Content Based Recommender TS

This is a fork from the content-based-recommender package with support for TS.

This is a simple content-based recommender implemented in javascript to illustrate the concept of content-based recommendation. Content-based recommender is a popular recommendation technique to show similar items to users, especially useful to websites for e-commerce, news content, etc.

After the recommender is trained by an array of documents, it can tell the list of documents which are more similar to the input document.

The training process involves 3 main steps:

• content pre-processing, such as html tag stripping, stopwords removal and stemming
• document vectors formation using tf-idf
• find the cosine similarity scores between all document vectors

Special thanks to the library natural helps a lot by providing a lot of NLP functionalities, such as tf-idf and word stemming.

⚠️ Note:

I haven't tested how this recommender is performing with a large dataset.

Installation

npm install content-based-recommender

or

yarn add content-based-recommender

Then import the ContentBasedRecommender class

1const ContentBasedRecommender = require('content-based-recommender')

What's New

Check the CHANGELOG.

Usage

Single collection

1const ContentBasedRecommender = require('content-based-recommender')
2const recommender = new ContentBasedRecommender({
3  minScore: 0.1,
4  maxSimilarDocs: 100
5})
6
7// prepare documents data
8const documents = [
9  { id: '1000001', content: 'Why studying javascript is fun?' },
10  { id: '1000002', content: 'The trend for javascript in machine learning' },
11  { id: '1000003', content: 'The most insightful stories about JavaScript' },
12  { id: '1000004', content: 'Introduction to Machine Learning' },
13  { id: '1000005', content: 'Machine learning and its application' },
14  { id: '1000006', content: 'Python vs Javascript, which is better?' },
15  { id: '1000007', content: 'How Python saved my life?' },
16  { id: '1000008', content: 'The future of Bitcoin technology' },
17  { id: '1000009', content: 'Is it possible to use javascript for machine learning?' }
18]
19
20// start training
21recommender.train(documents)
22
23//get top 10 similar items to document 1000002
24const similarDocuments = recommender.getSimilarDocuments('1000002', 0, 10)
25
26console.log(similarDocuments)
27/*
28  the higher the score, the more similar the item is
29  documents with score < 0.1 are filtered because options minScore is set to 0.1
30  [
31    { id: '1000004', score: 0.5114304586412038 },
32    { id: '1000009', score: 0.45056313558918837 },
33    { id: '1000005', score: 0.37039308109283564 },
34    { id: '1000003', score: 0.10896767690747626 }
35  ]
36*/

Multi collection

This example shows how to automatically match posts with related tags

1const ContentBasedRecommender =  require('content-based-recommender')
2
3const posts = [
4  {
5    id: '1000001',
6    content: 'Why studying javascript is fun?'
7  },
8  {
9    id: '1000002',
10    content: 'The trend for javascript in machine learning'
11  },
12  {
13    id: '1000003',
14    content: 'The most insightful stories about JavaScript'
15  },
16  {
17    id: '1000004',
18    content: 'Introduction to Machine Learning'
19  },
20  {
21    id: '1000005',
22    content: 'Machine learning and its application'
23  },
24  {
25    id: '1000006',
26    content: 'Python vs Javascript, which is better?'
27  },
28  {
29    id: '1000007',
30    content: 'How Python saved my life?'
31  },
32  {
33    id: '1000008',
34    content: 'The future of Bitcoin technology'
35  },
36  {
37    id: '1000009',
38    content: 'Is it possible to use javascript for machine learning?'
39  }
40]
41
42const tags = [
43  {
44    id: '1',
45    content: 'Javascript'
46  },
47  {
48    id: '2',
49    content: 'machine learning'
50  },
51  {
52    id: '3',
53    content: 'application'
54  },
55  {
56    id: '4',
57    content: 'introduction'
58  },
59  {
60    id: '5',
61    content: 'future'
62  },
63  {
64    id: '6',
65    content: 'Python'
66  },
67  {
68    id: '7',
69    content: 'Bitcoin'
70  }
71]
72
73const tagMap = tags.reduce((acc, tag) => {
74  acc[tag.id] = tag
75
76  return acc
77}, {})
78
79const recommender = new ContentBasedRecommender()
80
81recommender.trainBidirectional(posts, tags)
82
83for (const post of posts) {
84  const relatedTags = recommender.getSimilarDocuments(post.id)
85  const tags = relatedTags.map(t => tagMap[t.id].content)
86  console.log(post.content, 'related tags:', tags)
87}
88
89
90/*
91Why studying javascript is fun? related tags: [ 'Javascript' ]
92The trend for javascript in machine learning related tags: [ 'machine learning', 'Javascript' ]
93The most insightful stories about JavaScript related tags: [ 'Javascript' ]
94Introduction to Machine Learning related tags: [ 'machine learning', 'introduction' ]
95Machine learning and its application related tags: [ 'machine learning', 'application' ]
96Python vs Javascript, which is better? related tags: [ 'Python', 'Javascript' ]
97How Python saved my life? related tags: [ 'Python' ]
98The future of Bitcoin technology related tags: [ 'future', 'Bitcoin' ]
99Is it possible to use javascript for machine learning? related tags: [ 'machine learning', 'Javascript' ]
100*/
101

API

constructor([options])

To create the recommender instance

• options (optional): an object to configure the recommender

Supported options:

• maxVectorSize - to control the max size of word vector after tf-idf processing. A smaller vector size will help training performance while not affecting recommendation quality. Defaults to be 100.
• minScore - the minimum score required to meet to consider it is a similar document. It will save more memory by filtering out documents having low scores. Allowed values range from 0 to 1. Default is 0.
• maxSimilarDocs - the maximum number of similar documents to keep for each document. Default is the max safe integer in javascript.
• debug - show progress messages so can monitor the training progress

train(documents)

To tell the recommender about your documents and then it will start training itself.

• documents - an array of object, with fields id and content

trainBidirectional(collectionA, collectionB)

Works like the normal train function, but it creates recommendations between two different collections instead of within one collection.

getSimilarDocuments(id, [start], [size])

To get an array of similar items with document id

• id - the id of the document
• start - the start index, inclusive. Default to be 0
• size - the max number of similar documents to obtain. If it is omitted, the whole list after start index will be returned

It returns an array of objects, with fields id and score (ranging from 0 to 1)

export

To export the recommender as json object.

1const recommender = new ContentBasedRecommender()
2recommender.train(documents)
3
4const object = recommender.export()
5//can save the object to disk, database or otherwise

import(object)

To update the recommender by importing from a json object, exported by the export() method

1const recommender = new ContentBasedRecommender()
2recommender.import(object) // object can be loaded from disk, database or otherwise

Test

1npm i
2npm run test

Authors

• Stanley Fok
• AnthonyLzq

License

MIT