@recoengine/diversity

Diversification and blending for @recoengine/core: MMR, attribute quotas, exploration blending, and similarity providers.

npm i @recoengine/diversity @recoengine/core

Already using recoengine? These are re-exported from it — no separate install needed.

What it is

Ranking alone gives a correct but samey feed: twenty tracks by one artist, formally perfect and practically useless. These reorder the ranked list toward variety, and mix in exploration.

| Factory | What it does | | --- | --- | | mmrDiversifier | Maximal Marginal Relevance — trades relevance against similarity to already-picked items | | attributeQuotaDiversifier | caps how many items per group (artist, brand, category…) can appear | | bucketBlender | mixes buckets by slot quota, e.g. reserve N slots for exploration | | cosineSimilarity / jaccardSimilarity | similarity providers MMR uses to judge "how alike are these two" |

Similarity is a separate port on purpose: how alike are two items is domain-shaped, while MMR is not — keeping them apart is what lets one MMR serve music and e-commerce.

Usage

import { createEngine } from '@recoengine/core'
import { mmrDiversifier, attributeQuotaDiversifier, cosineSimilarity } from '@recoengine/diversity'

const engine = createEngine<Track>()
  .use(myProvider)
  .use(myStrategies)
  .use(cosineSimilarity())                                       // how MMR measures closeness
  .use(mmrDiversifier({ lambda: 0.7 }))                          // 0.7 relevance / 0.3 variety
  // or cap repeats by a group feature instead of / in addition to MMR:
  .use(attributeQuotaDiversifier({ feature: 'artist_group', max: 2 }))
  .configure({ limits: { maxCandidates: 5_000, maxLimit: 100, timeoutMs: 200 } })
  .build()

Options: MmrOptions, AttributeQuotaOptions, BucketBlenderOptions, CosineSimilarityOptions, WeightedJaccardSimilarityOptions.

Links

• Repository & full docs: https://github.com/waleron07/recommendationEngine

MIT