ArticlePDF Available

Abstract and Figures

This paper presents a hybrid document recommender system intended for use in digital libraries and institutional repositories that are part of the Slovenian Open Access Infrastructure. The recommender system provides recommendations of similar documents across different digital libraries and institutional repositories with the aim to connect researchers and improve collaboration efforts. The hybrid recommender system makes use of document processing techniques, document metadata, and the similarity ranking function BM25 to provide content-based recommendations as a primary method. It also uses collaborative-filtering methods as a secondary method in a cascade hybrid recommendation technique. We also provide a real-world data feedback collection analysis for our hybrid recommender system on an academic digital repository in order to be able to identify suitable time-frames for direct feedback collection during the year.
Content may be subject to copyright.
Document Recommendations and Feedback
Collection Analysis within the Slovenian
Open-Access Infrastructure
Mladen Boroviˇc * , Marko Ferme, Janez Brezovnik, Sandi Majninger, Klemen Kac and
Milan Ojsteršek
Faculty of Electrical Engineering and Computer Science, University of Maribor, 2000 Maribor, Slovenia; (M.F.); (J.B.); (S.M.); (K.K.); (M.O.)
*Correspondence:; Tel.: +386-2-220-74-60
Received: 24 September 2020; Accepted: 21 October 2020; Published: 23 October 2020
This paper presents a hybrid document recommender system intended for use in digital
libraries and institutional repositories that are part of the Slovenian Open Access Infrastructure.
The recommender system provides recommendations of similar documents across different
digital libraries and institutional repositories with the aim to connect researchers and improve
collaboration efforts. The hybrid recommender system makes use of document processing
techniques, document metadata, and the similarity ranking function BM25 to provide content-based
recommendations as a primary method. It also uses collaborative-filtering methods as a secondary
method in a cascade hybrid recommendation technique. We also provide a real-world data feedback
collection analysis for our hybrid recommender system on an academic digital repository in order to
be able to identify suitable time-frames for direct feedback collection during the year.
hybrid recommender systems; feedback collection; digital libraries; information retrieval;
real-world data; open-access
1. Introduction
Recommender systems are a part of everyday experience on the web, especially while using
online stores and search engines. The main objective of these systems is to provide the user with
relevant and interesting content. In digital repositories, the obvious task for a recommender system is
to provide recommendations to relevant documents. Digital repositories are usually used by students,
researchers, and other interested parties, with an objective to research a certain topic and broaden
their knowledge in that domain. A recommender system can be very helpful in achieving that, since it
helps discover relevant documents, while the user does not need to browse and review a large amount
of documents.
Recommender systems in academic digital repositories are becoming prominent as the number of
produced academic documents in electronic format grows. There are many types of documents present
in academic digital repositories, including, but not limited to, undergraduate theses, postgraduate
theses (master’s theses and doctoral theses), journal articles, conference articles, workbooks, study
books, manuals, collections of problems, course slides, and other teaching and research materials.
In Slovenia, universities, colleges, other higher education institutions, and research institutions have
joined efforts to form the Slovenian Open-Access Infrastructure where documents from all partners
would be publicly available. Naturally, this also provides a framework for recommender systems as it
is possible to recommend documents between different institutions. Another positive side effect of this
Information 2020,11, 497; doi:10.3390/info11110497
Information 2020,11, 497 2 of 14
is that researchers from different institutions that are in the same field of expertise can see the work of
their colleagues more transparently, encouraging cooperation between them. With this goal in mind,
a recommender system for the Slovenian Open-Access Infrastructure was designed as a part of the
infrastructure to support the goals of the nationwide project. The novelty of this recommender system
is that it is currently the only recommender system in Slovenia that includes all Slovenian universities
and their electronic publications. In practice, over 200,000 electronic publications originating from any
of the Slovenian universities can be recommended using our system.
This paper presents a cascade type hybrid recommender system which is implemented in the
Slovenian Open-Access Infrastructure with the aim to serve relevant document recommendations
across all digital libraries and institutional repositories which are currently included in the
infrastructure. The second section briefly reviews related work. The third section presents the
current state of the Slovenian Open Access Infrastructure. The inner workings and the architecture of
our recommender system are presented in the fourth section. In the fifth section, we give details on
the feedback collection analysis for our implemented hybrid recommender system using the digital
repositories established within the Slovenian Open Access Infrastructure. The sixth section contains
conclusions and ideas for further work.
2. Related Work
Document recommender systems can be applied in many practical scenarios. Specifically, for the
scenario of document recommendations where the documents are news, Reference [
] demonstrates
the use of recommendations for job postings, in Reference [
], cloud computing was used for
recommendations and Reference [
] demonstrates a semantic web approach to recommending news.
Many document recommender systems have been extensively covered by the research field especially
for use with news. References [
] provide a survey of news recommendation systems. In [
fuzzy logic is used to recommend news using content-based methods. Rich feedback is used to
recommend news to users in [
], while Reference [
] compares information retrieval algorithms in
news recommendation scenarios. In some cases, semantic approaches such as Wordnet are used to aid
in semantic recommendations [9,10].
Research paper recommender systems are also prominent when it comes to document
recommendations [
]. A tag-based research paper recommender system framework is presented
in [
], and a similar tag-based approach was used in [
]. A collaborative filtering approach using
contexts was used to recommend research papers in [
]. An extensive comparison of offline
and online evaluation approaches of research paper recommender systems is presented in [
Specifically for digital repositories, several recommender systems have been developed. In [
keyphrases were used as a basis for research paper recommendations and, in [
], a social bookmarking
service CiteULike was used for recommendations. A recommender system specifically tailored for
advising research publications as a part of digital libraries in a university environment was presented
in [
]. Another study [
] introduces a Recommendation-as-a-Service (RaaS) platform used for
recommendations in academia and its integration into the reference manager JabRef [
]. Similarly,
CORE Recommender [
] was developed specifically for use in digital libraries and repositories.
As shown in [
], such recommender systems have also been implemented in academic social networks,
namely Mendeley.
When faced with researching, implementing, and maintaining recommender systems, challenges
do occur. Some major challenges were outlined in [
]. These include data quality, the lack
of appropriate data sets, choice of appropriate recommendation techniques, evaluation of
recommendations, and even the number of recommended items. In addition to these challenges,
we also encountered challenges while processing documents in the Slovenian language. Being a
morphologically rich language, it is required to take different approaches to natural language
processing when processing documents in Slovenian. Very little research has been done in
recommending documents in the Slovenian language, mostly because there was very few structured
Information 2020,11, 497 3 of 14
datasets of documents in Slovenian. With the introduction of the Slovenian Open Access
Infrastructure [
], this has improved greatly due to the creation of a large structured dataset,
containing over 200,000 documents [
]. It features segmented metadata consisting of titles, abstracts,
keywords as well as full-texts and other document metadata. From it, other datasets of the Slovenian
language have formed [
], which allows for further research options not only in the research of
recommender systems, but also other tasks in information retrieval and natural language processing,
specific to the Slovenian language.
