tidaldb/docs/planning/milestone-3/phase-3/task-03-cold-start-and-exploration.md

# Task 03: Cold Start and Exploration

## Context

**Milestone:** 3 -- Personalized Ranking
**Phase:** m3p3 -- Personalized Ranking Profiles
**Depends On:** Task 02 (Personalized Profiles: `for_you` profile with `exploration: 0.1`), Task 01 (`UserContext::is_cold_start`), m2p4 (diversity enforcement, `DiversitySelector`)
**Blocks:** m3p4 (User State Filters + UAT need complete for_you behavior)
**Complexity:** M

## Objective

Deliver cold-start handling for new users and new items, plus the exploration budget injection for the `for_you` profile. These mechanisms prevent the personalization system from collapsing into a filter bubble or failing silently for users/items with no history.

**Cold-start users**: A new user with no engagement history has no preference vector. The `for_you` profile falls back to population-level signals: trending, quality (completion rate), and recency. This ensures new users see a reasonable feed on their first visit.

**Cold-start items**: New items with no signal history have no engagement data. Without intervention, they would never appear in personalized feeds, creating a chicken-and-egg problem. The exploration window gives new items a brief period (configurable, default 24h) where they are eligible for inclusion in the exploration budget of for_you feeds.

**Exploration budget**: The `for_you` profile reserves 10% of its result set (e.g., 5 of 50 items) for exploration candidates: items from creators the user does NOT follow. This prevents the feed from becoming a closed loop of familiar content and exposes users to new creators.

## Requirements

### Cold-Start User Fallback
- When `UserContext::is_cold_start` is true, `for_you` scoring uses population-level signals only
- Fallback formula: `trending_velocity * 0.4 + completion_rate * 0.3 + recency * 0.3`
- No preference match (no vector to compare against)
- No creator affinity (no interaction history)
- Candidate strategy: full corpus scan sorted by trending (not ANN, since no query vector)
- Diversity constraints still apply

### Cold-Start Item Exploration Window
- `ExplorationWindow` struct: items created within the last N hours (default 24h) with fewer than M signals (default 10) are eligible for exploration
- Eligible items are added to an `exploration_pool` bitmap
- The exploration budget draws from this pool when filling exploration slots
- After the exploration window expires, items must earn organic engagement to appear
- The window is configurable via `ExplorationConfig`

### Exploration Budget
- The `for_you` profile has `exploration: 0.1` (10%)
- For `LIMIT 50`, 5 items come from the exploration budget
- Exploration candidates are items NOT from followed creators
- Selection from the exploration pool: random shuffle, then score by population signals
- Exploration candidates are placed at positions throughout the result set (interleaved), not clustered at the end
- The remaining 90% (45 items) come from the standard personalized scoring

## Technical Design

### Module Structure

```
tidal/src/
  ranking/
    exploration.rs -- ExplorationConfig, ExplorationWindow, budget injection
```

### ExplorationConfig

```rust
// === ranking/exploration.rs ===

use std::time::Duration;

/// Configuration for cold-start and exploration behavior.
#[derive(Debug, Clone)]
pub struct ExplorationConfig {
    /// How long new items are eligible for exploration.
    /// Default: 24 hours.
    pub item_window: Duration,
    /// Maximum signal count for an item to be considered "cold start".
    /// Items with more signals than this are no longer in the exploration pool.
    /// Default: 10.
    pub max_signals_for_cold_item: u64,
    /// Fraction of results reserved for exploration.
    /// Default: 0.1 (10%).
    pub exploration_fraction: f64,
}

impl Default for ExplorationConfig {
    fn default() -> Self {
        Self {
            item_window: Duration::from_secs(24 * 3600),
            max_signals_for_cold_item: 10,
            exploration_fraction: 0.1,
        }
    }
}
```

