Relevance Ranking

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

BM25 Score for document D given query Q:
 
                    n
score(D, Q) = Σ    IDF(qi) × f(qi, D) × (k1 + 1)
                   i=1     ────────────────────────────────────────
                           f(qi, D) + k1 × (1 - b + b × |D| / avgdl)
 
Where:
  qi = query term i
  f(qi, D) = frequency of qi in document D
  |D| = length of document D (in words)
  avgdl = average document length in collection
  k1 = term frequency saturation parameter (typical: 1.2)
  b = length normalization parameter (typical: 0.75)
  
IDF formula (BM25 variant):
                    N - n(qi) + 0.5
  IDF(qi) = log ( ─────────────────── )
                     n(qi) + 0.5
 
Where:
  N = total documents in collection
  n(qi) = documents containing qi

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

TF Saturation in BM25 (k1 = 1.2):
 
Raw TF | BM25 TF Component
-------|------------------
   1   |      1.29
   2   |      1.71
   3   |      1.95
   5   |      2.22
  10   |      2.55
  20   |      2.78
  50   |      2.94
 100   |      3.00  (approaching limit of k1+1 = 2.2)
 
Key insight: 
- First occurrence is most valuable
- Additional occurrences have diminishing returns
- After ~20 occurrences, score barely increases
 
Compare to raw TF where 100 occurrences = 100× weight!
BM25's saturation prevents frequency spam from dominating.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

Length Normalization Factor: (1 - b + b × |D| / avgdl)
 
Assume avgdl = 500 words, b = 0.75:
 
Doc Length | Factor | Effect
-----------|--------|-------
   100     | 0.40   | Short doc: Boost (divide by smaller number)
   250     | 0.625  | Below avg: Moderate boost
   500     | 1.0    | Average: Neutral
  1000     | 1.75   | Long: Penalty
  2500     | 4.0    | Very long: Strong penalty
 
Why? Long documents naturally contain more term occurrences
just by having more words. Length normalization levels the field.
 
With b = 0: Factor = 1 always (no normalization)
With b = 1: Factor = |D| / avgdl (full normalization)
With b = 0.75: Balanced middle ground

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

Typical Document Fields by Importance:
 
Field               | Suggested Weight | Rationale
--------------------|------------------|----------------------------------
Title               | 4.0 - 10.0       | Highest: Directly describes content
H1/H2 Headings      | 2.0 - 4.0        | Structure indicates importance
Abstract/Summary    | 2.0 - 3.0        | Concentrated information
Keywords/Tags       | 2.0 - 4.0        | Explicit topic markers
First paragraph     | 1.5 - 2.0        | Often contains key points
Body text           | 1.0 (baseline)   | Default weight
Footnotes/Citations | 0.5 - 0.8        | Less relevant to main topic
Comments/Metadata   | 0.3 - 0.5        | User-generated, less reliable
 
Example scoring:
Query: "machine learning"
 
Doc A: "machine learning" in title → score × 5.0
Doc B: "machine learning" in body (5 times) → score × 1.0
 
Doc A likely ranks higher despite fewer occurrences!

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31

// Multi-field query with boosting
 
GET /articles/_search
{
  "query": {
    "multi_match": {
      "query": "database optimization",
      "fields": [
        "title^5",          // Title matches worth 5x
        "headings^3",       // Heading matches worth 3x
        "abstract^2",       // Abstract matches worth 2x
        "body",             // Body is baseline (1x)
        "footnotes^0.5"     // Footnotes discounted
      ],
      "type": "best_fields", // Score = max of field scores
      "tie_breaker": 0.3     // Plus 0.3 × other field scores
    }
  }
}
 
// Alternative: Combine with bool query
{
  "query": {
    "bool": {
      "should": [
        { "match": { "title": { "query": "database", "boost": 5 }}},
        { "match": { "body": { "query": "database", "boost": 1 }}}
      ]
    }
  }
}

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

-- PostgreSQL uses weight labels A, B, C, D
-- By default: A=1.0, B=0.4, C=0.2, D=0.1
 
-- Create weighted tsvector
SELECT 
    setweight(to_tsvector('english', COALESCE(title, '')), 'A') ||
    setweight(to_tsvector('english', COALESCE(abstract, '')), 'B') ||
    setweight(to_tsvector('english', COALESCE(body, '')), 'C') ||
    setweight(to_tsvector('english', COALESCE(footnotes, '')), 'D')
    AS document_vector
FROM articles;
 
-- Custom weights in ts_rank
SELECT 
    title,
    ts_rank(
        document_vector,
        to_tsquery('database & optimization'),
        ARRAY[0.1, 0.2, 0.4, 1.0]  -- Custom weights for D,C,B,A
    ) AS relevance
FROM articles
WHERE document_vector @@ to_tsquery('database & optimization')
ORDER BY relevance DESC;

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

Query: "database optimization"
 
Document A: "This section covers database optimization techniques."
  Span: 2 words (adjacent)
  Score boost: Maximum (exact phrase)
 