3. Overview of the Slovenian Open Access Infrastructure
The Slovenian Open Access Infrastructure was established in 2013 and has since enabled the
interested parties in Slovenia (researchers, students, companies, and the public) access to the intellectual
production of Slovenian educational and research organizations. Simultaneously, it has enabled the
researchers to fulfill the requirements for open access to publications from publicly financed research.
Structurally (Figure 1), the infrastructure consists of a national portal [
], institutional
repositories for each of the four Slovenian universities (Digital Library of University of Maribor
(DLUM) [
], Repository of University of Ljuljana (RUL) [
], Repository of University of Primorska
(RUP) [
], Repository of University of Nova Gorica (RUNG) [
]), a repository for research institutions
(Digital Repository of Slovenian Research Organizations (DiRROS) [
]), and a repository for colleges
and higher education institutions (ReVIS) [34]).
- Federated search
- Similar content detection
- Recommender system
National open access
(Open Science Slovenia)
Digital Library of
University of Maribor
Repository of University of
Repository of University of
Nova Gorica
Digital Repository of
Slovenian Research
Repository of Colleges and
Higher Education Institutions
Other digital archives
Repository of University of
Figure 1. Structure of the Slovenian Open-Access Infrastructure.
The infrastructure also aggregates metadata from other digital archives such as [
], Social Science Data Archives [
], Digital Library of Slovenia [
], NUK Web
Archive [
], and the Ministry of Defense Library and Information System [
]. The types of
publications that are stored in the infrastructure include diploma, master’s and doctoral theses,
journal and conference articles, proceedings, datasets, scientific and technical reports, books, lecture
materials, and videos of lectures. Since a great majority of publications are in Slovenian, a side product
of this infrastructure was a large-scale corpus of full-text documents in the Slovenian language,
covering several different domains of research. It also spawned some research datasets for use in
linguistic studies [
]. More importantly, it currently represents the largest corpus of segmented
texts in the Slovenian language, giving several options for research not only in linguistics but also
in natural language processing. Due to interests for cooperation between the four universities and
several research institutions in Slovenia, a recommender system was integrated in the infrastructure.
The aim was to notify users about similar studies being done at different institutions through digital
libraries and institutional repositories.
Information 2020,11, 497 4 of 14
4. Document Recommendations
There are a few different approaches to recommendation in existence. The most common
approaches are content-based and collaborative filtering [
]. Other approaches include
demographic, utility-based, and knowledge-based techniques to recommendation. There is no optimal
approach for every situation. Each approach has advantages and disadvantages in certain scenarios.
While content-based filtering works well when a good description of an object is provided and
when starting out with recommendations, collaborative filtering tends to provide more contextually
appropriate recommendations once enough user feedback is provided. Hybrid systems aim to resolve
the disadvantages of both approaches by combining them in different ways [
]. Several hybridization
methods exist [
]. Weighted hybrids compute a score for a recommended item using outputs of all
recommendation approaches available in the system. Switching hybrids employ a mechanism to switch
between recommendation approaches. In this type of hybrid, approaches in the system are usually
given priorities. If an approach with a higher priority cannot give a sufficient score, the recommender
system switches to an approach with a lower priority as an attempt to provide a more recommendation
with a more sufficient score. Mixed hybrids provide recommendations from different approaches at
the same time. In cascade hybrids, one approach is used first to produce an initial set of recommended
items; then, a second approach is used to fine-pick the most suitable items from that initial set, in order
to provide a final recommendation.
Our recommender system is a cascade hybrid, incorporating content-based filtering as a primary
recommendation technique and collaborative filtering as a secondary re-ranking method. It consists
of three fundamental modules (Figure 2). The user activity log module provides the information
on user activities such as view count, download count, document ratings, and document referrals.
The document processing module ensures a unified feature representation of all documents in a triplet
representation consisting of a title, keywords, and an abstract. Simultaneously, this module performs
the calculation of BM25 values for each document pair, which forms a document index. The latter is
a similarity matrix for all documents. Documents are periodically processed as new documents are
added to the system daily. This way, the index is kept updated and the recommendations include
new documents.
Document processing
User activity log module
Document ranking module
<Title, Keywords, Abstract>
Document similarity index
BM25 calculation
User views
User downloads
User ratings
Document referrals
CBF recommendations
CF recommendations
Hybridization point
Final recommendations
Recommended documents
Figure 2. The architecture of our hybrid recommender system.
Information 2020,11, 497 5 of 14
The user activity data and the calculated similarities between the documents are the input to
the document ranking module, where similar documents are chosen depending on the document
that is viewed by the user. This is also the hybridization point, where content-based filtering and
collaborative filtering methods are applied in cascade to output the final list of recommendations,
which is served to the end-user.
4.1. Processing Documents in Slovenian
A variety of different metadata were obtained from previous established repositories.
These included information about authors, titles, keywords, abstracts, publishing year, and other
bibliographic information. The metadata standards were different and included COMARC, MARC 21,
and Dublin Core Metadata. We merged the different metadata schemes in our own metadata scheme to
enable collection of as much metadata as possible. Our own metadata scheme consists of all metadata
fields from the established standards with some extra fields for internal use. We use our metadata
scheme to represent documents and use it with the recommender system as well as some other services
within the Slovenian Open-Access Infrastructure.
For the recommender system, the documents are represented by titles, keywords, and abstracts.
Most documents are in the Slovenian language; however, there are also documents in English, German,
Italian, Croatian, and Hungarian. The documents that are not written in Slovenian have at least the
abstract and keywords translated to Slovenian to conform with the publication and cataloguing rules.
In the case of these documents, the available metadata in Slovenian are used with higher priority
than the metadata in other languages. First, the most common words in the Slovenian language
are removed from the text, since they do not contribute to semantic information. These are mainly
conjunctions, prepositions, particles, and interjections; however, common verbs and nouns are also
included. The common word list was built using word counts in documents. This is a periodic task,
which is run each time after a recommendation index is updated. Additionally, we used lemmatization
to help when dealing with conjugations and declensions in the text. Lemmatization is the process
of determining the basic lexical form (i.e., lemma) to the words in a text. A very similar process
to lemmatization is stemming. The main difference between lemmatization and stemming is that
stemming does not convert the word into its dictionary form but simply cuts off the ending of the
word. In text mining, lemmatization can be used to detect contexts of texts. It is used in our text
processing step to group semantically similar words and to avoid the difficult process of grouping with
declension and conjugation rules. Furthermore,
-grams for
N= [
1, 2, 3, 4, 5
are generated and used
with the
t f
id f
based ranking function BM25 to perform content-based filtering within our hybrid
approach to recommendation.
4.2. Document Ranking
For document ranking, we used the BM25 ranking function [
] along with additional weights,
which were obtained from document metadata and user activities. BM25 is a ranking function,
which enables the ranking of documents by the similarity of terms that are contained within those
documents. It is a family of functions, which differs by weighting schemes and parameter values.
