@recoengine/strategies

The nine standard, domain-neutral scoring strategies for @recoengine/core.

npm i @recoengine/strategies @recoengine/core

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

What it is

A scoring strategy answers "how well does this candidate fit this user" by reading one or more feature columns and folding them into a single score column plus its reasons. Each strategy here is a factory function returning a ScoringStrategy you register with .use(...). None of them looks inside your item payload — they read named feature columns a domain extractor produced — so the same strategy serves music, commerce, or news unchanged.

The strategies

| Factory | Scores by | | --- | --- | | historyStrategy | repeat interaction — count × recency of past events | | affinityStrategy | match to a user-affinity feature (artist, brand, genre…) | | popularityStrategy | blended global/cohort popularity percentiles; a cold-start fallback | | recencyStrategy | how new the item is (exponential decay of age) | | similarityStrategy | similarity to recent items and to the user profile | | coOccurrenceStrategy | "people who interacted with X also…" co-occurrence | | noveltyStrategy | unfamiliar items, scaled by how saturated the profile is | | discoveryStrategy | rewards distance from the usual — exploration | | contextStrategy | match to request-time context signals |

Usage

import { createEngine } from '@recoengine/core'
import { popularityStrategy, historyStrategy, affinityStrategy } from '@recoengine/strategies'

const engine = createEngine<Track>()
  .use(myProvider)
  .use(myFeatureExtractors)                              // produce the features these read
  .use(historyStrategy())
  .use(popularityStrategy({ cohortFeature: null }))
  .use(affinityStrategy({ id: 'artist', feature: 'affinity_artist' }))
  .use(affinityStrategy({ id: 'genre', feature: 'affinity_genre' }))
  .configure({ limits: { maxCandidates: 5_000, maxLimit: 100, timeoutMs: 200 } })
  .build()

Each factory takes an options object (e.g. HistoryStrategyOptions, PopularityStrategyOptions) to tune thresholds, feature names, and weights. Register two instances of the same strategy under different ids (as with the two affinityStrategy calls above) so they don't compete for the same weight.

Every strategy declares the feature keys it requires; get one wrong and build() throws before the first request. The history-based features these expect can be produced for free by @recoengine/features.

Links

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

MIT