Easy, fast, greedy recommender

"Will it scale?" is a less important question than "will it ever matter?" (David Kadavy)

NB: We have re-written good part of the recommender.

The APIs have changed, and the webapp is now a separate package, called cold-start-recommender-webapp, which can be installed via pip. You can still access the old version with:

pip install cold-start-recommender==0.3.15

or from the source folder (same folder of the setup.py file):

pip install .

To Uninstall the package:

pip uninstall csrec

Any comment sent to info@elegans.io will be appreciated.

We developed Cold Start Recommender because we needed a recommender with the following characteristics:

• Greedy. Useful in situations where no previous data on Items or Users are available, therefore any information must be used --not just which Item a User likes, but also --in the case of a book-- the corresponding category, author etc.

• Fast. Any information on Users and Item should be stored and used immediately. A rating by any User should improve recommendations for this User, but also for other Users. This means in-memory database and no batch computations.

• Ready to use. Take a look at cold-start-recommender-webapp to start a webapp that POSTs information and GETs recommendations.

CSRec should not (yet) be used for production systems, but only for pilots, where statistics are so low that filters (e.g. loglikelihood filter on the co-occurence matrix) are premature. It aims to gather data in order to immediately personalise the user experience.

CSRec is written in Python, and under the hood it uses the Pandas_ library.

Dependencies

The following python packages are needed in order to run the recommender:

• pickle
• pandas
• numpy

Since version 4, the web service has been taken out of the package. You need to install elegans.io's package csrec-webapp

Features

The Cold Start Problem

The Cold Start Problem originates from the fact that collaborative filtering recommenders need data to build recommendations. Typically, if Users who liked item 'A' also liked item 'B', the recommender would recommend 'B' to a user who just liked 'A'. But if you have no previous rating by any User, you cannot make any recommendation.

CSRec tackles the issue in various ways.

Selective profiling

CSRec allows profiling with well-known Items without biasing the results.

For instance, if a call to insert_rating is done in this way:

engine.db.insert_item_action(user_id='user1', item_id='item1', code=4, item_meaningful_info=['author', 'tags'], only_info=True)

CSRec will only register that user1 likes a certain author, certain tags, but not that s/he might like item1. This is of fundamental importance when profiling users through a "profiling page" on your website. If you ask users whether they prefer "Harry Potter" or "The Better Angels of Our Nature", and most of them choose Harry Potter, you would not want to make the Item "Harry Potter" even more popular. You might just want to record that those users like children's books marketed as adult literature.

CSRec does that because, unless you are Amazon or a similar brand, the co-occurence matrix is often too sparse to compute decent recommendations. In this way you start building multiple, denser, co-occurence matrices and use them from the very beginning.

Store any possible information

Any information is used. You decide which information you should record about a User rating an Item. This is similar to the previous point, but you also register the item_id.

Use everything you can, now

Any information is used immediately. The co-occurence matrix is updated as soon as a rating is inserted.

Efficient users' tracking

It tracks anonymous users and merges their preferences into profiles. E.g. an anonymous visitors of a website likes a few items before the sign in/ sign up process. After sign up/ sign in the information can be reconciled --information relative to the session ID is moved into the correspondent user ID entry.

Mix recommended items and popular items

What about users who would only receive a couple of recommendations? No problem! CSRec will fill the list with the most popular items (nor rated by such users).

Algorithms

At the moment CSRec only provides purely item-based recommendations (co-occurence matrix dot the User's ratings array). In this way we can provide recommendations in less than 200msec for a matrix of about 10,000 items.