### Cold-Start User Scoring

```rust
use crate::db::user_context::UserContext;
use crate::schema::{EntityId, Timestamp, Window};
use crate::signals::SignalLedger;

/// Score a candidate for a cold-start user.
///
/// Uses population-level signals only: trending velocity, completion rate,
/// and recency. No personalization factors.
pub fn cold_start_score(
    entity_id: EntityId,
    now: Timestamp,
    ledger: &SignalLedger,
    item_created_at: &dyn Fn(EntityId) -> Option<u64>,
) -> f64 {
    let view_vel = read_agg(entity_id, "view", &SignalAgg::Velocity, Window::TwentyFourHours, ledger);
    let share_vel = read_agg(entity_id, "share", &SignalAgg::Velocity, Window::TwentyFourHours, ledger);
    let trending = (view_vel + 2.0 * share_vel).min(1.0);

    let completion = read_agg(entity_id, "completion", &SignalAgg::DecayScore, Window::AllTime, ledger);
    let completion_rate = completion.min(1.0);

    let recency = item_created_at(entity_id)
        .map_or(0.5, |ts| recency_score(ts, now));

    trending * 0.4 + completion_rate * 0.3 + recency * 0.3
}
```

### Exploration Budget Injection

```rust
/// Inject exploration candidates into a personalized result set.
///
/// Takes the scored personalized results and injects exploration candidates
/// at regular intervals (interleaved, not appended).
///
/// # Parameters
///
/// - `personalized`: scored candidates from the personalized pipeline (sorted desc)
/// - `exploration_candidates`: candidates from unfollowed creators, scored by population signals
/// - `total_limit`: target result count (e.g., 50)
/// - `exploration_fraction`: fraction of results for exploration (e.g., 0.1)
///
/// # Returns
///
/// A merged result set with exploration candidates interleaved.
pub fn inject_exploration(
    personalized: &[ScoredCandidate],
    exploration_candidates: &[ScoredCandidate],
    total_limit: usize,
    exploration_fraction: f64,
) -> Vec<ScoredCandidate> {
    let exploration_count = ((total_limit as f64) * exploration_fraction).ceil() as usize;
    let personalized_count = total_limit.saturating_sub(exploration_count);

    let personalized_slice = &personalized[..personalized_count.min(personalized.len())];
    let exploration_slice = &exploration_candidates[..exploration_count.min(exploration_candidates.len())];

    if exploration_slice.is_empty() {
        // No exploration candidates available: return personalized only.
        return personalized[..total_limit.min(personalized.len())].to_vec();
    }

    // Interleave: place exploration candidates at regular intervals.
    let mut result = Vec::with_capacity(total_limit);
    let step = if exploration_slice.is_empty() {
        usize::MAX
    } else {
        total_limit / exploration_slice.len()
    };

    let mut p_idx = 0;
    let mut e_idx = 0;

    for i in 0..total_limit {
        if e_idx < exploration_slice.len() && i > 0 && i % step == 0 {
            result.push(exploration_slice[e_idx].clone());
            e_idx += 1;
        } else if p_idx < personalized_slice.len() {
            result.push(personalized_slice[p_idx].clone());
            p_idx += 1;
        } else if e_idx < exploration_slice.len() {
            result.push(exploration_slice[e_idx].clone());
            e_idx += 1;
        }
    }

    result
}

/// Select exploration candidates from the item corpus.
///
/// Exploration candidates are items NOT from followed creators.
/// They are scored by population-level signals and returned in
/// descending score order.
pub fn select_exploration_candidates(
    all_item_ids: &[EntityId],
    user_ctx: &UserContext,
    exploration_config: &ExplorationConfig,
    now: Timestamp,
    ledger: &SignalLedger,
    item_creator_lookup: &dyn Fn(EntityId) -> Option<EntityId>,
    item_created_at: &dyn Fn(EntityId) -> Option<u64>,
    limit: usize,
) -> Vec<ScoredCandidate> {
    let now_ns = now.as_nanos();
    let window_ns = exploration_config.item_window.as_nanos() as u64;

    let mut candidates: Vec<ScoredCandidate> = all_item_ids.iter()
        .filter(|&&item_id| {
            // Exclude items from followed creators.
            let creator = item_creator_lookup(item_id);
            let from_followed = creator
                .map_or(false, |c| user_ctx.followed_creators.contains(&c.as_u64()));
            !from_followed
        })
        .filter(|&&item_id| {
            // Exclude hidden items and items from blocked creators.
            if user_ctx.hidden_items.contains(&(item_id.as_u64() as u32)) {
                return false;
            }
            let creator = item_creator_lookup(item_id);
            if let Some(c) = creator {
                if user_ctx.blocked_creators.contains(&c.as_u64()) {
                    return false;
                }
            }
            true
        })
        .map(|&item_id| {
            let score = cold_start_score(item_id, now, ledger, item_created_at);
            ScoredCandidate {
                entity_id: item_id,
                score,
                signal_snapshot: vec![],
                creator_id: item_creator_lookup(item_id),
                format: None,
            }
        })
        .collect();

    // Sort by score descending.
    candidates.sort_by(|a, b| b.score.partial_cmp(&a.score).unwrap_or(std::cmp::Ordering::Equal));
    candidates.truncate(limit);
    candidates
}

/// Check if an item is within the cold-start exploration window.
pub fn is_cold_start_item(
    item_id: EntityId,
    now: Timestamp,
    config: &ExplorationConfig,
    item_created_at: &dyn Fn(EntityId) -> Option<u64>,
    total_signal_count: &dyn Fn(EntityId) -> u64,
) -> bool {
    let now_ns = now.as_nanos();
    let window_ns = config.item_window.as_nanos() as u64;

    // Created within window.
    let created = item_created_at(item_id).unwrap_or(0);
    if now_ns.saturating_sub(created) > window_ns {
        return false;
    }

    // Below signal threshold.
    total_signal_count(item_id) < config.max_signals_for_cold_item
}
```

