20IO International Conference On Computer Design And Appliations (ICCDA 2010)

Research Paper Recommendation with Topic Analysis

Chenguang Pan, Wenxin Li

Computer Science Department
Peking University
Beijing, China
{cgp, lwx}@[Link]

Abstract-With the collaborative filtering techniques becoming there is even no appropriate dataset available for the overall
more and more mature, recommender systems are widely used evaluation of research paper recommendation algorithms.
nowadays, especially in electronic commerce and social A recommender system primarily exploits two kinds of
networks. However, the utilization of recommender system in information: users' ratings for items, profiles of users and/or
academic research itself has not received enough attention. A items. Memory-based recommendation makes use of only
research paper recommender system would greatly help
the users' ratings for items, say, the user-item matrix, with
researchers to find the most desirable papers in their fields of
each entry the rating of a user for a certain item. There are
endeavor. Due to the textual nature of papers, content
basically two kinds of memory-based recommendations:
information could be integrated into existed recommendation
user-based recommendation and item-based recommendation
methods. In this paper, we proposed that by using topic model
[6] [7]. User-based recommendation attaches higher weights
techniques to make topic analysis on research papers, we could
introduce a thematic similarity measurement into a modified
for users who share the similar rating patterns with the active
version of item-based recommendation approach. This novel
user and calculate the utilities of new items from these
recommendation method could considerable alleviate the cold weighted users. Item-based recommendation evaluate an
start problem in research paper recommendation. Our item's utility for a user by first selecting its most similar
experiment result shows that our approach could recommend items that have been rated by the user, then compute the
highly relevant research papers. utility as the weighted average of the rating of these similar
items. For an online recommender system, when the number
Keywords: collaborative jiltering, research paper of the users grows to a considerable amount, the computation
recommendation, latent dirichlet allocation, topic model, cold of similarity of every user pair's rating patterns would be
start very time consuming, thus item-based recommendation
could provide a feasible approach for it computes the
I. INTRODUCTION similarities between items offline. Reference [8] gives a nice
comprehensive survey on recommender systems.
Recommender system appeared in the early 1990s
Research paper recommendation would not only utilize
[1] [2] [3] [4], and now they are playing a significant role in
the user-item matrix but also the content information of the
people's daily lives. Online shopping site [Link] and
research papers. We use topic model techniques to make
online movie rental company Netflix recommend
topic analysis of research papers and introduced the
commodities and services by analyzing customers'
similarity over topics between research papers as the
preferences and online behaviors. Social networking services
thematic similarity. By incorporating the thematic similarity
like [14] may suggest that you may know some people or
and a modified version of item-based method, we could
would possibly like someone to become your friend. There
successfully generate highly relevant recommendations and
are also interests-based social networking site like [15]
greatly relieve the cold start problem.
provides users with recommendations of books, music CDs,
movies, and articles, and recommendations of people who
might share the similar tastes based on users' ratings for the II. COLLABORATIVE FILTERING
mentioned items.
References [5][6] [7][8][9] describe the details of the two
In colleges, universities and research institutions,
approaches of memory-based recommendations: user-based
graduate students, professors, and other researchers need to
find the papers most relevant to their research projects. Thus, recommendation and item-based recommendation, very well.
For later convenience we will still elaborate on these two
looking for the right papers to read becomes a very important
part of their academic lives. A research paper recommender approaches.
system will benefit these people in helping the most relevant In a typical recommender system, we have a user set U
consists of N users, an item set I consists of M items and a
papers and saving their precious time. Unluckily, paper
rating set R in which rlli stands for user u's rating for item i.
SII � I stands for the set of items user u has rated.
research recommender systems have not received enough
attention. The currently rare available work includes [16],
[17], [18], [19] etc. All these work have not taken thematic Collaborative filtering algorithms attempt to compute the
information of research papers into recommendation. And expected ratings or say, utilities of items that has not been
rated by the active user, then present the active user with the

978-1-4244-7164-5/$26.00 © 2010 IEEE V4-264 Volume 4

 2010 International Conference On Computer Design And Appliations (ICCDA 2010)

items with the highest utilities. We will describe below the Pearson correlation:
details of relevant memory-based recommendations: user­
based recommendation, item-based recommendation and our pearson(a, u) =
modified version of item-based recommendation.