Document B: "Database systems require optimization for performance."  
  Span: 4 words (database ... optimization)
  Score boost: High (very close)
 
Document C: "We will optimize the database connection pool."
  Span: 3 words (optimize ... database)
  Order: Reversed
  Score boost: Moderate (close but reordered)
 
Document D: "The database stores millions of records. Performance 
            optimization is critical for the application."
  Span: ~12 words across sentences
  Score boost: Low (same paragraph but far)
 
Document E: "Chapter 1: Database Design... Chapter 5: Optimization."
  Span: Hundreds of words
  Score boost: Minimal (different sections entirely)
 
Ranking: A > B > C > D > E

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48

def compute_minimum_span_score(query_terms, term_positions):
    """
    Compute proximity score based on minimum span containing all query terms.
    
    term_positions: dict mapping term -> list of positions in document
    """
    if not all(term in term_positions for term in query_terms):
        return 0  # Not all terms present
    
    # Create merged list of (position, term) sorted by position
    all_positions = []
    for term in query_terms:
        for pos in term_positions[term]:
            all_positions.append((pos, term))
    all_positions.sort()
    
    min_span = float('inf')
    terms_needed = set(query_terms)
    
    # Sliding window to find minimum span
    left = 0
    terms_in_window = {}
    
    for right, (pos, term) in enumerate(all_positions):
        terms_in_window[term] = terms_in_window.get(term, 0) + 1
        
        # Shrink window from left while all terms covered
        while len(terms_in_window) == len(terms_needed):
            left_pos, left_term = all_positions[left]
            span = pos - left_pos + 1
            min_span = min(min_span, span)
            
            terms_in_window[left_term] -= 1
            if terms_in_window[left_term] == 0:
                del terms_in_window[left_term]
            left += 1
    
    # Convert span to score (smaller span = higher score)
    if min_span == float('inf'):
        return 0
    
    # Score formula: Inverse of span, with minimum cap
    return len(query_terms) / max(min_span, len(query_terms))
 
# Example: query = ["database", "optimization"]
# positions = {"database": [5, 42, 100], "optimization": [7, 105]}
# Best span: positions 5-7 → span = 3
# Score: 2 / 3 = 0.67

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44

def compute_final_score(doc, query, user_context):
    """
    Combine multiple signals into final ranking score.
    Production systems may have 100+ features!
    """
    
    # Text relevance (BM25)
    text_score = compute_bm25(doc, query)
    
    # Field-weighted matching
    field_score = (
        5.0 * match_in_title(doc, query) +
        2.0 * match_in_headings(doc, query) +
        1.0 * match_in_body(doc, query)
    )
    
    # Proximity bonus
    proximity_score = compute_proximity(doc, query)
    
    # Document quality signals
    quality_score = (
        0.3 * doc.pagerank +
        0.2 * doc.avg_rating +
        0.1 * log(doc.view_count + 1)
    )
    
    # Freshness (decay older content)
    days_old = (today - doc.published_date).days
    freshness_score = 1.0 / (1 + 0.01 * days_old)  # Exponential decay
    
    # Personalization
    personal_score = user_affinity(user_context, doc.topics)
    
    # Combine with learned weights
    final_score = (
        0.40 * text_score +
        0.20 * field_score +
        0.15 * proximity_score +
        0.10 * quality_score +
        0.10 * freshness_score +
        0.05 * personal_score
    )
    
    return final_score

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34

Learning to Rank Feature Vector:
 
For each (query, document) pair, extract features:
 
Feature Category      | Example Features
----------------------|------------------------------------------
Text Matching         | BM25 score, TF-IDF score, term overlap
Field Matching        | Title match, heading match, URL match
Proximity             | Min span, phrase match count, same sentence
Query Features        | Query length, term count, rare term count
Document Features     | PageRank, length, freshness, ratings
Match Statistics      | Terms matched, % coverage, IDF sum
 
Example feature vector:
query: "machine learning tutorial"
document: "Introduction to Machine Learning - Stanford"
 
features = [
    4.23,   # BM25 score
    0.85,   # Title match (85% terms match)
    0.67,   # Heading match score  
    2,      # Query length (words)
    3,      # Document matches 3 of 3 terms
    0.92,   # Phrase proximity score
    7.2,    # PageRank score
    30,     # Days since publication
    4.5,    # Average user rating
    12500,  # View count
    ...
]
 
Label: 4 (highly relevant) on 0-4 scale
 
Training data: Thousands of such labeled examples

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

LambdaMART Algorithm Overview:
 
1. Initialize: Simple ranking (e.g., BM25 only)
 
2. For each boosting iteration:
   a. Compute gradients based on ranking quality
      - Consider all pairs of documents
      - Gradient focuses on pairs where swap improves ranking
      - Weight by how much nDCG would improve from swap
      
   b. Train regression tree to predict gradient
   
   c. Update model: Add tree with learning rate
 
3. Final model: Ensemble of trees
 
Why it works:
- Trees handle non-linear feature interactions
- Lambda gradients directly optimize ranking metrics (nDCG)
- Focuses training on important ranking decisions
 
