In recent years we can observe a steady growth of scienti c publications in increasingly diverse scienti c elds. Current digital libraries and retrieval systems make searching these publications easy, but determining which of these are relevant for a speci c person remains a challenge. This becomes even harder if we constrain ourselves to publicly available meta-data, as complete information (in particular the fulltext) is rarely accessible due to licensing issues. In this paper we propose to model researcher pro le as a multivariate linear regression problem leveraging meta-data like abstracts and titles in order to achieve e ective publication recommendation. We also evaluate the proposed approach and show its e ectiveness compared with competing approaches.
Modern research is remarkably boosted by contemporary research-supporting
tools. Thanks to digital libraries, researchers support their work by accessing
a large part of the complete human knowledge with little e ort. However, the
sheer amount of rapidly published scienti c publications overwhelms researchers
with a large number of potentially relevant pieces of information. Recommender
systems have been introduced as an e ective tool in pointing researchers to
important publications [
The rest of this paper is organized as follows. In Section 2 we review work related to our approach. Section 3 presents the problem de nition and outlines the presented approach. Section 4 demonstrates the pro le building model employing the anti-aging factor. In Section 5 we explain the conducted experiments and discuss the results. Finally we conclude the paper in Section 6 2
Research paper recommendation has been a hot topic for more than a decade.
Several works addressed this problem proposing ideas from di erent
recommendation directions [
We propose a content-based research publications recommender (PubRec) that models both the active user (the researcher) and the candidate publications in terms of domain-related keywords. This section introduces the basic concepts of PubRec along with the formal problem de nition. 3.1
Digital libraries like ACM, IEEE, Springer, etc. publish meta-data about research publications publicly. Out of this meta-data, we are interested in title, abstract, keyword list and publication year. The rst three can be e ectively exploited to build a pro le for each publication p as a keyword vector, which represents p in terms of domain-related keywords: V!p = hwp;k1 ; wp;k2 ; :::; wpkn i, where ki is a domain-related keyword from the set of all keywords K, and wp;ki is the weight of ki in p with range of [0; 1].
All keywords from the keyword list are added to V!p with the maximum weight value of 1 by virtue of their source. As they are assigned to publications explicitly by the authors, we consider them the most precise domain description for the underlying publication. This list, however, contains usually up to 10 keywords, which is a small number for modeling a publication, thus, we aim to extend this list. Titles and abstracts hold a great essence of the ideas presented in publications. Therefore, we treat them as the second source of keywords and for each publication we apply keyword extraction from the concatenation of its title and abstract with weights correspond to the TF-IDF weighting scheme. 3.2
Given a researcher r with a set of her/his publications, we construct a researcher pro le V!r = hsr;k1 ; sr;k2 ; :::; sr;kn i such that ki 2 K is a domain-related keyword, and sr;ki is the importance of ki to r. Our proposed pro le construction method ensures that r's Interest Extent (IE) in a publication p is achieved by computing the dot product between the researcher's vector and the publication's vector: IE(V!r; V!p) = V!r Vp ! (1) 3.3
Our problem can be formally de ned as: Given a researcher r along with the corresponding set of publications Pr and a candidate set of publications Pcand, nd k publications from Pcand with the maximum IE. The presented approach can be summarized in the following steps: { First, we build the researcher pro le V!r using previous publications by modeling the problem as a multivariate linear regression problem (Section 4) { Each candidate publication p 2 Pcand is modeled as a keyword-vector V!p { We use Formula 1 to calculate IE(V!r; V!p) for candidate publication p 2 Pcand { Candidate publications are ordered by their Interest Extents and the top k are recommended to r.
We utilize researchers' publications to draw conclusions about their interests. A key aspect of PubRec consists in considering the di erent interest researchers have in their publications. After all, this interest might vary from paper to paper depending on several factors. Moreover, the importance of these factors vary among researchers. Thus, we believe that the publication age is an important factor in this regard since a ve years old publication, for example, might not re ect the author's current interest as much a publication of the current year. Based on that, we introduce a scoring function for estimating the a nity of a researcher r towards one of her publications p 2 Pr by engaging the publication's age, which is expressed by the number of years elapsed after the publication's date and represented by in the following function:
IEr;p = e ( )2 : (2) Here, is the researcher-speci c anti-aging factor. As depicted in Figure 1, the curve of IE is plotted for three di erent values of : 4, 20 and 50. There we can see how regulates the steepness of this curve. As the values of increase, the curve becomes less steep and results in higher IE values for older publications. For example consider researcher r0, the Interest Extent of r0 for p0, a 3 years old publication, can be modeled in three di erent ways upon three di erent values of : IEr0;p0 = 0:1 for = 4, IEr0;p0 = 0:63 for = 20 and IEr0;p0 = 0:83 for = 50. This behavior helps in modeling di erent types of researchers based on their a nity model. Such that, researchers who tend to stick to the same research topics longer time are modeled using larger values compared to other researchers who tend to change their topics of interest more rapidly. Choosing the best for each researcher is done empirically in this work, but further investigations about the correlation between researcher characteristics and the optimal value are left for future work.
