Exploring a weighted ensemble of different recommendation engines such as User based, Item based, Slope-one based and Content based for the MovieLens 100K, 1M, 10M datasets. Achieved an improvement of 11.59% with the ensemble. Also implemented the item based recommender in a distributed manner using Apache Mahout.

1. MovieLens Recommendation
Engine

Outline:

Task & Dataset

Techniques

Results

Scalability

Conclusion
­ Ambarish Hazarnis
­ Vibhor Mathur

2. Task
Predict the rating, a user will give to a movie
which he hasn’t seen yet.
Recommend the movies with the highest scores.

3. Dataset
MovieLens 100k
•
100000 ratings by 943 users on 1682 items. Each user has rated at
least 20 movies.
•
Movies can be in several genres at once.
•
Demographic information about the users (age, gender, occupation).
Evaluation

Root Mean Squared Error

4. Techniques

Collaborative

User Based

Item Based

Slope one

Content Based

User Based – Age, Occupation, Gender

Item Based – Genre

Ensemble

Committee

Weighted

Distributed

5. Results

RMSE
Recommender Error
User Based 1.227
Item Based 0.664
Slope One 0.587
User Content Based 0.649
Item Content Based 0.639

6. Ensemble

Commitee
Recommender RMSE
Collaborative Based 0.595
Content Based 0.612
Collaborative + Content 0.594

Weighted
Recommender RMSE
Collaborative Based 0.747
Content Based 0.612
Collaborative + Content 0.663

7. Slope One

Principle:
Preferences for new items is based on average difference in the
preference value between a new item and the other items the user
prefers.

For two items I1 and I2, rating of user1 for I2 who has rated I1,

Count Weighting- Weight heavily those differences that are based on
more data.

Standard Deviation- A low std dev means will translate to a higher
weight.

8. User Content Based
User: Gender, Occupation, Age
Principle - Two users having similar gender, occupation or age group share similar taste.
Similarity -
Taking advantage of user-specific knowledge.
Custom Similarity metric for user similarity.
Assigning different weightage to gender, occupation and age similarities to deduce this
custom similarity.
This custom similarity metric can be paired with a standard
GenericUserBasedRecommender.
Discard all rating related information from metric computation.

9. Item Content based
Item: Multiple genre
Principle - Two movies of similar multiple genres will be similar.
Similarity -
Taking advantage of item-specific knowledge.
Custom Similarity metric for movie similarity.
Similarity is deduced based on the degree of similarity of genres.
This custom movie similarity metric can be paired with a standard
GenericItemBasedRecommender.

10. Ensemble

Ensemble

Uses phenomenon of 'Wisdom of crowds'

Commitee
Unweighted average of predicted ratings of all recommenders

Weighted

Higher weights for better recommenders

If Ei is the error of recommender, let Ai and Wi denote its accuracy and
weight respectively.

11. Scalability-1

Case Study: Item Based Recommender using Coocurrence as similarity.
4(2.0) + 3(0.0) + 4(0.0) + 3(4.0) + 1(4.5) + 2(0.0) + 0(5.0) = 24.5
Distributed computation helps
by breaking up a problem that’s too big for
one server into pieces that several smaller
servers can handle

12. Scalability-2

Sums the products of co-occurrences and preference values.

How is it suitable for distributed?
Computing the resulting recommendation vector only requires
loading one row or column of the matrix at a time
User's
Ratings
Cooccurence
Matrix
Item Based Rec
Top N
Recommendations
Apache Mahout: Provides scalable Machine learning
libraries
Package:
org.apache.mahout.cf.taste.hadoop.item.RecommenderJob
(5 MapReduce jobs)
Recommendations for User 122:
[ 9 : 5.0, 546 : 5.0, 568 : 5.0, 527 : 5.0, 515 : 5.0, 514 : 5.0, 511 : 5.0, 498 : 5.0]

13. Conclusion

Slope one recommender worked best but it is also computationally
very expensive.

Content based approach gave better results than plain collaborative
approach. However, the former is domain-specific.

A ensemble of simple learners gave comparable result.

More learners in a ensemble results in better predictions.

14. Thank YouThank You