Automated Topic Detection and Social Recommendation using LDA

Vinay Mandal

2015-09-01

Motivation

  1. Most recommendation don't use review texts
  2. Real World data
  3. Graphs

The dataset

Number of Businesses: 61184
Number of Users: 366715
Number of Reviews: 1569264

Business Categories

height

Ratings by day of week

height

Average rating text length

height

Method Overview

  1. Uncover hidden dimensions of ratings and review text combined
  2. Use the hidden dimensions from review texts to find user preference
  3. Build a graph database of user - prefers - latent topic
  4. Use the graph to recommend new friends

Training

Tokenize Reviews → Remove stop words → POS Tagging → Lemmatize → Train LDA → Label latent topics

Tokenization and removal of stop words

  1. Process of breaking stream of text into words, symbols, phrases or other meaningful elements called tokens.
  2. Certain characters like punctuations might be excluded
    An example: "Few things are harder to put up with than a good example."
    can be tokenized into:
    ["few", "things", "are", "harder", "to", "put", "up", "with", "than", "a", "good", "example"]
  3. Stop words like are, to, up and a can be ignored as they are not important

POS Tagging

Process of classifying words into their parts of speech based on its definition as well as its context.

[('Few', 'JJ'), ('things', 'NNS'), ('harder', 'JJR'), ('put', 'VB'), ('up', 'RP'), ('than', 'IN'), ('good', 'JJ'), ('example', 'NN')]

Lemmatization

"Process of grouping diffrent inflected forms of word so that they can be analyzed as single word"

wordnet_lemmatizer.lemmatize(‘aardwolves’) # u'aardwolf'
wordnet_lemmatizer.lemmatize(‘is’, pos=’v’) # u'be'
wordnet_lemmatizer.lemmatize(‘are’, pos=’v’) # u'be

Latent Dirichlet Allocation

"...is a generative model that allows sets of observations to be explained by unobserved groups that explain why some parts of the data are similar."

  • In LDA, a document is a mixture of a small number of topics and that each word's creation is attributable to one of the document's topics
  • LDA is similar to pLSA except in LDA topic distribution is assumed to have Dirichlet prior.
  • Using python package "gensim"

Model prediction example

"I love this place and eat here at least once a week. Their pad si yew, pad Thai, dumplings, Siam rolls, and their curries are their best dishes. Especially the curry. You can't go wrong at Thai moon, so I suggest you go and enjoy the food!"

Model prediction:
['scores': u'0.060*thai + 0.036*food + 0.032*curry + 0.028*lunch + 0.027*spicy + 0.023*place + 0.021*tea + 0.019*dish + 0.017*chicken + 0.015*time']

User → Likes → Topic

Why graph db?

  1. Relationships are first-class citizens of graph database
  2. High performance online query
  3. Good for building recommendation engine

User → Likes → Topic

Neo4j with "py2neo". Example:

from py2neo import Node, Relationship, Graph
graph = Graph()
userNode = Node('User', name='name', id=jsonData'user_id')
topicNode = Node('topics', 'key', topicKey)
likes = Relationship(userNode, 'topiclike', topicNode)
graph.create(likes)

User → Likes → Topic

height

Recommendation Engine

query = ('match (u1 {id:"UZyNf0h26jNVvpHT-GF2oA"})'
         ' - [rel:topiclike]-> r <- [re:topiclike] - (u2) '
         'with count(r) as topiccounts, u2 as user2 '
         'where topiccounts > 2 return user2, topiccounts'
         ' order by topiccounts desc limit 10')
result = graph.cypher.execute(query)

Result

userid topiccounts
1 kGgAARL2UmvCcTRfiscjug 20
2 4ozupHULqGyO42s3zNUzOQ 19
3 DrWLhrK8WMZf7Jb-Oqc7ww 16
4 90a6z--_CUrl84aCzZyPsg 15
5 4G68oLRY3aHE5XUt_MUUcA 12
6 iwUN95LIaEr75TZE_JC6bg 11
7 pEVf8GRshP9HUkSpizc9LA 10
8 Ovpa3S8xD96dLE5eDxcxJg 10
9 WmAyExqSWoiYZ5XEqpk_Uw 9
10 lmiDCrmas8TxRsbIGZX9Pg 9

-

Future Work

  1. Find trendsetters
  2. Location aware recommendation
  3. Predict usefulness of review texts