A. User-based Recommendation (4)

In a collaborative filtering algorithm we are to compute

the utility Pai of the active user a on a given item i that has
not been rated by this user. Since the item i is not rated by
user a, thus we have irl. Sa. For user-based recommendation, Simple cosine:
the utility of an unrated item for a user is based on the ratings
of that item of the nearest neighbors of the user. A similarity
measure between users need to be defined and a set of (5)
nearest neighbors to be selected. Then we combine the
ratings of these neighbors on the item in some way.
The way we can define the similarity between users will
be described in A.I). Let sim(a, u) be similarity between There are also other similarity definitions as Manhattan
users a and u. The number of neighbors to be considered is similarity, Jaccard similarity, Tanimoto score, but we will
often set by a system parameter that we denote by K, K is not describe them here.
often set to N, the total number of users in the system. So the
set of neighbors of a give user a, denoted by Ta, is made of B. Item-based Recommendation
the K users that maximize their similarity to user a. In the item-based approach, Pai, the utility of an unrated
We define the utility of item i for user a as the weighted item i for the user a is computed as follows:
sum of the ratings of the nearest neighbors uE Ta that have We choose the most similar rated K items by user a for
already rated item i: item i. We denote the set of these K items as h. rk stands for
rating of item h E h. So the utility of the unrated item i is:

Pai=
L rk X sim(i, ik) (6)
L sim(i, ik)
(1)

Like the user-based approach, we also have a method to

compute utility of i using deviation from the mean rating
In order to take into account the difference in user of
given by user a.
rating scale by different users, predictions based on
If we use the recommendation provided by formula (6),
deviations from the mean ratings have been proposed. In that
we could easily find out that when the user has just rated one
case, Pai is computed using the sum of the user mean rating
item or has given a few items the same rating, utilities of all
and the weighted sum of deviations from their mean rating of
the unrated items will be the same as the given rating. In that
the neighbors that have rated item i:
case no unrated item could be given preference in
- L{uqYEsJsim(a,u). (rUi - rJ recommendation.
Pai� r + (2) Since we hope our recommendation could give reliable
L�,qYESjlsim(a,u)1
a
recommendations even when the user has just rated very few
items or even one item. We will use a modified version of
item-based recommendation:
r
u
represents the mean rating of user u:
Pai=
Lrkxsim(i, ik) (7)
K
This formula simply weights the ratings of items
(3) according to their respective similarity with the current item
being evaluated, and computes an average.

III. TOPIC ANALYSIS

1) Similarity Measurements Topic models [10] [11] are based upon the idea that with
The similarity defined between users or items include but in text corpora, a document is a mixture of topics, where a
are not limited to the following ways: topic is a multinomial distribution over words. The latent
dirichlet allocation (LDA) model (or "topic model") is a
generative model for documents: it specifies a simple
probabilistic procedure by which documents can be
generated. The mixing coefficients for each document and

V4-265 Volume 4
 2010 International Conference On Computer Design And Appliations (ICCDA 2010)

the word-topic distributions are unobserved and are learned rp [k, w], the multinomial coefficient for word w in topic
from data using unsupervised learning methods. k.
LDA can be described as finding a mixture of topics for jE[I, D], WE[I, W), kE[I, K].
each document, i.e., P(z I d), with each topic described by
words following another probability distribution, i.e., pet I z). We define the thematic similarity between two research
This can be formalized as papers p and as the cosine similarity defined over their
z q
P(1i I d)= L P(ti I Zi= j)P(Zi= j I d) (8) topic mixture coefficient vector and 8 [q :
8[p) ]
j='l
where P(ti I d) is the probability of the ith word for a given . . 8[p] . 8[q ]
Thematic Slm(p, = (12)
q
�

