In this paper, we present work-in-progress on applying user pre-filtering to speed up and enhance recommendations based on Collaborative Filtering. We propose to pre-filter users in order to extracta smaller set of candidate neighbors, who exhibit a high number of overlapping entities and to compute the final user similarities based on this set. To realize this, we exploit features of the high-performance search engine Apache Solr and integrate them into a scalable recommender system. We have evaluated our approach on a dataset gathered from Foursquare and our evaluation results suggest that our proposed user pre-filtering step can help to achieve both a better runtime performance as well as an increase in overall recommendation accuracy.

1. Neighborhood Troubles:
On the Value of User Pre-Filtering To Speed Up and Enhance
Recommendations
Emanuel Lacic, Dominik Kowald & Elisabeth Lex
Social Computing
Know-Center GmbH / Graz University of Technology

2. Motivation
RecSys in the Big Dataera:
• Need to analyze a lot of data
– With many different data features and recommendable entities
• Handle frequent streams of new data
• Demand for real-time performance

3. User-based Collaborative Filtering
• UB-CF is (still) one of the most commonly utilized approaches
in both industry and academia
• Accomplished in two steps (both parallelizable):
– Determine k-nearest neighbors (similar users) for a target user 𝑢 𝑡
– Recommend entities of these users that the target user 𝑢 𝑡
has not yet consumed

4. User-based Collaborative Filtering
• UB-CF is (still) one of the most commonly utilized approaches
in both industry and academia
• Accomplished in two steps (both parallelizable):
– Determine k-nearest neighbors (similar users) for a target user 𝑢 𝑡
– Recommend entities of these users that the target user 𝑢 𝑡
has not yet consumed
• A lot of data? Parallelize and scale UB-CF?
– What if it’s not desirable or possible to
allocate additional computing resources?

5. Bottleneck: Neighbor processing
• We need to fetch the history of every neighbor and calculate the
similarity with the target user 𝑢 𝑡
• Similarities need to be sorted in order to pick the top-k similar users
– Common implementations of such operations have a
complexity of 𝑂 𝑛 × log 𝑛
– Could also be improved to a complexity of 𝑂 𝑛 + 𝑘 × log 𝑛
by additionally implementing a partial sort

6. Bottleneck: Neighbor processing
The larger the neighborhood of 𝑢 𝑡 is, the larger the impact
on the runtime performance could be!

7. Adaptation: User Pre-Filtering
• Adapt the 1st step of UB-CF by pre-filtering the candidate set
of possible similar users
• Greedy strategy
– Find the top-N candidate users which have the highest overlap
of entities that were interacted with
• Aim to positively influence the runtime performance of those users,
which exhibit many neighbors
– Increase the probability that users with a high overlap will in the end
be picked as the top-k similar users
𝑂𝑉(𝑢 𝑡, 𝑢 𝑐) = ∆(𝑢 𝑡) ∩ ∆(𝑢 𝑐)

8. Implementation details: ScaR
• Scalable entity recommender
framework that leverages the
Apache Solr search engine
http://scar.know-center.tugraz.at

9. Implementation details: ScaR
• Data Modification Layer
– Agent between the framework
and Apache Solr
• Recommender Customizer
– Configuration depending on
the entity recommendation
scenario (e.g., similarity
metric definition)
• Recommender Engine
𝑝𝑟𝑒𝑑 𝑢 𝑡, 𝑒 = 𝑠𝑖𝑚(𝑢 𝑡, 𝑢 𝑐)
𝑢 𝑐∈𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑟𝑠(𝑢 𝑡,𝑒)

10. Implementation details: ScaR
• Need Efficient User Pre-Filtering
• Utilize Solr’s facet functionality
– Arrangement of results into
categories
– Possible to restrict with filters
• Reduce the search space and
exactly get the desired 𝑂𝑉(𝑢 𝑡, 𝑢 𝑐)
– Facet on users
– Filter on entities of target user 𝑢 𝑡

11. Experimental Setup
• Foursquare Dataset
– Sparse with a right-tailed distribution that
has sharper peaks and broader tails
• Neighborhood size
– Average: 764
– Median: 4
– Maximum: 125,046
• Evaluation on all users that rated at
least 11 venues
– 58,046 users in total
– Withheld 10 rated venues from the dataset to be
predicted in the test set
– Evaluate on ranking accuracy (Precision, Recall, nDCG) and runtime
Measure Value
# Users 2,153,471
# Entities 2,809,581
# Ratings 1,143,092
Density 0.000015
Skewness 227.53
Kurtosis 62,844.43

12. Experimental Setup
• Hypothesis: the same interval of values for top-k similar users is valid
for a greedy pick of candidate neighbors (i.e., between 20 and 60)
Approach Description
Most Popular A baseline that recommends most popular items
𝐶𝐹𝐹𝑢𝑙𝑙 UB-CF which calculates similarities for all neighbors
𝐶𝐹𝑂𝑉=20 Greedy pick of top-20 overlapping users
𝐶𝐹𝑂𝑉=40 Greedy pick of top-40 overlapping users
𝐶𝐹𝑂𝑉=60 Greedy pick of top-60 overlapping users
𝐶𝐹𝑂𝑉=80 Greedy pick of top-80 overlapping users
𝐶𝐹𝑂𝑉=100 Greedy pick of top-100 overlapping users

13. Results

14. Results
Approach 𝑇 (𝑚𝑠) 𝜎 (𝑚𝑠) P@10 R@10 nDCG@10 UC
Most Popular 78.59 20.00 .0285 .0285 .0232 100 %
CollaborativeFiltering
𝐶𝐹𝐹𝑢𝑙𝑙 2,053.45 9,600.63 .0611 .0527 .0316 66.56 %
𝐶𝐹𝑂𝑉=20 59.56 60.08 .0586 .0541 .0318 65.87 %
𝐶𝐹𝑂𝑉=40 65.47 69.61 .0689 .0645 .0378 66.21 %
𝐶𝐹𝑂𝑉=60 74.62 85.83 .0724 .0678 .0396 66.10 %
𝐶𝐹𝑂𝑉=80 82.40 102.75 .0707 .0661 .0386 65.62 %
𝐶𝐹𝑂𝑉=100 87.38 115.17 .0693 .0646 .0373 65.70 %

15. Conclusion and Future Work
• Integrating User Pre-Filtering can speed up and enhance
recommendations
• Depending on the data, it can also increase the overall accuracy
– Potentially can help to reduce noise out of the candidate entities
• Need to validate results in a more comprehensive study
using data with different types of entities
• Plan to validate our approach in course of an online study
within the Analytics for Everyday Learning (AFEL) project
– See our AFEL-REC paper in the SIR workshop @ CIKM

16. Questions / suggestions ?
Emanuel Lacić
elacic@know-center.at
Social Computing / Know-Center