A simple script

from csrec import Recommender
engine = Recommender()

# Insert items with their properties (e.g. author, tags...)
# NB lists can be passed as json-parseable strings or strings
engine.db.insert_item(item_id='item1', attributes={'author': 'Author A', 'tags': '["nice", "good", "new"]'})

# The author field is a list, even if it was passed as a simple string:
assert engine.db.items_tbl['item1']['author'] ==  ['Author A']

engine.db.insert_item(item_id='item2', attributes={'author': '["Author B", "Author Z"]', 'tags': '["nice", "fair"]'})
engine.db.insert_item(item_id='item3', attributes={'author': 'Author B', 'tags': '["nice", "good"]'})
engine.db.insert_item(item_id='item4', attributes={'author': 'Author C', 'tags': '["new", "fashion"]'})

# The following lines tell the recommender that user1 likes items 1 and 2 but also "Author A", "B", "Z"
# and tags "nice", "good" and "fair"

engine.db.insert_item_action(user_id='user1', item_id='item1', code=4, item_meaningful_info=['author', 'tags'])
engine.db.insert_item_action(user_id='user1', item_id='item2', code=5, item_meaningful_info=['author', 'tags'])

# user1 has given a total of 4 points to Author A, 5 to Author B and Z, 4 to tag good, 5 to fair, and 9 to nice:
assert engine.db.tot_categories_user_ratings == {'author': {'user1': {'Author A': 4, 'Author B': 5, 'Author Z': 5}},
'tags': {'user1': {'fair': 5, 'good': 4, 'new': 4, 'nice': 9}}}

# ...and user2 likes item3, "Author B", "nice" and "good" items:
engine.db.insert_item_action(user_id='user2', item_id='item3', code=5, item_meaningful_info=['author', 'tags'])

# ...and user3 likes item4, "Author C", but we give no information about the tag!
engine.db.insert_item_action(user_id='user3', item_id='item4', code=5, item_meaningful_info=['author'])

# ...and user4 only goes through the profiling page, and say she likes books tagged as 'new' and 'fashion'
engine.db.insert_item_action(user_id='user4', item_id='item4', code=5, item_meaningful_info=['tags'], only_info=True)

# We should recommend to user1 items 3 and then 4, etc etc
assert engine.get_recommendations('user1') == ['item3', 'item4']

# 'user2' signs in and we discover that it's 'user1' who was browsing anonymously
engine.db.reconcile_user('user2', 'user1')

# now we know user1 liked item1, 2, 3
assert engine.db.users_ratings_tbl['user1'] == {'item1': 4, 'item2': 5, 'item3': 5}

# so we can only recommend item4
assert engine.get_recommendations('user1') == ['item4']

Remember that the cold start recommender is now only in memory, which means that you must implement a periodic saving of the data:

# Save the data from the engine from above
engine.db.serialize('pippo.db')

# create a new engine with the same data:
new_engine = Recommender()
new_engine.db.restore('pippo.db')

Versions

v 0.4.2 No backward compatibility with 3

Small fixes for Pypi

v 0.4.0 No backward compatibility with 3

• Action of users on users can be saved (see insert_social_action in dal.py)
• Various new metrics to monitor users' interaction (see e.g. get_social_actions in dal.py)
• No more embedded web service: use csrec-webapp
• TODO: make "social" recommendations based on users saving actions on each other
• Heavy refactoring
• Serialization and de-serialization of the data in a file for backup
• Data Abstraction Layers for memory and mongo.

v 0.3.15

• It is now a singleton, improved performance when used with, eg, Pyramid

v 0.3.14

• Minor bugs

v 0.3.13

• Added self.drop_db

v 0.3.12

• Bug fixed

v 0.3.11

• Some debugs messsages added

v 0.3.10

• Categories can now be a list (or passed as json-parseable string). This is important for, eg, tags which can now be passed in a REST API as:

  curl -X POST "http://127.0.0.1:8081/insertitem?id=Boo2&author=TheAuthor&cathegory=Horror&tags=scary,terror"
  

• Fixed bug in recommender_api example file

v 0.3.8

• Sync categories' users and items collections in get_recommendations

v 0.3.7

• Bug fixing for in-memory

v 0.3.5

• Added logging
• Added creation of collections for super-cold start (not even one rating, and still user asking for recommendations...)
• Additional info used for recommendations (eg Authors etc) are now stored in the DB
• _sync_user_item_ratings now syncs addition info's collections too
• popular_items now are always returned, even in case of no rating done, and get_recommendations eventually adjusts the order if some profiling has been done

.. _Pandas: http://pandas.pydata.org