In general,
t f
id f
weights are used [
]. The term frequency (
t f
) is the occurrence count of a term
within a document
while the inverse document frequency (
id f
) is the importance of the term
in the
given document collection
). Composite nonlinear
t f
normalizations and the family of
BM25 ranking functions have been used extensively in search engines to rank documents:
id f (t) = log ||D|| − n(t) + 0.5
n(t) + 0.5 (1)
s(d,Q) =
id f (qi)·t f (qi,d)·(k1+1)
t f (qi,d) + k1·B,qiQ,dD(2)
Information 2020,11, 497 6 of 14
avgdl (3)
It is a state-of-the-art
t f
id f
based ranking function and has spawned many variants including
BM25L, BM25+, BM25-adpt and BM25T [
], which bring improvements on very specific datasets.
It has also been implemented in open source and commercial solutions such as Apache Lucene,
Apache Solr, and Xapian as well as in Microsoft SQL Server and MySQL database implementations
as a default full-text search solution. We decided to implement BM25 ourselves on a Microsoft SQL
Server platform to have research options while studying parameters of the original ranking function
and its variants, since commercial solutions do not allow enough customization. Another reason for
this is that our documents are in the Slovenian language, for which only limited support exists in these
open source and commercial solutions.
in Equation
is the length of the collection
is the number of documents which
contain the term
. The BM25 value
depends on the weights
t f
id f
as well as parameters
. A general BM25 calculation for a document
and a query
with terms
is given with
, where
is the size of the query
given with the number of terms and
is a
normalization factor (Equation
). In Equation
is the length of document
is the
average length of the document in the corpus D.
The parameter
regulates the importance of the
t f
weight and the parameter
regulates the
importance of document length. The values for these two parameters can be set using advanced
optimization approaches, but usually values
1.2, 2.0
0.75 are used [
]. Currently,
we use empirically determined fixed values
1.2 and
0.75, but further study of the corpus
properties and parameter effects is underway. An automated adaptive technique of choosing the
parameters using an optimization method such as in [
] is desired. Additionally, we are also working
on including alternative weighting schemes such as t f *pd f [52] and t f -iduf[53].
4.3. Hybrid Approach to Recommendation
The input to our content-based filtering approach is a collection of metadata which describes
the documents. A document feature is represented with a vector of terms obtained from titles,
keywords, and abstracts. As we also have full-texts available, we empirically found that it is better
to use semantically dense metadata rather than full-text due to two important disadvantages. Firstly,
full-texts contain more terms which slows down the process of ranking similar documents. Secondly,
semantically important contexts diminish even after applying pre-processing with stop-word lists
t f
id f
filtering. However, when compared to a simpler document feature assembled from titles,
keywords and abstracts do not significantly improve recommendation results. We further enrich the
document feature with metadata including document typology [
], issue year, authors, repository ID,
and document language.
With all the metadata considered, we calculate a BM25 score based on the enriched document
features. We also use the Jaro–Winkler distance [
], in order to define a document typology
similarity. The Jaro–Winkler similarity is suitable when dealing with short strings and when the
similarity between them should be greater if the two strings match from the beginning. First, the Jaro
similarity is calculated by including the number of matching characters m and half the number of
transpositions t between strings
and their respective lengths
). Then,
the Jaro–Winkler similarity is calculated by including the common prefix length
and a scaling factor
0.1 to adjust the value depending on the common prefix length (Equation
). In our situation,
the document typologies are denoted with a short string of up to five characters (e.g.,
5). The first
character of the typology defines the kind of document and the following characters define the variant
of the document. Some examples of document types are provided in Table 1.
Information 2020,11, 497 7 of 14
Table 1.
Examples of document typologies and their metadata notation. Full typology is available
in [54].
Document Typology (Notation) Document Typology (Meaning)
1.01 Original scientific article
1.02 Review article
1.03 Short scientific article
1.04 Professional article
2.08 Doctoral dissertation
2.09 Master’s thesis
2.11 Undergraduate thesis
2.23 Patent application
2.24 Patent
2.25 Other monographs and completed works
Using the Jaro-Winkler distance (Equation
), we compare the typologies of two documents in
order to rank the documents with the similar typology higher. The final content-based filtering score
) is calculated as a product between the BM25 score on the document feature vector and
the Jaro–Winkler similarity on the document typology:
simj(s1,s2) =
0 if m=0
motherwise (4)
simjw(s1,s2) = simj(s1,s2) + λp(1simj(s1,s2)) (5)
djw(s1,s2) = 1simjw(s1,s2)(6)
ScoreCBF =B M25(dA,dB)·djw (tdA,tdB)(7)
Our collaborative filtering approach is collaborative in the sense that we use user interactions to
re-rank the content-based filtering recommendations with the goal of improving recommendations.
The input to our collaborative filtering approach is the user activity data regarding a document
Views and download counts for documents are kept and regularly updated. The values for actions were
set to 1 if a view occurs and 10 if a download occurs, meaning that a download action is as significant
as 10 view actions (Equation
). A feedback value
is calculated by summing all values of actions.
Furthermore, we also store a similar feedback value for actions
on recommended documents
give higher weight to the documents which were interesting to end-users (Equation
). The values for
boosts were set to 5 if a view on a recommended document occurs and 50 if a recommended document
is downloaded. Action significance values for
were set empirically, with an idea in mind
that a download is worth 10 times as significant as a view, and a recommended view is five times as
significant as a regular view.
ad,iad,i=(1 document view
10 document download (8)
rd,ird,i=(5 document view
50 document download (9)
Information 2020,11, 497 8 of 14
We can provide adaptive recommendations using actions from users by combining feedback
values for actions and recommendations with the download rate
), which is the
ratio between downloads and views of a document. The logic is the same for the download rate of
the recommended documents
, but only views and downloads on the recommended document
are considered. The feedback value for actions on recommended documents makes the clicked
recommendations rank higher in the recommendation list. The final collaborative filtering score
) is calculated as a product of the document download rate and the sum of action
feedback values on the document and actions on recommendations:
ScoreCF =fad·hd+frd·hr(11)
With both approaches combined into a hybrid approach, we use recommendation strategies,
which can be customized depending on the type or purpose of recommendations. Some
recommendation strategies that we used in production are »latest + relevant«, »same repository
+ relevant« and »more from same authors«. These strategies can also be merged into a single strategy
using priority factors. For example, a strategy »latest from same repository and from same authors«
would first pick the latest documents and would then filter them according to their repository primarily
and according to their authors secondarily.:
The workflow of our hybrid recommender system consists of four steps (Figure 3). First, the results
from our content-based approach are obtained. Second, an exponential temporal decay mechanic
) is implemented to increase the ranks of recently published documents. The parameter
controls the exponential temporal decay. The similarity score of the document is multiplied by the
temporal decay and the recommendations in the results are re-ranked. Documents contained in the
result set are then input into our collaborative filtering approach which re-ranks the results again.
Currently, the output result length of our content-based approach is 25 documents. Finally, the list of
recommendations is shortened to N documents for better presentation of the result on the web. In
practice, we shorten the list to five documents.
Figure 3. The workflow of our hybrid cascade approach to recommendations.