## Test Strategy

### Unit Tests

```rust
#[test]
fn cold_start_score_uses_population_signals() {
    let ledger = test_ledger_with_signals();
    let entity = EntityId::new(1);
    let now = Timestamp::now();

    let score = cold_start_score(entity, now, &ledger, &|_| Some(now.as_nanos()));
    assert!(score > 0.0, "cold start score should be positive: {}", score);
    assert!(score <= 1.0, "cold start score should be <= 1.0: {}", score);
}

#[test]
fn inject_exploration_correct_count() {
    let personalized: Vec<ScoredCandidate> = (0..45)
        .map(|i| make_candidate(i + 1, (45 - i) as f64, Some(i as u64 + 1), None))
        .collect();
    let exploration: Vec<ScoredCandidate> = (0..5)
        .map(|i| make_candidate(i + 100, (5 - i) as f64, Some(i as u64 + 50), None))
        .collect();

    let result = inject_exploration(&personalized, &exploration, 50, 0.1);
    assert_eq!(result.len(), 50);

    // Count exploration items (IDs >= 100).
    let explore_count = result.iter().filter(|c| c.entity_id.as_u64() >= 100).count();
    assert_eq!(explore_count, 5, "should have 5 exploration items");
}

#[test]
fn inject_exploration_empty_exploration_pool() {
    let personalized: Vec<ScoredCandidate> = (0..50)
        .map(|i| make_candidate(i + 1, (50 - i) as f64, Some(1), None))
        .collect();
    let exploration: Vec<ScoredCandidate> = vec![];

    let result = inject_exploration(&personalized, &exploration, 50, 0.1);
    assert_eq!(result.len(), 50);
    // All items from personalized.
    assert!(result.iter().all(|c| c.entity_id.as_u64() <= 50));
}

#[test]
fn inject_exploration_interleaves() {
    let personalized: Vec<ScoredCandidate> = (0..45)
        .map(|i| make_candidate(i + 1, (45 - i) as f64, Some(1), None))
        .collect();
    let exploration: Vec<ScoredCandidate> = (0..5)
        .map(|i| make_candidate(i + 100, 1.0, Some(50), None))
        .collect();

    let result = inject_exploration(&personalized, &exploration, 50, 0.1);

    // Exploration items should be spread through the list, not all at the end.
    let first_half = &result[..25];
    let second_half = &result[25..];
    let explore_first = first_half.iter().filter(|c| c.entity_id.as_u64() >= 100).count();
    let explore_second = second_half.iter().filter(|c| c.entity_id.as_u64() >= 100).count();

    // At least one exploration item should be in each half.
    assert!(explore_first > 0 || explore_second > 0);
}

#[test]
fn select_exploration_excludes_followed_creators() {
    let user_state = UserStateIndex::new();
    let user = EntityId::new(1);
    user_state.add_follow(user, EntityId::new(10));

    let iw_ledger = InteractionWeightLedger::new(InteractionWeightConfig::default());
    let ctx = UserContext::load(user, &user_state, &iw_ledger, &|_| None, Timestamp::now());

    let all_items: Vec<EntityId> = (1..=20).map(EntityId::new).collect();
    let creator_lookup = |id: EntityId| -> Option<EntityId> {
        // Items 1-10 from creator 10 (followed), 11-20 from creator 20 (not followed).