q]
1 11 1
document d and Zi is the latent topic. P(ti I Zi = j) is the 8[p]
probability of ti within topicj. P(Zi =j I d) is the probability of ) fJ[
generating a word from topic j in the document d. The As we know, there are a variety of ways to define the
number of latent topics Z is defined in advance and controls similarity between vectors, but in our experiment we simply
the granularity of differences among latent topics. LDA choose simple cosine similarity for the computation of
estimates the topic-word distribution pet I z) and the recommendations.
document-topic distribution P(z I d) from an unlabeled In our recommendation method we use thematic
corpus of documents using Dirichlet priors for the similarity to replace the original similarity get from user-item
distributions and a fixed number of topics. matrix, combined with modified item-based method, we
Reference Blei et al [11] introduced the LDA model could make research paper recommendations for users. This
within a general Bayesian framework and developed a recommendation method could considerably alleviate cold
variational algorithm for learning the model from data. start problem of recommender systems. In a typical
Reference Griffiths and Steyvers [12] subsequently proposed recomender system, cold start problem has two situations.
a learning algorithm based on Gibbs sampling. The Gibbs One is that when a user has just ranked very few items, it is
sampling approach iterates multiple times over each word ti very difficult to generate recommendations accurately due to
in document di, and samples a new topicj for the word based the little information of the user we have. The other situation
on Equation (8), until the LDA model parameters converge. is that when a new item has just been registered in the system,
no user has yet got time to rate the item, so that this item will

CTZ+ fJ DZ not get the slightest chance to be recommended. For most

Cd +a recommender system the first situation is more severe and
P(Zi =j I ti, di, Z_i) ex:
t,J ;l
" DZ (9) more common than the second situation. Our method could
L..t C1Z
If
+TfJ"
L..z Cd,z+Za relieve both the two situations greatly, especially the first
situation.
CTZ maintains a count of all topic-word assignments, (!Yz
counts the document-topic assignments, Z_i represents all IV. EXPERIMENT
topic-word and document-topic assignments except the
Our experiment is carried in the following procedure and
current assignment Zi for word ti, a and fJ are the we will demonstrate the result all along, and we will make
hyperparameters for the Dirichlet priors, serving as data and result available online for the convenience of other
smoothing parameters for the counts. Based on the counts researchers.
the posterior probabilities in Equation (8) can be estimated as We take 122 research papers from the research paper
follows: collection of our laboratory, mainly on biometrics research,
as the research papers that are going to be used in our
recomender system. They are all in Portable Document
(10) Format (PDF).
Then all these PDF files are converted into text files

CDZ
using a Java library called Apache PDF Box [22]. These text
. I di) = d,;+a can be download at [20].
P(Zi = ] ---':-::--
- - (11) We preprocessed these text files and use a perl script by
Lz C;z+Za David Newman to make these research papers into uel Bag­
Of-Words Dataset format [21]. The dataset generated by us
Simply put, we are going to use Gibbs sampling methods is available here [20].
to compute the mixture coefficients of each topic in a We implemented the Gibbs sampling algorithm to
document, and the coefficients of a topic's multinomial process the dataset we generated. We took 40 topics, and
distribution over words.
a =2.0/40, fJ =0.01 as the parameters in our inference. We
Say we have D documents in the corpora, W unique
words appeared within, N words in total, and we suppose successfully computed the coefficient of topic mixture in
there are K topics in the corpora. Using Gibbs sampling for each research paper and the multinomial coefficient for each
LDA we could obtain the results as follows:
8 U, k], the mixture coefficient for topic k in documentj.

V4-266 Volume 4
 2010 International Conference On Computer Design And Appliations (ICCDA 2010)