5. Feedback Collection Analysis
Collecting feedback from users is an important part of recommender systems design because
it can directly influence the resulting recommendations. The overall user experience with regard to
recommendations can be greatly improved if feedback is regularly collected from users. This can
be done directly using surveys, questionnaires, and quick questions or indirectly by analyzing user
activity. To achieve sufficient feedback, an appropriate time for feedback collection must be determined.
The quality of feedback depends on the mood of the user, but, with careful planning, there is more
Information 2020,11, 497 9 of 14
chance that the user will be willing to give good quality feedback. Another perspective is to collect
feedback at a certain time, where we are sure that users might be more inclined to express their
opinions (e.g., a week after something changed) as they have had enough time to form an opinion.
Furthermore, a good feedback collection approach can lead to an organized approach to evaluation of
recommender systems. With it, evaluation metrics can be better defined and used to measure the true
performance of the recommender system.
We performed an analysis of time-frames during the year, when feedback collection would make
sense within the Slovenian Open-Access Infrastructure. In our case, the recommendations are focused
on documents and are meant to help students, academic staff, and researchers find more similar
documents to their interest. The recommendations are therefore accessed as the users are using the
recommender system, which is linked to different time-frames during the year. We found that several
spikes in usage occur during the year and we tried to link them to specific events that occur in the
academic year (e.g., thesis defenses, summer vacations, etc.).
We limited our data to data from four universities in Slovenia and their institutional repositories in
the Slovenian Open-Access Infrastructure. University of Maribor was included with DLUM, University
of Ljubljana with RUL, University of Primorska with RUP and University of Nova Gorica with RUNG.
All institutional repositories store view and download counts for documents. During this analysis,
we treated viewed documents as mildly interesting and downloaded documents as very interesting.
We did this because a download can occur only after the document is viewed; therefore, if a user
downloaded the document, they must have viewed its detailed description with metadata and made a
conscious decision that it is interesting enough for them to download it.
We encountered a major limitation with the accessibility of the traffic data on each institutional
repository. DLUM was the only repository that we were able to get the data from, since other
repositories opted not to be included in the analysis by their maintenance teams. Furthermore,
the maintenance teams of DLUM, RUL, RUP, and RUNG decided to exclude all traffic tracking options
on repositories after 2016. As for DLUM data that we were able to obtain, it was Google Analytics
traffic data between January 2013 and December 2016. With all limitations considered, we performed
an analysis using data only from DLUM (Figure 4). It proved to be a suitable institutional repository
for this task, since it is the first university institutional repository in Slovenia, running since 2008 and
serving as a basis for all other institutional repositories in the national open-access infrastructure.
Figure 4. Weekly user visits to DLUM between January 2013 and December 2016.
In the data set time-frame of user activity between January 2013 and December 2016 (Figure 4),
special events have occurred. In November 2014, DLUM saw a major update and was offline for two
Information 2020,11, 497 10 of 14
weeks (weeks 48 to 50) due to this. It was updated at this time because it had to run stable for most
of the year, due to a regular influx of new theses. This influx annually reaches a peak in September
and October (weeks 40 to 42), when the theses are catalogued by the librarians. It was decided to run
DLUM without interruption between March and November 2014 because most users during that time
are students researching for their theses and researchers searching for related work for their articles.
An increase in weekly user visits can be observed in 2015. This increase seems to be attributed
to the marketing efforts of the Slovenian Open-Access Infrastructure and the cross-repository
recommendations; however, this cannot be confirmed due to the lack of traffic tracking capabilities on
repositories RUL, RUP, and RUNG.
Furthermore, in 2016, we can observe another increase in weekly user visits, which lasts from
January (week 1) to September (week 40). This unusual additional traffic was generated by students
enrolled in pre-Bologna process study programs at the University of Maribor. These students had
to complete and defend their theses by October 2016 as directed by the University of Maribor and
were most likely collecting research on DLUM in order to achieve this. This reason holds, as the traffic
increase stops in September 2016 (week 40).
By observing traffic fluctuation during the year, we found a decrease in weeks that correspond to
holidays. This occurs in several time-frames which are visible in Figure 4and denoted with letters:
A—January; the first week of the year (consequence of New Year),
B—February; weeks 7 and 8, around February 8th (national holiday “Prešeren Day”),
C—April and May; week 18 and 19, starting around April 27th (national holiday “Day of uprising
against occupation”) and ending around May 1st (national holiday “International Workers’ Day”),
D—June, July and August; weeks 26 to 36, summer holiday season,
E—October, November; weeks 44 and 45, around October 31st (national holiday “Reformation
Day”) and November 1st (national holiday “All Saint’s Day”),
F—December; weeks 50 to 53, around December 25th (national holiday “Christmas”),
26th (national holiday “Independence and Unity Day”) and December 31st (national holiday
“New Year’s Eve”).
We conclude that these time-frames are suitable for maintenance work on institutional repositories.
Time-frames B, C, and E show the potential for smaller updates and minor changes, while time-frame
D shows the potential for large-scale maintenance.
We also observed the peak traffic occurring between some before mentioned time-frames:
X—weeks 9 and 17 (from February to April),
Y—weeks 20 to 25 (from May to June),
Z—weeks 37 to 43 (from August to October).
We conclude that these time-frames are suitable for feedback collection campaigns, surveys,
and questionnaires. Namely, time-frames X and Y are more suitable for active user feedback
collection (e.g., validation of recommended documents), since users are actively researching during
that time. Time-frame Z is more suitable for general feedback collection (e.g., general surveys regarding
user experience).
An extensive evaluation study of our recommender system is currently still underway as
it requires successful collaboration of several institutions that maintain their own repositories.
Several metrics for recommendation system evaluation exist. In general, there are two ways of
evaluating any recommendation system: online and offline [
]. Offline evaluation makes use of
preferably labelled data which is split into training and test sets. The recommendation system uses the
training set ratings to try and predict the ratings in the test set. Actual users are not needed in this type
of evaluation. This makes offline evaluation fast and easy to perform on a large amount of data. It can
also be performed using many different datasets and with multiple different algorithms. The main
disadvantage of this approach is that it cannot measure true user satisfaction.
Information 2020,11, 497 11 of 14
In an online evaluation scenario, users interact with a running recommendation system and
respond to it naturally, while feedback is being collected from them. Feedback is obtained by either
asking the users directly or observing their actions. This approach measures true user satisfaction but
can take a long time to set-up and run from beginning to end.
The choice of metrics differs depending on the approach of recommendation. Information retrieval
metrics such as accuracy, recall, precision, and F-measure are usually considered preferable when
evaluating content-based recommendation systems. Other metrics for this type of recommendation
system include normalized discounted cumulative gain [
], rank-biased precision [
], and expected
reciprocal rank [
]. Collaborative filtering recommendations are usually evaluated using approaches
that measure novelty, serendipity, diversity, and coverage [
]. Currently, there are several different
metrics [
] that can be used to evaluate recommendation systems. When dealing with hybrid
recommendation systems, this must be carefully considered, since the type of hybridization can also
affect the evaluation process, making it complex due to implementation in multiple stages.