E I
1
0:8
0:6
0:4
0:2
0
0
2
p0
p0
p0
4
=4
=20
=50
6
8
10
The second contribution in this work is to model the problem of measuring
the importance of domain related keywords for a researcher r as a multivariate
linear regression problem as follows: Given the set of r's publications Pr, for
each publication pi 2 Pr we build the underlying publication pro le as described
in section 3.1: V!pi = hwpi;k1 ; wpi;k2 ; :::; wpi;kn i. Furthermore, the Interest Extent
IEr;pi is calculated using Formula 2 as shown in Figure 2. Let the set of keywords'
weights of the paper pi: wpi;k1 ; wpi;k2 ; :::; wpi;kn be the set of predictors related
to the response variable IEr;pi , then the multivariate linear regression model [
V!pi = 0 + 1wk1 + ::: + nwkn .
k1 k2 k3 kn IE p1 99K w1;1 w1;2 w1;3 . . . w1;n IEr;p1 p2 99K w2;1 w2;2 w2;3 . . . w2;n IEr;p1
. . . pm 99K wm;1 wm;2 wm;3 . . . wm;n IEr;pm 99K 1 2 3
. . . n Where ! is the regression coe cient vector and 0; 1; : : : ; n are the regression coe cients. Each coe cient value j ; j 2 1; : : : n de nes the relation between the researcher r and the keyword kj , or in other words the importance of kj for r. Consequently, the user pro le is modeled by means of !. Meaning, that in order to nd the user pro le V!r, we should solve the previously mentioned regression problem and nd the vector !. This problem is solved by minimizing the cost function:
J ( ) =
This is a well known optimization problem and there exist a couple of algorithms
such as gradient descent or Normal equation to solve it [
We conducted experiments to validate our approach and compared it against some state-of-the-art approaches. In the following we describe the used dataset along with the used evaluation metrics. Finally, we show and discuss the results. 5.1
To evaluate the presented approach, we used the Scholarly publication Recommendation dataset1. It covers information about 50 anonymous researchers, enclosing their publication set, in addition to a set of publications of interest for each researcher. The interest lists are subsets of a larger collection of 100,531 publications called the candidate publications which is also provided. 1 https://www.comp.nus.edu.sg/~sugiyama/dataset2.html To the best of our knowledge, this is the only available dataset which provides the interest list for such a number of researchers. However, we had to resolve a major obstacle before we could use the dataset. That is, publications in the dataset are named by unique IDs without titles or author names, hence they cannot be identi ed and no meta-data was provided.
In order to make the dataset usable for our evaluation, we needed to identify the publications to be able to retrieve their meta-data. This was achieved by the following steps: (a) requesting and obtaining original pdf les from the dataset authors; (b) extracting publications' titles from the pdf les and using them to nd publication identities within the DBLP2 register; and nally (c) having the electronic edition pointer (ee) from DBLP publication's attributes, we retrieved needed information from corresponding publisher web site3. The result is a rich dataset that contains meta-data for 69,762 candidate publications, and more importantly the full publications and interest sets for 49 researchers. Lastly, for all publications in this dataset we applied the keyword extraction. Keywords extraction and weighting. We use Topia's Term Extractor 4 because of its e ciency and usability. It is a tool that uses Parts-Of-Speech (POS) and statistical analysis to determine the terms and their strength in a given text. Yet we extended this tool in order to extract keywords with higher quality and make the best use out of the limited available resources. Our extensions to Topia are: (a) we apply post ltering on the resulting terms by choosing only those terms which appear in a white list of computer science terms; (b) the weights of extracted terms is calculated based on the normalized TF-IDF weighting scheme.
We report the quality of our method with two important and widely adopted
metrics for evaluating ranking algorithms in information retrieval. For the
following metrics r is a researcher from the set of researchers R:
Mean Reciprocal Rank (MRR). MRR measures the method's quality by
checking the rst correct answer's position in the ranked result. For each
researcher r, let pr be the position of the rst interesting publication from the
recommended list, then MRR is calculated as M RR = jR1 j Pr2R p1r
Normalized Discounted Cumulative Gain (nDCG)[
Using the previously described dataset and evaluation metrics, we conducted
quality evaluations for our method with the following setup: given a set of
candidate publications, and a set of researchers with their publications set, the
system should correctly predict the interesting publications for each researcher.
The results are demonstrated in the rst row of Table 1. It shows that PubRec
manages to achieve a high MRR score of 0.717. Looking deeper into the details
of this metric by examining results for individual researchers gives more insights:
for 29 out of 49 researchers the rst relevant publication appeared at the rst
position of the recommended list, and at the second position for 7 researchers.
We compared our approach with two state-of-the-art publication recommender
systems [
We have proposed a novel approach on recommending scienti c publications. By
exploiting only publicly available meta-data from digital libraries the quality of
the predictions is superior to state-of-the-art approaches, which require access
to the full text of the paper. The focus is primarily on the user pro ling, where
a strategy to determine the trend of interests of a user in her own publications
over time is integrated into a multivariate linear regression problem. The e cacy
5 This method applies a light-weight variation of the Rocchio algorithm [
Acknowledgments. Work by Anas Alzoghbi was partially supported by the German Federal Ministry of Economics and Technology (BMWi) (KF2067905BZ). We thank Kazunari Sugiyama for his e orts in providing us with the complete Scholarly dataset. We also thank ACM for providing meta-data for our dataset.