Using Neighborhood Pre-computation to Increase Recommendation
Efficiency
Vreixo Formoso, Diego Fern
´
andez, Fidel Cacheda and V
´
ıctor Carneiro
University of A Coru
˜
na, Campus de Elvi
˜
na s/n, 15017 A Coru
˜
na, Spain
Keywords:
Collaborative Filtering, Nearest Neighbors, Performance.
Abstract:
Collaborative filtering is a very popular recommendation technique. Among the different approaches, the k-
Nearest Neighbors algorithm stands out by its simplicity, and its good and explainable results. This algorithm
bases its recommendations to a given user on the opinions of similar users. Thus, selecting those similar
users is an important step in the recommendation, known as neighborhood selection. In real applications with
millions of users and items, this step can be a serious performance bottleneck because of the huge number of
operations needed. In this paper we study the possibility of pre-computing the neighbors in an offline step,
in order to increase recommendation efficiency. We show how neighborhood pre-computation reduces the
recommendation time by two orders of magnitude without a significant impact in recommendation precision.
1 INTRODUCTION
Recommender systems are a popular technique in
fields such as e-commerce, where they help users to
find the products they need. A particularly success-
ful technique is collaborative filtering, that computes
high-quality recommendations to a user, based on the
opinions of other users with similar tastes or interests.
Among the different algorithms developed, the k-
Nearest Neighbors (kNN) approach is very popular
because it is simple, intuitive (which allows to justify
a recommendation decision), and does not require a
training step (Desrosiers and Karypis, 2011). It first
selects a set of neighbors, that is, the set of users most
similar to the user the system is generating a recom-
mendation to (known as the active user). The items
most highly rated by those neighbors are the ones rec-
ommended.
The neighborhood selection is a computationally
intensive step, that can take a long time. It requires to
compute the similarity between the active user and ev-
ery remaining user. Each similarity computation also
requires to compare the opinions or ratings of both
users. In real applications, with millions of users and
items, this can take several seconds, which is unac-
ceptable in many cases where the recommendations
need to be generated in real time. Although tech-
niques such as compression (C
¨
oster and Svensson,
2002) have been proposed to optimize it, neighbor-
hood computation remains a very expensive step.
A practical solution is to pre-compute the neigh-
borhood in an offline step, storing it in an index-like
structure for later usage. This can significantly re-
duce the recommendation time. However, the neigh-
borhood should be updated to include new user opin-
ions. Otherwise, the neighbors used for recommen-
dation may not reflect the current user tastes, and that
might negatively influence the quality of the recom-
mendations.
In this paper we study the impact of neighbor-
hood pre-computation, both on computational effi-
ciency and recommendation quality. We show how
it is a very effective technique to reduce recommen-
dation time, without significantly reducing the recom-
mendation quality.
2 EXPERIMENTS AND RESULTS
For our experiment we have used the Netflix dataset
(Bennett and Lanning, 2007), a popular dataset from
the movie recommendation domain. It contains over
100 million ratings from 480,189 users to 17,770
movies, collected between October 1998 and Decem-
ber 2005. In order to evaluate the impact of neighbor-
hood pre-computation, we have studied the evolution
of the precision of the results with the time elapsed
after pre-computation.
From all ratings available in the dataset, we have
taken those ratings given before January 1st, 2005,
333
Formoso V., Fernández D., Cacheda F. and Carneiro V..
Using Neighborhood Pre-computation to Increase Recommendation Efficiency.
DOI: 10.5220/0004139703330335
In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2012), pages 333-335
ISBN: 978-989-8565-29-7
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
and we have pre-computed the neighborhood taking
only those ratings into account. Then, we evalu-
ate the algorithm considering all ratings up to Febru-
ary 1st, March 1st, and so on. For the evaluation,
we have considered two situations: one with neigh-
borhood pre-computation, where the previously com-
puted neighborhood (with ratings up to January 1st)
is used, and another where the neighborhood is com-
puted at recommendation time (and thus all ratings up
to the given month are considered). In either case, for
recommendation we consider all the ratings available
at that time. This way, we simulate an environment
where the neighborhood is computed once at the be-
ginning of the year, but the rating matrix is being up-
dated constantly. We have performed the evaluation
with 1, 000 randomly selected users.
First, we have studied how neighborhood pre-
computation can speed up recommendation time. For
our experiments, we have used a PC with a Intel Pen-
tium 4 CPU at 3.20 GHz and 256 MiB of RAM. Us-
ing an old machine is an approach commonly used for
efficiency evaluation in Information Retrieval (Badue
et al., 2007) when the dataset used is significantly
smaller than the amount of data in real applications.
