@recoengine/features

Reusable, domain-neutral feature extractors and transforms for @recoengine/core.

npm i @recoengine/features @recoengine/core

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

What it is

The standard strategies read features with fixed names — interaction_count, interaction_recency, and so on. Someone has to produce those. For the features that come from the interaction history rather than from your item payload, that producer is domain-neutral: counting events keyed by item id, and decaying their recency, is the same arithmetic whether the items are tracks or products. This package ships those producers, so you get history-based scoring without writing any domain code.

| Factory | Kind | Produces | | --- | --- | --- | | interactionCountExtractor | extractor | interaction_count — how often the user interacted with each item | | interactionRecencyExtractor | extractor | interaction_recency — decayed recency of those interactions | | logTransform | transform | log1p compression of a column (tame heavy tails) | | decayTransform | transform | a decay curve (exponential / linear) over a column |

• Extractors read ctx.history and ctx.now — never the payload — which is what keeps them reusable across domains.
• Transforms are pure maths over columns, domain-neutral by construction.

Features that genuinely need the payload — an item's own age, its category, a precomputed co-occurrence score — stay in your own domain extractors, because only you know where in the payload they live.

Usage

import { createEngine } from '@recoengine/core'
import { interactionCountExtractor, interactionRecencyExtractor, logTransform } from '@recoengine/features'
import { historyStrategy } from '@recoengine/strategies'

const engine = createEngine<Track>()
  .use(myProvider)
  .use(interactionCountExtractor())
  .use(interactionRecencyExtractor({ halfLife: 30 }))     // days
  .use(logTransform({ source: 'plays', target: 'plays_log' }))
  .use(historyStrategy())                                 // reads the features above
  .configure({ limits: { maxCandidates: 5_000, maxLimit: 100, timeoutMs: 200 } })
  .build()

Options: InteractionCountOptions, InteractionRecencyOptions, LogTransformOptions, DecayTransformOptions.

Links

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

MIT