6. Conclusions
In this article, we present a cascade hybrid recommender system implemented in institutional
repositories that is part of the Slovenian National Open-Access Infrastructure. We outlined the
recommender system architecture, document pre-processing, and ranking approaches. A feedback
collection analysis has been presented on real-world data from one of our longest running repositories.
With the analysis, we were able to identify different time-frames during the year where it is suitable
to consider feedback collection on an academic digital repository. An extensive evaluation study is
currently underway and we conclude that, for an extensive evaluation of our recommender system’s
contribution to knowledge exchange and spread across the Slovenian Open-Access Infrastructure,
a unified framework should be developed in addition to institutional repository management processes
regarding logging user activities and using traffic tracking scripts. Only with such an approach can a
definitive contribution of the recommender system be confirmed and further researched. It would also
allow the observation of any significant cooperation between institutions, as it is already suspected that
the institutions in the two largest institutional repositories in the national open-access infrastructure be
in accordance with the majority of research cooperation efforts in Slovenia.
Author Contributions:
Conceptualization, M.B., M.F., and M.O.; methodology, M.B. and M.F.; software, M.B.,
M.F., and J.B.; validation, M.B., M.F., J.B., S.M., K.K., and M.O.; writing—original draft preparation, M.B. and
M.O.; writing—review and editing, M.B., M.F., J.B., S.M., K.K., and M.O. All authors have read and agreed to the
published version of the manuscript.
Funding: This research received no external funding.
Conflicts of Interest: The authors declare no conflict of interest.
Elsafty, A.; Riedl, M.; Biemann, C. Document-based Recommender System for Job Postings using Dense
Representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association
for Computational Linguistics: Human Language Technologies, New Orleans, LA, USA, 1–6 June 2018;
Volume 3, pp. 216–224, doi:10.18653/v1/N18-3027.
del Campo, J.V.; Pegueroles, J.; Hernández-Serrano, J.; Soriano, M. DocCloud: A document
recommender system on cloud computing with plausible deniability. Inf. Sci.
,258, 387–402,
Cantador, I.; Bellogín, A.; Castells, P. News@hand: A Semantic Web Approach to Recommending News. In
International Conference on Adaptive Hypermedia and Adaptive Web-Based Systems, Proceedings of the
AH 2008: Adaptive Hypermedia and Adaptive Web-Based Systems, Hannover, Germany, 29 July–1 August
2008; Nejdl, W., Kay, J., Pu, P., Herder, E., Eds.; Springer: Berlin/Heidelberg, Germany, 2008; pp. 279–283.
Karimi, M.; Jannach, D.; Jugovac, M. News recommender systems—Survey and roads ahead.
Inf. Process. Manag. 2018,54, 1203–1227, doi:10.1016/j.ipm.2018.04.008.
Information 2020,11, 497 12 of 14
Borges, H.L.; Lorena, A.C., A Survey on Recommender Systems for News Data. In Smart Information and
Knowledge Management: Advances, Challenges, and Critical Issues; Szczerbicki, E., Nguyen, N.T., Eds.; Springer:
Berlin/Heidelberg, Germany, 2010; pp. 129–151, doi:10.1007/978-3-642-04584-4_6.
Adnan, M.N.M.; Chowdury, M.R.; Taz, I.; Ahmed, T.; Rahman, R.M. Content based news recommendation
system based on fuzzy logic. In Proceedings of the 2014 International Conference on Informatics, Electronics
Vision (ICIEV), Dhaka, Bangladesh, 23–24 May 2014; pp. 1–6.
Ardissono, L.; Petrone, G.; Vigliaturo, F. News Recommender Based on Rich Feedback. In International
Conference on User Modeling, Adaptation, and Personalization, Proceedings of the UMAP 2015: User Modeling,
Adaptation and Personalization, Dublin, Ireland, 29 June–3 July 2015; Ricci, F., Bontcheva, K., Conlan, O., Lawless,
S., Eds.; Springer International Publishing: Cham, Switzerland, 2015; pp. 331–336.
Bogers, T.; van den Bosch, A. Comparing and Evaluating Information Retrieval Algorithms for News
Recommendation. In Proceedings of the 2007 ACM Conference on Recommender Systems (RecSys ’07),
Minneapolis, MN, USA, 19–20 October 2007; pp. 141–144. doi:10.1145/1297231.1297256.
Capelle, M.; Hogenboom, F.; Hogenboom, A.; Frasincar, F. Semantic News Recommendation Using Wordnet
and Bing Similarities. In Proceedings of the 28th Annual ACM Symposium on Applied Computing (SAC ’13),
Coimbra, Portugal, 18–22 March 2013; Association for Computing Machinery: New York, NY, USA, 2013; pp.
296–302, doi:10.1145/2480362.2480426.
Capelle, M.; Moerland, M.; Hogenboom, F.; Frasincar, F.; Vandic, D. Bing-SF-IDF+: A Hybrid
Semantics-Driven News Recommender. In Proceedings of the 30th Annual ACM Symposium on Applied
Computing (SAC ’15), Salamanca, Spain, 13–17 April 2017; pp. 732–739, doi:10.1145/2695664.2695700.
Beel, J.; Gipp, B.; Langer, S.; Breitinger, C. Research-paper recommender systems: A literature survey. Int. J.
Digit. Libr. 2016,17, 305–338, doi:10.1007/s00799-015-0156-0.
Jomsri, P.; Sanguansintukul, S.; Choochaiwattana, W. A Framework for Tag-Based Research Paper
Recommender System: An IR Approach. In Proceedings of the 2010 IEEE 24th International Conference on
Advanced Information Networking and Applications Workshops, Perth, WA, Australia, 20–23 April 2010;
pp. 103–108.
Choochaiwattana, W. Usage of tagging for research paper recommendation. In Proceedings of the 2010
3rd International Conference on Advanced Computer Theory and Engineering(ICACTE), Chengdu, China,
20–22 August 2010; Volume 2, pp. V2–439–V2–442.
Winoto, P.; Tang, T.; McCalla, G. Contexts in a Paper Recommendation System with Collaborative Filtering.
Int. Rev. Res. Open Distance Learn. 2012,13, 56–75, doi:10.19173/irrodl.v13i5.1243.
Beel, J.; Langer, S. A Comparison of Offline Evaluations, Online Evaluations, and User Studies in the
Context of Research-Paper Recommender Systems. In International Conference on Theory and Practice of Digital
Libraries, Proceedings of the TPDL 2015: Research and Advanced Technology for Digital Libraries, Poznan, Poland,
14–18 September 2015; Kapidakis, S., Mazurek, C., Werla, M., Eds.; Springer International Publishing: Cham,
Switzerland, 2015; pp. 153–168.
Ferrara, F.; Pudota, N.; Tasso, C. A Keyphrase-Based Paper Recommender System. In Digital Libraries and
Archives; Agosti, M., Esposito, F., Meghini, C., Orio, N., Eds.; Springer: Berlin/Heidelberg, Germany, 2011;
pp. 14–25.