Results are shown in Figure 1. As expected,
the usage of neighborhood pre-computation signif-
icantly reduces recommendation time. In average,
recommendation is computed two orders of magni-
tude faster, which is a very important achievement.
Moreover, with neighborhood pre-computation the
required time remains more or less constant among
months, even though the number of ratings increases.
On the other hand, with no pre-computation it signif-
icantly increases with the number of ratings. That is,
the neighborhood computation time is more affected
by the number of ratings than the final recommenda-
tion step, which makes sense because in that final step
only a few users (the neighbors) are actually consid-
ered.
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Jany Jun Dec
0 20 40 60 80 100 120
No pre−computation
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Jan Jun Dec
0.0 0.1 0.2 0.3 0.4
With pre−computation
Recommendation time (s.)
Figure 1: Recommendation time (seconds) with and with-
out neighborhood pre-computation. Note the different scale
in each chart.
We have also evaluated the precision and recall of
the recommendations, in order to study the evolution
of the quality with the time elapsed after neighbor-
hood pre-computation. If the precision dropped very
fast, this technique would be not very useful, because
the pre-computation step would need to be done very
often. However, as seen in Table 1, this is not the
case. Both precision and recall remain similar with
and without pre-computation
1
, without statistical sig-
nificant differences between them. While updating
the rating matrix is very important (in order to recom-
mend new products, for example), dealing with an old
neighborhood seems to have almost no impact in rec-
ommendation quality. Of course, the actual threshold
where an outdated neighborhood begins to negatively
impact quality is domain-dependent. While a several
months old neighborhood is not a problem in the stud-
ied case, other domains might require a shorter neigh-
borhood update time.
Table 1: Precision@5 and Recall@5 with and without pre-
computation.
P@5 R@5
With Without With Without
Jan 1.28 1.28 0.13 0.13
Feb 0.99 0.90 0.12 0.08
Mar 1.16 1.39 0.13 0.15
Apr 0.98 1.21 0.12 0.17
May 0.72 0.76 0.07 0.08
Jun 0.75 0.81 0.09 0.12
Jul 1.03 0.65 0.46 0.12
Ago 0.12 0.39 0.03 0.05
Sep 0.26 0.32 0.08 0.27
Oct 0.22 0.21 0.14 0.12
3 CONCLUSIONS
In this paper we have evaluated the benefits of neigh-
borhood pre-computation. We have shown how this
technique can reduce the recommendation time of
k-Nearest Neighbors algorithms by two orders of
magnitude, without a significant impact in the rec-
ommendation list quality. These results show that
real applications can benefit from neighborhood pre-
computation techniques with no important drawback
in terms of precision. In the future, we plan to ex-
tend this research to further domains. We also plan to
study the impact on different metrics, and with differ-
ent update strategies.
1
Note that bad precision in the last months is related to
the evaluation methodology, as there are few relevant rat-
ings after that time. This is a well-known limitation of of-
fline evaluation (Cacheda et al., 2011).
KDIR2012-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
334
ACKNOWLEDGEMENTS
This research was supported by the Ministry of Edu-
cation and Science of Spain and FEDER funds of the
European Union (Project TIN2009-14203).
REFERENCES
Badue, C. S., Baeza-Yates, R., Ribeiro-Neto, B., Ziviani,
A., and Ziviani, N. (2007). Analyzing imbalance
among homogeneous index servers in a web search
system. Inf. Process. Manage., 43:592–608.
Bennett, J. and Lanning, S. (2007). The netflix prize. In
KDDCup ’07: Proceedings of KDD Cup and Work-
shop, page 4, San Jose, California, USA. ACM.
Cacheda, F., Carneiro, V., Fern
´
andez, D., and Formoso, V.
(2011). Comparison of collaborative filtering algo-
rithms: Limitations of current techniques and propos-
als for scalable, high-performance recommender sys-
tems. ACM Trans. Web, 5:2:1–2:33.
C
¨
oster, R. and Svensson, M. (2002). Inverted file search
algorithms for collaborative filtering. In SIGIR ’02:
Proceedings of the 25th annual international ACM SI-
GIR conference on Research and development in in-
formation retrieval, pages 246–252, New York, NY,
USA. ACM.
Desrosiers, C. and Karypis, G. (2011). A comprehensive
survey of neighborhood-based recommendation meth-
ods. In Ricci, F., Rokach, L., Shapira, B., and Kan-
tor, P. B., editors, Recommender Systems Handbook,
pages 107–144. Springer.
UsingNeighborhoodPre-computationtoIncreaseRecommendationEfficiency
335