        if id.as_u64() <= 10 {
            Some(EntityId::new(10))
        } else {
            Some(EntityId::new(20))
        }
    };

    let ledger = empty_test_ledger();
    let config = ExplorationConfig::default();
    let now = Timestamp::now();

    let candidates = select_exploration_candidates(
        &all_items, &ctx, &config, now, &ledger,
        &creator_lookup, &|_| Some(now.as_nanos()), 10,
    );

    // Only items from creator 20 (unfollowed) should appear.
    assert!(candidates.iter().all(|c| c.creator_id == Some(EntityId::new(20))),
        "exploration should exclude followed creators");
}

#[test]
fn select_exploration_excludes_blocked_and_hidden() {
    let user_state = UserStateIndex::new();
    let user = EntityId::new(1);
    user_state.add_block(user, EntityId::new(30));
    user_state.add_hide(user, EntityId::new(15));

    let iw_ledger = InteractionWeightLedger::new(InteractionWeightConfig::default());
    let ctx = UserContext::load(user, &user_state, &iw_ledger, &|_| None, Timestamp::now());

    let all_items: Vec<EntityId> = (1..=20).map(EntityId::new).collect();
    let creator_lookup = |id: EntityId| -> Option<EntityId> {
        if id.as_u64() <= 10 { Some(EntityId::new(20)) }
        else { Some(EntityId::new(30)) } // Items 11-20 from blocked creator
    };

    let ledger = empty_test_ledger();
    let config = ExplorationConfig::default();
    let now = Timestamp::now();

    let candidates = select_exploration_candidates(
        &all_items, &ctx, &config, now, &ledger,
        &creator_lookup, &|_| Some(now.as_nanos()), 20,
    );

    // Item 15 hidden, items 11-20 from blocked creator 30.
    assert!(candidates.iter().all(|c| c.entity_id.as_u64() != 15),
        "hidden items excluded");
    assert!(candidates.iter().all(|c| c.creator_id != Some(EntityId::new(30))),
        "blocked creator items excluded");
}

#[test]
fn is_cold_start_item_within_window() {
    let config = ExplorationConfig::default(); // 24h window
    let now = Timestamp::now();

    // Item created 1 hour ago with 0 signals: cold start.
    let one_hour_ago = now.as_nanos() - 3600 * 1_000_000_000;
    assert!(is_cold_start_item(
        EntityId::new(1), now, &config,
        &|_| Some(one_hour_ago), &|_| 0,
    ));
}

#[test]
fn is_cold_start_item_outside_window() {
    let config = ExplorationConfig::default();
    let now = Timestamp::now();

    // Item created 48 hours ago: outside 24h window.
    let forty_eight_hours_ago = now.as_nanos() - 48 * 3600 * 1_000_000_000;
    assert!(!is_cold_start_item(
        EntityId::new(1), now, &config,
        &|_| Some(forty_eight_hours_ago), &|_| 0,
    ));
}

#[test]
fn is_cold_start_item_too_many_signals() {
    let config = ExplorationConfig::default(); // max 10 signals
    let now = Timestamp::now();

    // Item created recently but has 20 signals: not cold start.
    assert!(!is_cold_start_item(
        EntityId::new(1), now, &config,
        &|_| Some(now.as_nanos()), &|_| 20,
    ));
}
```

