 images — the optimization selected entirely different images than naive ranking.
  6. Method B (cluster-limited) scored highest on the objective function — it kept most of top-N's popularity while enforcing diversity through cluster constraints. Method C (per-cluster) achieved perfect diversity (1.0) but lower popularity.

Key Insights

Examples

Results from 12,217 images, k=25 visual clusters:

Method Objective Popularity Diversity Shared with Top-13
A: Naive top-13 14.19 4.32 0.77 13/13
B: Cluster-limited 14.26 4.32 0.85 11/13
C: Per-cluster 13.96 4.01 1.00 5/13
E: MMR greedy 11.89 2.79 0.77 0/13

Method E selected entirely different images — it aggressively penalizes redundancy via CLIP cosine similarity.

Applicability

This lesson applies to any "select K items from N" problem where the items interact:

Does NOT apply when items are independent (e.g., top-10 search results where each result is consumed individually).

Related Lessons