TABLE!. TOPICS MIXTURE IN Doc. 59 TABLE Ill. WORDS GENERATION PROBABILITY IN TOPIC 32

TOPIC ID TOPIC WEIGHT WORDID PROBABILITY WORD

32 0.70430677 3240 0.08928722 sensor
27 0.12807462 745 0.05155874 circuit
26 0.06082450 731 0.04073623 chip
5 0.04884846 1565 0.03562560 fingerprint
38 0.02535698 2704 0.03487404 pixel
8 0.01476278 3237 0.02705778 sensing
35 0.00831414 1870 0.02014340 image
23 0.00370797 2056 0.01773840 japan
6 0.00324735 1545 0.01668621 fig
10 0.00140488 633 0.01398058 calibration
33 0.00048365 3301 0.01232714 signal
11 0.00002303 1562 0.01067370 finger
word to be generated III each tOpIC. Results are aVaIlable at 3541 0.01067370 surface
[20). Table l. shows the coefficients of topic mixture in Doc. 2712 0.01007245 plate
404 0.00992214 area
59: "A Novel Scheme for Fingerprint Identification" by
781 0.00992214 cmos
Tsong-Liang Huang et aI. Only 12 topics with highest
1323 0.00977183 element
mixture weight are show in Table I. It is very obvious that 412 0.00886995 array
this research paper is mainly about Topic 32 and Topic 27. 401 0.00871964 architecture
Topics 32 and 27's probability to generate each word is 1551 0.00856932 film
shown in Table 3 and 4. Only 20 words with the highest
probability are shown.
Then we computed the thematic similarities between all TABLE IV. WORDS GENERAnON PROBABILITY IN TOPIC 27
possible pairs of research papers. Results are available here
WORDID PROBABILITY WORD
[20). We can easily find that the thematic similarity between 1857 0.05555788 ieee
Doc. 59 and Doc.61 is 0.99821355. The topic weights of 2844 0.04356642 processing
Doc. 61 are shown in Table. 2. Doc. 61 is "A Pixel-Level 3622 0.02695439 test
Automatic Calibration Circuit Scheme for Capacitive 3301 0.02618429 signal
Fingerprint Sensor LSIs" by Hiroki Morimura et aI. We can 3085 0.02508416 research
easily find out that these two papers are quite similar in their 2969 0.02255385 received
2841 0.02211380 process
main topics.
1355 0.02035358 engineering
In the recommendation scenario we consider a user that 3597 0.01760325 technology
has only rated 3 papers: Doc. 14, 31, 59, with respective 3795 0.01617307 university
rating 1, -2, 2. We take rating -2, -1, 0, 1, 2 to stand for 2311 0.01320271 member
"awful", "bad", "so-so", "good", "awesome". 1317 0.01155251 electrical
We successfully use our recommendation algorithm to 1948 0.01133248 information
generate the recommendations with the highest predicted 1047 0.01100244 data
utility for this user. The recommendation result is available 1094 0.01089243 degree
1135 0.01056239 design
here [20] and partly shown in Table 5 on next page.
1189 0.00946225 digital
Relevant Doc. 14, 31, 58, 55, 49, 48 are "Estimation and 1992 0.00946225 interest
sample size calculations for correlated binary error rates of 1111 0.00935224 department
biometric identification devices" by Michael E. Schuckers, 872 0.00891219 computer
"Validating a Biometric Authentication System: Sample Size
Requirements" by Sarat C. Dass et aI., "A Multichannel Approach to
Fingerprint Classification" by Anil K. Jain et aI., "A
TABLE I!. TOPICS MIXTURE IN Doc. 61
Fingerprint Verification System Based on Triangular
TOPIC ID TOPIC WEIGHT Matching and Dynamic Time Warping" by Zsolt Mikos
32 0.65057699 Kovacs-Vajna, "A 600-dpi Capacitive Fingerprint Sensor
27 0 13236567 Chip and Image-Synthesis Technique" by Jeong-Woo Lee et
26 0.06168410 aI., "Theoretical statistical correlation for biometric
38 0.03968626
identification performance" respectively.
5 0.03247386
The recommendation list in Table 5 shows us that under
24 0.02490083
35 0.01660656
the influence of Doc. 59, which got the highest rating 2, the
21 0.01011540 Doc.61 which is highly thematically similar to Doc. 59 as
28 0.00903354 demonstrated above, rank the first in the recommendation list.
8 0.00867292 The utility of Doc. 61 is not as high as close to Doc. 59's
16 0.00578796 rating is due to the influence of the worst rated item: Doc. 31.
23 0.00398485

V4-267 Volume 4
 2010 International Conference On Computer Design And Appliations (ICCDA 2010)