Bogers, T.; van den Bosch, A. Recommending Scientific Articles Using Citeulike. In Proceedings of the 2008
ACM Conference on Recommender Systems (RecSys ’08), Lausanne, Switzerland, 23–25 October 2008; pp.
287–290, doi:10.1145/1454008.1454053.
Porcel, C.; Moreno, J.; Herrera-Viedma, E. A multi-disciplinar recommender system to advice
research resources in University Digital Libraries. Expert Syst. Appl.
,36, 12520–12528,
Beel, J.; Aizawa, A.; Breitinger, C.; Gipp, B. Mr. DLib: Recommendations-as-a-Service (RaaS) for Academia.
In Proceedings of the 2017 ACM/IEEE Joint Conference on Digital Libraries (JCDLToronto, ON, Canada,
19–23 June 2017; pp. 1–2, doi:10.1109/JCDL.2017.7991606.
Feyer, S.; Siebert, S.; Gipp, B.; Aizawa, A.; Beel, J. Integration of the Scientific Recommender System Mr.
DLib into the Reference Manager JabRef. In European Conference on Information Retrieval, Proceedings of the
ECIR 2017: Advances in Information Retrieval, Aberdeen, UK, 8–13 April 2017; Springer: Cham, Switzerland,
2017, doi:10.1007/978-3-319-56608-5_80.
Information 2020,11, 497 13 of 14
Knoth, P.; Anastasiou, L.; Charalampous, A.; Cancellieri, M.; Pearce, S.; Pontika, N.; Bayer, V. Towards
effective research recommender systems for repositories. arXiv 2017, arXiv:1705.00578.
Vargas, S.; Hristakeva, M.; Jack, K. Mendeley: Recommendations for Researchers. In Proceedings of the
10th ACM Conference on Recommender Systems (RecSys ’16), Boston, MA, USA, 15–19 September 2016;
pp. 365–365.
Beel, J.; Dinesh, S. Real-World Recommender Systems for Academia: The Pain and Gain in Building,
Operating, and Researching them [Long Version]. arXiv 2017, arXiv:1704.00156.
Ojsteršek, M.; Brezovnik, J.; Kotar, M.; Ferme, M.; Hrovat, G.; Bregant, A.; Boroviˇc, M. Establishing
of a Slovenian open access infrastructure: A technical point of view, Program
,48, 394–412.
OpenScience Slovenia Dataset. Available online: (accessed on
23 October 2020).
Erjavec, T.; Fišer, D.; Ljubeši´c, N.; Arhar Holdt, Š.; Bren, U.; Robnik Šikonja, M.; Udoviˇc, B. Terminology
Identification Dataset KAS-Term 1.0. Available online:
11356/1198 (accessed on 23 October 2020).
Erjavec, T.; Fišer, D.; Ljubeši´c, N.; Bitenc, M. Bilingual Terminology Extraction Dataset KAS-Biterm 1.0;
Available online: (accessed on 23 October
28. OpenScience Slovenia. Available online: (accessed on 23 October 2020).
Digital Library of University of Maribor-DLUM. Available online:
(accessed on 23 October 2020).
Repository of the University of Ljubljana-RUL. Available online:
php/eng (accessed on 23 October 2020).
Repository of the University of Primorska-RUP. Available online: https://repozitorij.upr/info/index.php/
eng (accessed on 23 October 2020).
Repository of the University of Nova Gorica-RUNG. Available online:
php/eng (accessed on 23 October 2020).
Digital repository of Slovenian research organizations. Available online:
index.php/eng (accessed on 23 October 2020).
Repository of colleges and higher education institutions-ReVIS. Available online: https://revis.openscience.
si/info/index.php/eng (accessed on 23 October 2020).
35. Available online: (accessed on 23 October 2020).
Social Science Data Archives. Available online: (accessed on on 23
October 2020).
Digital Library of Slovenia. Available online: (accessed on on 23 October
38. NUK Web Archive. Available online: (accessed on on 23 October 2020).
Ministry of Defence Library and Information System. Available online:
en (accessed on on 23 October 2020).
Jakubíˇcek, M.; Fiser, D.; Suchomel, V. Terminology Extraction for Academic Slovene Using Sketch Engine.
In Proceedings of the Tenth Workshop on Recent Advances in Slavonic Natural Language Processing,
Karlova Studanka, Czech Republic, 2–4 December 2016; Volume 10.
Ljubeši´c, N.; Fiser, D.; Erjavec, T. KAS-term: Extracting Slovene Terms from Doctoral Theses via Supervised
Machine Learning. In International Conference on Text, Speech, and Dialogue, Proceedings of the TSD 2019: Text,
Speech, and Dialogue, Ljubljana, Slovenia, 11–13 September2019; Springer: Cham, Switzerland, 2019; pp. 115–126,
Adomavicius, G.; Tuzhilin, A. Toward the next generation of recommender systems: A survey of
the state-of-the-art and possible extensions. IEEE Trans. Knowl. Data Eng.
,17, 734–749,
Bobadilla, J.; Ortega, F.; Hernando, A.; Gutiérrez, A. Recommender systems survey. Knowl.-Based Syst.
46, 109–132, doi:10.1016/j.knosys.2013.03.012.
Burke, R. Hybrid Recommender Systems: Survey and Experiments. User Model.-User-Adapt. Interact.
12, 331–370, doi:10.1023/A:1021240730564.
Information 2020,11, 497 14 of 14
Burke, R., Hybrid Web Recommender Systems. In The Adaptive Web: Methods and Strategies of Web
Personalization; Brusilovsky, P., Kobsa, A., Nejdl, W., Eds.; Springer: Berlin/Heidelberg, Germany, 2007;
pp. 377–408, doi:10.1007/978-3-540-72079-9_12.
Robertson, S.; Zaragoza, H. The Probabilistic Relevance Framework: BM25 and beyond. Found. Trends Inf.
Retr. 2009,3, 333–389.
Jones, K.; Walker, S.; Robertson, S. A probabilistic model of information retrieval: Development and
comparative experiments: Part 2. Inf. Process. Manag.
,36, 809–840, doi:10.1016/S0306-4573(00)00016-9.
Géry, M.; Largeron, C. BM25t: A BM25 extension for focused information retrieval. Knowl. Inf. Syst.
32, 217–241, doi:10.1007/s10115-011-0426-0.
Trotman, A.; Puurula, A.; Burgess, B. Improvements to BM25 and Language Models Examined.
In Proceedings of the 2014 Australasian Document Computing Symposium (ADCS ’14), Melbourne, VIC,
Australia, 27–28 November 2014; ACM: New York, NY, USA, 2014; pp. 58–65, doi:10.1145/2682862.2682863.
Manning, Christopher D. and Raghavan, Prabhakar and Schütze, Hinrich. In Introduction to Information
Retrieval; Cambridge University Press: New York, NY, USA, 2008.
Bollegala, D.; Noman, N.; Iba, H. RankDE: Learning a Ranking Function for Information Retrieval Using
Differential Evolution. In Proceedings of the 13th Annual Conference on Genetic and Evolutionary
Computation (GECCO ’11), Dublin, Ireland, 12–16 July 2011; pp. 1771–1778, doi:10.1145/2001576.2001814.