### Property Tests

```rust
use proptest::prelude::*;

proptest! {
    #[test]
    fn inject_exploration_preserves_total_count(
        n_personalized in 0usize..100,
        n_exploration in 0usize..20,
        total_limit in 1usize..100,
        frac in 0.0f64..0.3,
    ) {
        let personalized: Vec<ScoredCandidate> = (0..n_personalized)
            .map(|i| make_candidate(i as u64 + 1, (n_personalized - i) as f64, Some(1), None))
            .collect();
        let exploration: Vec<ScoredCandidate> = (0..n_exploration)
            .map(|i| make_candidate(i as u64 + 1000, 1.0, Some(50), None))
            .collect();

        let result = inject_exploration(&personalized, &exploration, total_limit, frac);

        let expected = total_limit.min(n_personalized + n_exploration);
        prop_assert!(result.len() <= total_limit,
            "result {} > limit {}", result.len(), total_limit);
    }

    #[test]
    fn cold_start_score_always_in_unit_range(
        entity_id in 1u64..1000,
    ) {
        let ledger = empty_test_ledger();
        let now = Timestamp::now();
        let score = cold_start_score(
            EntityId::new(entity_id), now, &ledger,
            &|_| Some(now.as_nanos()),
        );
        prop_assert!(score >= 0.0 && score <= 1.0,
            "cold start score out of range: {}", score);
    }
}
```

## Acceptance Criteria

- [ ] `ExplorationConfig` with configurable item window, signal threshold, and fraction
- [ ] `cold_start_score` returns population-level score in [0, 1]
- [ ] Cold-start users get `for_you` results ranked by trending + quality + recency
- [ ] `is_cold_start_item` correctly identifies new items within window and below signal threshold
- [ ] `inject_exploration` interleaves exploration candidates into personalized results
- [ ] Exploration count = `ceil(limit * fraction)` (e.g., 5 for limit=50, fraction=0.1)
- [ ] Exploration candidates exclude: followed creators, blocked creators, hidden items
- [ ] Exploration candidates are interleaved, not clustered at the end
- [ ] Empty exploration pool gracefully falls back to personalized-only results
- [ ] `select_exploration_candidates` returns items from unfollowed creators scored by population signals
- [ ] Property test: result count <= total_limit
- [ ] Property test: cold_start_score always in [0, 1]
- [ ] `cargo clippy -- -D warnings` passes
- [ ] All tests pass

## Research References

- [VISION.md](../../../../VISION.md) -- Exploration budget, cold-start handling
- [USE_CASES.md](../../../../USE_CASES.md) -- UC-01 (For You: exploration budget prevents filter bubbles)
- [thoughts.md](../../../../thoughts.md) -- Part V.16 (cold-start user defaults to population signals)

## Implementation Notes

- The exploration budget is enforced at the RETRIEVE executor level, after personalized scoring and before the final result assembly. The executor calls `select_exploration_candidates` to get exploration items, then `inject_exploration` to merge them with personalized results.
- For cold-start users, the candidate strategy switches from ANN (no query vector available) to full corpus scan sorted by `cold_start_score`. This is correct because without a preference vector, ANN retrieval has no meaningful query.
- The interleaving strategy is simple: place one exploration candidate every `total_limit / exploration_count` positions. This distributes exploration items evenly. More sophisticated interleaving (e.g., placing them at positions where the personalized score drops) is deferred to M6.
- `select_exploration_candidates` does a linear scan of all items. At M3 scale (10K items), this is fast. At larger scale, maintaining an exploration pool bitmap would be more efficient.
- The `ExplorationConfig` is stored as a field on `TidalDb`, initialized from defaults in `TidalDbBuilder::open()`. Custom values can be set via `builder.with_exploration_config(config)`.
- Cold-start item detection (`is_cold_start_item`) is used during exploration candidate selection to prioritize genuinely new items. However, all unfollowed items are eligible for exploration, not just cold-start items. Cold-start items get a small score boost within the exploration pool.