1. Google News Personalization: Scalable Online Collaborative Filtering
Abhinandan Das, Mayur Datar, Ashutosh Garg
WWW2007, May 8-12, 2007, Banff, Alberta, Canada
Presented By,
Naveenkumar Selvaraj

2. Google News A news aggregator website Personalized news-
Records search queries and news clicks
Makes them accessible online
Recommended for you – top stories based on the user’s past click history

4. Collaborative Filtering
Learn User Preferences
Make recommendations based on
User Data
Community Data
Ex : Amazon.com

5. Contribution of this paper
Aims to present recommendations to signed-in users for Google News based on
User’s click history
Click history of the community
Generate recommendations using three approaches- collaborative filtering using
MinHash Clustering
Probabilistic Latent Semantic Indexing (PLSI)
Covisitation counts

6. Problem Statement
Given click history for N users and M items and given a specific user u with click history Cu consisting of stories, recommend K stories to the user, which he/she may be interested in reading.

7. Challenges
Scalability – has several million users visiting it and several million items
Large number of item churns
Most systems including Amazon assume amount of churn – static or minimal
For Google News – large amount of churn. Stories of interest are the ones that appeared in the last couple of hours
Strict Timing Requirements
Strict Response time requirement – less than a second
Must also take into account – time for
Generating news story clusters by accessing various indexes storing the content
Generating the HTML content as a response for HTTP request
Hence only a few milliseconds left

8. Challenges ( Contd..) Treating click as positive vote
More noisy – than giving 1-5 ratings (like Amazon.com )
Clicks do not say anything about the user’s negative interest

9. Related Work Recommender Systems
Content-based Filtering – Items similar to the ones rated highly by the user recommended
Collaborative Filtering – Two types
Memory Based –
Rating predictions based on users’ past rating
Weighted average of the ratings given by other users ( weight proportional to similarity between users)
Challenge – Make it more scalable

10. Related Work (Contd..) Model-based
Model the users based on past ratings
Use the models to predict ratings on unseen items
Shortcomings of earlier work in Model-based
Categorizes each user into only one class
User may have different tastes for different topics

11. Work Done Mix of Memory Based and Model Based algorithms used.
Model Based – MinHash and PLSI ( Probabilistic Latent Semantic Indexing)
Memory Based – Covisitation
Each algorithm - assigns a numeric score to a story
Scores combined – obtain a ranked list of stories

12. MinHash Probabilistic Clustering technique.
Assigns pair of user to the same cluster with probability proportional to the overlap between the set of items users voted for
Similarity between users ui and uj represented as
Doing this in real-time – not scalable

13. MinHash LSI ( Locality Sensitive Hashing)
Randomly permute set of items S.
For each user ui, compute hash value h(ui) – index of the first item under the permutation that belongs to user’s item set.
Probability that two users will have same hash function is equal to similarity between users.
Each hash bucket corresponds to a cluster , which puts two users in the same cluster with similarity S(ui,uj)
Can concatenate more hash keys- underlying clusters more refined and average similarity of users within clusters more.
Has high precision but low recall

14. MinHash
Improve recall by repeating above step in parallel multiple times
Generating random permutations over millions of items not feasible
Generate set of independent random seed values, one for each MinHash function- map each news story to a hash-value computed using Id of news story, seed value
MapReduce used
Handles large amounts of data in short period of time and makes it more scalable

15. PLSI (Probabilistic Latent Semantic Indexing)
Models users and items as random variables –taking values from space of all possible users and items
Relationship learned by modeling as joint distribution of users and items
Hidden variable Z captures this relationship
where |Z|=L

16. PLSI Mapreducing EM algorithm
EM algorithm – used to learn maximum likelihood parameters of the model.
When the number of users and items are very large of the order millions, memory requirements for the Conditional Probability Distributions is huge.
Solution : MapReduce

17. PLSI
Map-reduce Framework

18. PLSI
R x K grid
Users and items sharded into R and K groups respectively
Click data corresponding to the (u,s) pair sent to (i,j) th machine – u belongs to ith shard and s belongs to jth shard
Each machine has to load only (1/R)th of user CPDs and (1/K)th of item CPDs
Reducer shard receives key-value pair and computes the values for the next iteration
Problems with PLSI :
If new user/items added – whole model needs to be retrained
Even after parallelizing, it is not real time

19. Covisitation
Event in which two stories are clicked by the same user within certain time interval
Graph
Nodes represent items
Weighted edges – number of Covisitation instances
Graph maintained as an adjacency list in Bigtable keyed by item-id (Bigtable maintains data in lexicographic order by row key)

20. Covisitation
Fetch every item in users click history, limited to past few hours
For every item si, Lookup for the pair si , s in the adjacency list for si stored in the Bigtable.
Add the value stored in the entry to the recommendation score normalized by the sum of all entries.
Covisitation scores normalized between 0 and 1 by linear scaling

21. System setup

22. System setup (contd..) Two types of requests handled Recommend Request
Update Statistics Request

23. Evaluation
Precision-recall curves for the Movie Lens Dataset

24. Evaluation
Precision-recall curves for the GoogleNews Dataset

25. Evaluation
Live traffic click ratios for different algorithms with baseline as Popular algorithm

26. Evaluation
Live traffic click ratios for comparing PLSI and MinHash algorithms

27. Conclusion
The paper present the algorithms behind scalable real time recommendation engines
Novel approaches to clustering over dynamic datasets using MinHash and PLSI with MapReduce Framework for scalability implemented
Evaluation on live traffic was done over a large fraction of Google News traffic over a period of several days

28. Future Work
Suitable learning techniques to determine how to combine the scores from different algorithms
Exploring cost-benefit tradeoffs of using higher order covisitation statistics

29. Questions?