Cold Start Recommender was developed because we needed a recommender with the following characteristics:
- Greedy. No previous data on Items or Users available, therefore
any information should be used --not just which Item a User likes, but also --in case of a book-- the correspondend 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 such User, but also for other Users. This means no batch computations, and in-memory database.
- Ready to use. Have a look at bin/recommender_api.py for starting
a webapp to POST information and GET recommendations.
CSRec should not (yet) be used for production system, but only for pilots, where statistics is so low that any filter (e.g. loglikelihood filter on the co-occurence matrix) do not make yet sense. It aims at gather data providing a recommendation experience.
TODO Future releases will include state of the art algorithms (about October 2014, a few months before our product is supposed to go public).
CSRec is written in Python, and under the hood it uses the Pandas library.
Table of Contents
from csrec.Recommender import Recommender
engine = Recommender()
# Insert Item with it properties (e.g. author, category...)
engine.insert_item({'_id': 'an_item', 'author': 'The Author'})
# Insert rating, indicating wich property of the Item should be used for producing recs
engine.insert_rating(user_id='a_user;, item_id='an_item', rating=4, item_info=['author'])
# Insert rating, indicating that only the property should be used for recs (e.g. initial users' profiling)
engine.insert_rating(user_id='another_user', item_id='an_item', rating=3, item_info=['author'], only_info=True)
You can use CSRec purely in-memory for testing or with MongoDB, which you can install on a tmpfs filesystem created in your RAM (on Linux, see http://edgystuff.tumblr.com/post/49304254688/how-to-use-mongodb-as-a-pure-in-memory-db-redis-style). If using a RAM partition, please make a replica set!
(Why using a replica set? Because you can have the primary DB in memory, and two other secondaries on disk. If the primary goes down, you still can use CSRec at lower performances, but without any data loss.)
Examples:
engine = Recommender() # Start in-memory recommender for testing
engine = Recommender(mongo_host='localhost', mongo_db_name='my_cold_rec') # ...with MongoDB, collections are created automatically
engine = Recommender(mongo_host='localhost', mongo_db_name='my_cold_rec', mongo_replica_set='recommender_replica') # as above, with replica
The Cold Start Problem originates from the fact that collaborative filtering recommenders needs 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:
1. It allows profiling with well-known Items without biasing the results. For instance, if a call to insert_rating is done in this way:
engine.insert_rating(user_id='another_user', item_id='an_item', rating=3, item_info=['author'], only_info=True)
CSRec will only register that 'another_user' likes a certain author, but not that s/he might like 'an_item'. This is of fundamental importance when profiling Users with a "profiling page" provided soon after log-in. If you ask Users if they like "Harry Potter" instead of "The Better Angel of Our Nature", and many chose Harry Potter, you don't want to make the Item "Harry Potter" even more popular than what alread is. You might only want to record that the user likes books for children sold as literature for adults.
CSRec does that because, unless you are Amazon or a similar brand, the co-occurence matrix is often too sparse to build any decent recommendation. In this way you start building multiple, denser, co-occurence matrices and use them from the beginning.
2. 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.
3. Any information is used *immediately*. The co-occurence matrix is updated as soon as a rating is inserted.
4. It tracks anonimous users, e.g. random visitors of a website 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.
What about users who would only receive a couple of recommended items? No problem! We'll fill the list with most popular items who were not recommended (nor rated by such user).
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.
v 3.7
Bug fixing for in-memory
v 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