Nguyen, K.; Byung-Joo Shin.; Seong Joon Yoo. Hot topic detection and technology trend tracking for patents
utilizing term frequency and proportional document frequency and semantic information. In Proceedings of
the 2016 International Conference on Big Data and Smart Computing (BigComp), Hong Kong, China, 18–20
January 2016; pp. 223–230.
Beel, J.; Langer, S.; Gipp, B. TF-IDuF: A Novel Term-Weighting Scheme for User Modeling based on Users’
Personal Document Collections. In Proceedings of the iConference 2017, Wuhan, China, 22–25 March, 2017,
doi: 10.9776/17217.
COBISS/IZUM, Typology of documents/works for bibliography management in COBISS, 2016. Available
online: (accessed on 23 October 2020).
Jaro, M.A. Advances in Record-Linkage Methodology as Applied to Matching the 1985 Census of Tampa,
Florida. J. Am. Stat. Assoc. 1989,84, 414–420, doi:10.1080/01621459.1989.10478785.
Winkler, W. String Comparator Metrics and Enhanced Decision Rules in the Fellegi-Sunter Model of Record
Linkage. Avaliable online: (accessed on 23 October 2020).
Hernández del Olmo, F.; Gaudioso, E. Evaluation of recommender systems: A new approach.
Expert Syst. Appl. 2008,35, 790–804, doi:10.1016/j.eswa.2007.07.047.
Silveira, T.; Zhang, M.; Lin, X.; Liu, Y.; Ma, S. How good your recommender system is? A survey on
evaluations in recommendation, Int. J. Mach. Learn. Cybern.
,10, 813–831, doi:10.1007/s13042-017-0762-9.
Wang, Y.; Wang, L.; Li, Y.; He, D.; Liu, T.Y. A Theoretical Analysis of NDCG Type Ranking Measures.
In Conference on Learning Theory; PMLR: Princeton, NJ, USA, 2013; Volume 30, pp. 25–54.
Moffat, A.; Zobel, J. Rank-Biased Precision for Measurement of Retrieval Effectiveness. ACM Trans. Inf. Syst.
2008,27, doi:10.1145/1416950.1416952.
Chapelle, O.; Metlzer, D.; Zhang, Y.; Grinspan, P. Expected Reciprocal Rank for Graded Relevance.
In Proceedings of the 18th ACM Conference on Information and Knowledge Management (CIKM ’09),
Hong Kong, China, 2–6 November 2009; Association for Computing Machinery: New York, NY, USA, 2009;
pp. 621–630, doi:10.1145/1645953.1646033.
Gunawardana, A.; Shani, G. A Survey of Accuracy Evaluation Metrics of Recommendation Tasks. J. Mach.
Learn. Res. 2009,10, 2935–2962.
Shani, G.; Gunawardana, A., Evaluating Recommendation Systems. In Recommender Systems Handbook;
Springer: Boston, MA, USA, 2011; pp. 257–297, doi:10.1007/978-0-387-85820-3_8.
Publisher’s Note:
MDPI stays neutral with regard to jurisdictional claims in published maps and institutional
2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (
... In the literature, recommender systems have been explored for different domains. Some examples include their use in e-learning settings (Marques et al. 2021;Amara and Subramanian 2020;Fernández-García et al. 2020;Chrysafiadi et al. 2018), entertainment websites (Kannikaklang et al. 2022;Gupta et al. 2020;Singla et al. 2020;Troussas et al. 2018), social environments (Yang et al. 2020;Wongkhamchan et al. 2019;Mughaid et al. 2019;Liang et al. 2019), digital repositories (Troussas et al. 2021;Borovič et al. 2020;Guan et al. 2019;Jomsri et al., 2018), tourism systems (Baker and Yuan 2021;Srisawatsakul and Boontarig 2020;Chen et al. 2020;Kbaier et al. 2017). The algorithmic techniques that have been primarily used in the research papers, described above, are collaborative filtering, model-based approaches, stereotypes, content-based filtering, machine learning and hybrid approaches. ...
Full-text available
Group formation is a complex task requiring computational support to succeed. In the literature, there has been considerable effort in the development of algorithms for composing groups as well as their evaluation. The most widely used approach is the Genetic Algorithm, as, it can handle numerous variables, generating optimal solutions according to the problem requirements. In this study, a novel genetic algorithm was developed for forming groups using innovative genetic operators, such as a modification of 1-point and 2-point crossover, the gene and the group crossover, to improve its performance and accuracy. Moreover, the proposed algorithm can be characterized as domain-independent, as it allows any input regardless of the domain problem; i.e., whether the groups concern objects, items or people, or whether the field of application is industry, education, healthcare, etc. The grouping genetic algorithm has been evaluated using a dataset from the literature in terms of its settings, showing that the tournament selection is better to be chosen when a quick solution is required, while the introduced gene and group crossover operators are superior to the classic ones. Furthermore, the combination of up to three crossover operators is ideal solution concerning algorithm’s accuracy and execution time. The effectiveness of the algorithm was tested in two grouping cases based on its acceptability. Both the students participated in forming collaborative groups and the professors participated in evaluating the groups of courses created were highly satisfied with the results. The contribution of this research is that it can help the stakeholders achieve an effective grouping using the presented genetic algorithm. In essence, they have the flexibility to execute the genetic algorithm in different contexts as many times as they want until to succeed the preferred output by choosing the number of operators for either greater accuracy or reduced execution time.
... Computer-aided process design can only apply artificial intelligence in the early stages of development, and it is limited to intelligent thinking at this time. In practice, there are no distinct varieties of artificial intelligence [16]. Instead, a variety of applications are combined to create a more complete form [17]. ...
Full-text available
Technological development has given a new dimension to the art and design area. This research implements the deep learning model for human-computer integration technology in art and design with wireless sensor networks. In this human-computer interaction, each person and the devices are treated as network nodes. The user will act as a node requesting access privileges for utilizing the calligraphy design to learn the strokes in the self-pace mode for studies. In this research, the human-computer interaction is implemented for recognizing the strokes and curves in the calligraphy images loaded by the user. This identification will aid the user in learning different styles of calligraphy. The proposed fruit-fly optimization algorithm (FFOA) method is analyzed for high classification accuracy, less delay time, and reduced noise parameters against specific existing algorithms.
... In the related scientific literature, there have been several research works that explore the development of recommender systems in different domains, e.g., digital repositories [6,[8][9][10][11][12], entertainment web portals [13][14][15][16], digital education [17][18][19][20], and social settings [21][22][23][24]. In the aforementioned research studies, the algorithmic approaches that have been mainly adopted are content-based filtering, stereotypes and model-based approaches, collaborative filtering, and hybrid techniques. ...