Typical model:
- 500-2000 trees
- Max depth 6-10
- 50-200 features

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

Precision and Recall (Binary Relevance):
 
Consider a query with results and relevance judgments:
R = Set of truly relevant documents (from human annotation)
A = Set of returned documents (search results)
 
                 |R ∩ A|          Relevant docs returned
Precision = ──────────── = ─────────────────────────────────
                  |A|            Total docs returned
 
                 |R ∩ A|          Relevant docs returned
Recall    = ──────────── = ─────────────────────────────────
                  |R|            Total relevant docs
 
Example:
True relevant: {doc1, doc2, doc3, doc5, doc8}  (5 relevant)
Returned:      {doc1, doc2, doc4, doc6, doc3}  (5 returned)
Intersection:  {doc1, doc2, doc3}              (3 correct)
 
Precision = 3/5 = 0.6  (60% of results are relevant)
Recall    = 3/5 = 0.6  (60% of relevant docs found)
 
Tradeoff: Returning more increases recall but may decrease precision.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

Precision@K: Consider only top K results
 
Results (ordered): [doc1✓, doc4✗, doc2✓, doc6✗, doc3✓, doc5✓, ...]
 
P@1 = 1/1 = 1.0     (top 1 result is relevant)
P@3 = 2/3 = 0.67    (2 of top 3 are relevant)
P@5 = 3/5 = 0.60    (3 of top 5 are relevant)
P@10 = 4/10 = 0.40  (4 of top 10 are relevant)
 
Why P@K matters:
- Users rarely look past first page (top 10)
- P@10 is often the most critical metric
- Different K values for different use cases:
  - Web search: P@10 (first page)
  - Mobile search: P@3 (small screen)
  - Voice search: P@1 (single answer)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32

Normalized Discounted Cumulative Gain (nDCG):
 
Graded relevance scale: 0 (irrelevant) to 4 (perfect)
 
Step 1: Compute Discounted Cumulative Gain (DCG)
                n
  DCG@n = Σ    relevance(i) / log₂(i + 1)
               i=1
               
Position discount: log₂(2)=1, log₂(3)=1.58, log₂(4)=2, log₂(5)=2.32
 
Example ranking with relevance scores:
Position:  1    2    3    4    5
Relevance: 4    2    0    3    1
 
DCG@5 = 4/log₂(2) + 2/log₂(3) + 0/log₂(4) + 3/log₂(5) + 1/log₂(6)
      = 4/1 + 2/1.58 + 0/2 + 3/2.32 + 1/2.58
      = 4.0 + 1.26 + 0 + 1.29 + 0.39
      = 6.94
 
Step 2: Compute Ideal DCG (IDCG) - best possible order
Ideal order: 4, 3, 2, 1, 0  (descending by relevance)
 
IDCG@5 = 4/1 + 3/1.58 + 2/2 + 1/2.32 + 0/2.58
       = 4.0 + 1.90 + 1.0 + 0.43 + 0
       = 7.33
 
Step 3: Normalize
nDCG@5 = DCG@5 / IDCG@5 = 6.94 / 7.33 = 0.947
 
Interpretation: 94.7% of ideal ranking quality achieved.
Perfect ranking: nDCG = 1.0

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

Two-Stage Ranking Architecture:
 
Stage 1: Candidate Generation (Fast)
-----------------------------------------
- Use inverted index for recall
- Apply simple scoring (BM25)
- Retrieve top 1000 candidates
- Speed: ~10ms for millions of docs
 
Stage 2: Re-Ranking (Sophisticated)  
-----------------------------------------
- Extract rich features for 1000 candidates
- Apply ML model (LambdaMART, Neural)
- Consider expensive signals (proximity, ML)
- Return top 10-20 results
- Speed: ~50-100ms for 1000 candidates
 
Why two stages?
- Can't run expensive ML model on millions of docs
- Inverted index efficiently filters to candidates
- Re-ranker provides quality polish on small set
 
Total latency: ~60-110ms for world-class results

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34

Result Diversity:
 
Problem: Top 10 results all from same website = poor experience
 
Solution: Maximal Marginal Relevance (MMR)
 
MMR(d) = λ × Relevance(d) - (1-λ) × max Similarity(d, selected)
 
At each position:
1. Pick document with highest MMR
2. Add to selected set
3. Repeat
 
Effect: Balance relevance against redundancy
 
Example:
Position 1: Pick most relevant doc
Position 2: Pick relevant doc that's different from #1
Position 3: Pick relevant doc different from #1 and #2
...
 
Result: Diverse top 10 even if similar docs score high
 
Freshness Boost:
 
For time-sensitive queries (news, events):
freshness_boost = 1.0 / (1 + decay_rate × days_old)
 
Score = relevance_score × freshness_boost
 
Setting decay_rate by query type:
- Breaking news: High decay (hours)
- General news: Moderate decay (days)
- Evergreen content: Low decay (months)