如何优化Elastic Search职业标题好友推荐的搜索相关性

1. Is your assumption correct? #

Not entirely—let’s clarify how minimum_should_match works here. This parameter controls the proportion of query terms (after tokenization) that need to match, not the percentage of overlapping characters between job titles.

Here’s what actually happens in your setup:

• For single-word titles like barber: Your search analyzer splits the query into a single term (barber). When you set minimum_should_match: 50%, Elasticsearch rounds this up to 100% for a single term—meaning it still requires an exact match to the indexed ngrams. So bartender won’t be recommended just because of shared characters, since its indexed ngrams don’t include the full term barber.
• For multi-word titles like Senior Software Developer: The query splits into 3 terms. A 90% threshold requires matching all 3 terms (since 3*0.9 = 2.7, rounded up to 3), which is why you’re getting too few results. Dropping to 50% would allow matching 2 out of 3 terms (e.g., Junior Software Developer or Senior Frontend Developer), which are still relevant to the original title—this isn’t about character similarity, it’s about partial term matching.

Your concern about character-based mismatches is valid in other scenarios, but it’s not the issue you’d face with lowering minimum_should_match here.

2. How to improve title search relevance while increasing match count? #

Here are practical, targeted solutions tailored to your use case:

a. Use dynamic minimum_should_match rules #

Instead of a fixed percentage, adjust the threshold based on how many terms are in the query. This balances strictness for long titles and flexibility for short ones. For example:

should: [
  { 
    match: { 
      title: { 
        query: user.title, 
        minimum_should_match: "3<90% 2<70% 1<100%",
        boost: 2 
      } 
    } 
  }
]

This rule means:

• If the query has 3+ terms, require 90% match (strict for multi-word titles)
• If 2 terms, require 70% match
• If 1 term, require 100% match (avoids accidental character-based matches)

b. Add synonym filters for related job titles #

Create a synonym filter to map similar job roles together, so Elasticsearch recognizes them as equivalent. For example:

filter: {
  // ... your existing filters
  job_synonyms: {
    type: "synonym",
    synonyms: [
      "barber, hairdresser",
      "bartender, mixologist",
      "developer, software engineer, coder",
      "blogger, content creator, writer"
    ]
  }
}

Update your analyzers to include this filter (both index and search analyzers). This way, searching for barber will also return users with hairdresser—boosting match count while keeping relevance high.

c. Combine multiple query types with weighted boosts #

Use a mix of exact matches, fuzzy matches, and prefix matches, each with different boosts to prioritize the most relevant results first. Example:

should: [
  // Exact term match (highest priority)
  { match: { title: { query: user.title, minimum_should_match: "100%", boost: 3 } } },
  // Fuzzy match for minor typos/variations
  { match: { title: { query: user.title, fuzziness: "AUTO", boost: 1.5 } } },
  // Prefix match for partial job titles (e.g., "dev" matching "developer")
  { match_phrase_prefix: { title: { query: user.title, boost: 1 } } }
]

d. Standardize job titles with a taxonomy #

For long-term improvement, map raw job titles to a standardized taxonomy (like O*NET or a custom internal list). Add a standardized_title field to your documents, then run recommendations against this field instead of raw titles. This eliminates inconsistencies (e.g., "Sr Dev" vs "Senior Software Developer") and ensures recommendations are based on actual job roles, not string matching.

内容的提问来源于stack exchange，提问作者Vivien Adnot