TABLE V. TOP 5 RECOMMENDATIONS [4) lA. Konstan et a!. "GroupLens: applying collaborative filtering to
Usenet news", Comm. ACM, VoI40,no.3,77-87,1997
DOC ID PREDIC TED UTILITY
[5) lS. Breese, D. Heckerman and C. Kadie, "Empirical analysis of
61 0.66960195
predictive algorithms for collaborative filtering", Proc. 14th Conf
58 0.66002356
Uncertainty in Artificial Intellligence, July 1998
55 0.49133007
[6) [Link], G. Karypis, lKonstan and I Riedl, "Item-Based
49 0.37691359
collaborative filtering recommendation algorithms", Proc. 10th In!'I
48 0.37416252
WWW Conf. 2001
[7) G. Linden, B. Smith and 1 York, "Amazon. com recommendations:
The thematic similarity between Doc. 61 and lowest rated item-to-item collaborative filtering", IEEE Internet Computing,
Doc. 3 I reduced the utility of Doc. 61 to below 1 but still [Link]. 2003
remains the highest. We can also see that utility of research [8) G. Adomavicius and A. Tuzhilin, "Toward the next generation of
papers with highest thematic similarity to Doc. 31 have very recommender systems: a survey of the state-of-the-art and possible
low or even negative utilities. For example, Doc. 21 named Extensions", IEEE Trans. on Knowledge and Data Engineering 17,
(2005), 634-749
"Effects of User Correlation on Sample Size Requirements"
[9) L. Candillier, F. Meyer and F. Fessant, "Designing specific weighted
by Sarat C. Dass and Anil K. Jain, is predicted a utility of -

similarity measures to improve collaborative filtering systems",

0.55714472, because this paper has a high thematic ICDM 2008, LNAI 5077, 242-255
similarity at 0.99470842 with Doc. 31. [10) T. Hoffman, "Probabilistic latent semantic analysis", UAI'99
[II) D. Blei, A. Ng and M. Jordan, "Latent dirichlet allocation", Journal of
V. CONCLUSION
Machine Learning Research, 3:993-1022, Jan. 2003
Recommender systems are a powerful technology for [12) T. Griffiths and M. Steyvers, "Finding scientific topics", PNASI USA,
electronic commerce and social networks. They can help a 101 Supp 115228-5235, Apr. 2004
business to increase sales revenue and help customers to find [13) I. Porteous et a!. "Fast collapsed Gibbs sampling for latent dirichlet
product to their liking, or help people to find friends who allocation", KDD'08 Aug. 2008
share the same taste for things. At the same time their [14) [Link]
potential impact on scientific research is largely unexplored. [15) [Link]
We proposed a novel method for research paper [16) C. Basu, H. Hirsh, W. W Cohen and C. Nevill-Manning, 'Technical
recommendation. We proved by our experiment that by paper recommendation: a study in combining multiple information
sources", Journal of Machine Learning Research 1 (2001) 231-252
making topic analysis on research papers and introducing
thematic similarity we could recommend highly relevant [17) T Y Tang and G. McCalla, "A multidimensional paper
recommender - experiments and evaluations", 2009 IEEE Internet
papers and considerably alleviate the cold start problem. Computing
Even when the user has just rated very few papers, we could
[18) M. Gori and A. Pucci, "Research paper recommender systems: a
still generate satisfactory recommendations. random-walk based approach", Proc. of 2006 IEEEIWICIACM Int'l
Conf on Web Intelligence.
ACKNOWLEDGMENT
[19) T. Bogers and A. van den Bosch, "Recommending scientific articles
We would like to thank the developers of Apache using CiteULike", ACM Recsys'08
PDFBox, and Dave Newman for his nicely working perl [20) [Link] com/uc?id=OBw3EfWDJYdmXZDg3NDViYzctN
2E3MiOOODJkLWI2YTMtZjYxNDY3NDNkOTQx&export=downlo
script. Thank you for your work, so that we don't need to do
ad&hl=en
everything from scratch.
[21) [Link] edulmlldatasets/Bag+of+Words

REFERENCES [22) [Link]

[I) D. Goldberg, D. Nichols, B. M. Oki and D. Terry, "Using

collaborative filtering to weave an information tapestry",
Communication of the ACM 35 (1992), 61-70
[2) P. Resnick et a!. , "GroupLens: an open architecture for collaborative
filtering of netnews", Proc. ACM 1994 Conf Computer Supported
Cooperative Work, ACM Press, 1994, 175-186
[3) UShardanand, P. Maes, "Social information filtering: algorithms for
automating 'word of mouth''', ACM 1995 Conf Human Factors in
Computing Systems, VolI, 210-217

V4-268 Volume 4