Full-text available
Digital repositories contain a large amount of content, which is available to heterogeneous groups of people. As such, in many cases people encounter difficulties in finding specific content which is related to their preferences. In view of this compelling need and towards advancing human-computer interaction, this paper presents a recommender system which is incorporated in a digital repository. The recommender system is designed using multiple-criteria decision analysis (MCDA) and more specifically the weighted sum model (WSM) in order to refine the delivered content to the users. It also considers several users’ characteristics (their preferences as depicted by the content they visited or searched and by the frequency of searches/visits) and features of the content (content types and traffic). The recommender system outputs the suggestions of content to users based on their preferences and interests. The presented recommender system was evaluated by real users, and the results show a high degree of accuracy in the recommended content and satisfaction by users.
... Sometimes, collaborative filtering methods are combined with content-based approaches to solve some problems of the former and obtain more reliable recommendations. This combination is used in a cascade hybrid proposal for document recommendation presented in this issue [2]. A content-based method that makes use of document processing techniques and document metadata is applied first to provide an initial list of recommendations. ...
Full-text available
The large amount of digital content available through web sites, social networks, streaming services, and other distribution media, allows more and more people to access virtually unlimited sources of information, products, and services [...]
Full-text available
Recommender Systems have become a very useful tool for a large variety of domains. Researchers have been attempting to improve their algorithms in order to issue better predictions to the users. However, one of the current challenges in the area refers to how to properly evaluate the predictions generated by a recommender system. In the extent of offline evaluations, some traditional concepts of evaluation have been explored, such as accuracy, Root Mean Square Error and P@N for top-k recommendations. In recent years, more research have proposed some new concepts such as novelty, diversity and serendipity. These concepts have been addressed with the goal to satisfy the users’ requirements. Numerous definitions and metrics have been proposed in previous work. On the absence of a specific summarization on evaluations of recommendation combining traditional metrics and recent progresses, this paper surveys and organizes the main research that present definitions about concepts and propose metrics or strategies to evaluate recommendations. In addition, this survey also settles the relationship between the concepts, categorizes them according to their objectives and suggests potential future topics on user satisfaction.
Full-text available
In this paper, we argue why and how the integration of recommender systems for research can enhance the functionality and user experience in repositories. We present the latest technical innovations in the CORE Recommender, which provides research article recommendations across the global network of repositories and journals. The CORE Recommender has been recently redeveloped and released into production in the CORE system and has also been deployed in several third-party repositories. We explain the design choices of this unique system and the evaluation processes we have in place to continue raising the quality of the provided recommendations. By drawing on our experience, we discuss the main challenges in offering a state-of-the-art recommender solution for repositories. We highlight two of the key limitations of the current repository infrastructure with respect to developing research recommender systems: 1) the lack of a standardised protocol and capabilities for exposing anonymised user-interaction logs, which represent critically important input data for recommender systems based on collaborative filtering and 2) the lack of a voluntary global sign-on capability in repositories, which would enable the creation of personalised recommendation and notification solutions based on past user interactions.
Full-text available
Research on recommender systems is a challenging task, as is building and operating such systems. Major challenges include non-reproducible research results, dealing with noisy data, and answering many questions such as how many recommendations to display, how often, and, of course, how to generate recommendations most effectively. In the past six years, we built three research-article recommender systems for digital libraries and reference managers, and conducted research on these systems. In this paper, we share some experiences we made during that time. Among others, we discuss the required skills to build recommender systems, and why the literature provides little help in identifying promising recommendation approaches. We explain the challenge in creating a randomization engine to run A/B tests, and how low data quality impacts the calculation of bibliometrics. We further discuss why several of our experiments delivered disappointing results, and provide statistics on how many researchers showed interest in our recommendation dataset.
Conference Paper
Full-text available
TF-IDF is one of the most popular term-weighting schemes, and is applied by search engines, recommender systems, and user modeling engines. With regard to user modeling and recommender systems, we see two shortcomings of TF-IDF. First, calculating IDF requires access to the document corpus from which recommendations are made. Such access is not always given in a user-modeling or recommender system. Second, TF-IDF ignores information from a user's personal document collection, which could – so we hypothesize – enhance the user modeling process. In this paper, we introduce TF-IDuF as a term-weighting scheme that does not require access to the general document corpus and that considers information from the users' personal document collections. We evaluated the effectiveness of TF-IDuF compared to TF-IDF and TF-Only and found that TF-IDF and TF-IDuF perform similarly (click-through rates (CTR) of 5.09% vs. 5.14%), and both are around 25% more effective than TF-Only (CTR of 4.06%) for recommending research papers. Consequently, we conclude that TF-IDuF could be a promising term-weighting scheme, especially when access to the document corpus for recommendations is not possible, and thus classic IDF cannot be computed. It is also notable that TF-IDuF and TF-IDF are not exclusive, so that both metrics may be combined to a more effective term-weighting scheme.
Conference Paper
Full-text available
This paper presents a description of integration of the Mr. DLib scientific recommender system into the JabRef reference manager. Scientific recommender systems help users identify relevant papers out of vast amounts of existing literature. They are particularly useful when used in combination with reference managers. Over 85% of JabRef users stated that they would appreciate the integration of a recommender system. However, the implementation of literature recommender systems requires experience and resources that small companies cannot afford. With the desires of users in mind, we integrated the Mr. DLib scientific recommender system into JabRef. Using Mr. DLib's recommendations-as-a-service, JabRef users can find relevant literature and keep themselves informed about the state of the art in their respective fields.
This paper presents a dataset and supervised learning experiments for term extraction from Slovene academic texts. Term candidates in the dataset were extracted via morphosyntactic patterns and annotated for their termness by four annotators. Experiments on the dataset show that most co-occurrence statistics, applied after morphosyntactic patterns and a frequency threshold, perform close to random and that the results can be significantly improved by combining, with supervised machine learning, all the seven statistic measures included in the dataset. On multi-word terms the model using all statistics obtains an AUC of 0.736 while the best single statistic produces only AUC 0.590. Among many additional candidate features, only adding multi-word morphosyntactic pattern information and length of the single-word term candidates achieves further improvements of the results. KeywordsTerminology extractionSupervised machine learningSlovene language
More and more people read the news online, e.g., by visiting the websites of their favorite newspapers or by navigating the sites of news aggregators. However, the abundance of news information that is published online every day through different channels can make it challenging for readers to locate the content they are interested in. The goal of News Recommender Systems (NRS) is to make reading suggestions to users in a personalized way. Due to their practical relevance, a variety of technical approaches to build such systems have been proposed over the last two decades. In this work, we review the state-of-the-art of designing and evaluating news recommender systems over the last ten years. One main goal of the work is to analyze which particular challenges of news recommendation (e.g., short item life times and recency aspects) have been well explored and which areas still require more work. Furthermore, in contrast to previous surveys, the paper specifically discusses methodological questions and today's academic practice of evaluating and comparing different algorithmic news recommendation approaches based on accuracy measures.
Conference Paper
This paper proposes to exploit author-defined tags and social interaction data (commenting and sharing news items) in news recommendation. Moreover it presents a hybrid news recommender which suggest news items on the basis of the reader’s short and long-term reading history, taking reading trends and short-term interests into account. The experimental results we carried out provided encouraging results about the accuracy of the recommendations.