=Paper=
{{Paper
|id=Vol-2311/paper_4
|storemode=property
|title=Unsupervised Topic Modelling in a Book Recommender System for New Users
|pdfUrl=https://ceur-ws.org/Vol-2311/paper_4.pdf
|volume=Vol-2311
|authors=Haifa Alharthi,Diana Inkpen,Stan Szpakowicz
|dblpUrl=https://dblp.org/rec/conf/sigir/AlharthiIS17
}}
==Unsupervised Topic Modelling in a Book Recommender System for New Users==
https://ceur-ws.org/Vol-2311/paper_4.pdf
Unsupervised Topic Modelling in a Book Recommender System
for New Users
Haifa Alharthi & Diana Inkpen & Stan Szpakowicz
EECS, University of Ottawa, Ottawa, Canada
halha060@uottawa.ca,Diana.Inkpen@uottawa.ca,szpak@eecs.uottawa.ca
ABSTRACT with higher ability of communication, empathy and social support
Book recommender systems (RSs) are useful in libraries, schools and [27, 28].
e-commerce applications. To our knowledge, no book RS exploits One challenge facing RSs is the user cold start. It happens when
social networks other than book-cataloguing websites. We propose new users with no rating history are introduced to the system. A
a recommendation component that learns the user’s interests from book RS may consider non-readers as new users and recommend
social media data and recommends books accordingly. Our new books to them, and so help encourage the practice of reading. An
method of modelling users’ interests acquires a user’s distinctive issue related to cold start is the lack of explicit feedback from
topics using tf-idf and represents them as word embeddings. Even existing users, who may find it burdensome to assign ratings to
though the system is designed to complement other systems, we items.
evaluated it against content-based RS, a traditional book RS, and This paper proposes an automatic personalization module that
obtained similar performance. So, the system’s new user would learns the users’ interests from social media data and recommends
receive recommendation as accurate as current users. books accordingly. The Topic-Model-Based book recommendation
component (TMB) would help existing RSs deal with new users
KEYWORDS with no user rating history. For each user, a topic profile is created
that summarizes subjects discussed on her social media account.
recommender systems, personalization, user modelling, social me-
User profiles are matched with descriptions of books, and the most
dia, Twitter
similar ones are suggested. To evaluate TMB, a dataset was collected
ACM Reference format: that encompasses user profiles on Twitter and Goodreads, a social
Haifa Alharthi & Diana Inkpen & Stan Szpakowicz. 2017. Unsupervised book cataloging Web site. We compared the top k recommendations
Topic Modelling in a Book Recommender System for New Users. In Pro-
made by TMB and content-based system (CB). Both retrieved a
ceedings of ACM Conference, Tokyo, Japan, August 2017 (SIGIR 2017 eCom),
comparable number of books, even though CB relied on users’
8 pages.
rating history while TMB only needed their social profiles. We
conclude that new users would receive recommendations (made by
1 INTRODUCTION TMB) as accurate as those for current users (made by CB).
The information flood on the Internet makes desirable a wide va- Content-based RS is a standard RS which is widely used for book
riety of applications, among them recommender systems (RSs). recommendations [22, 31, 35, 44]. A CB recommender is a classifier
They help limit users’ choices to the possibly most preferred items. that learns the patterns and similarities in the purchase history of
To make suggestions, existing RSs exploit users’ rating history, one user to predict her future interests.
product features, user social-media content and relationships, user Many book RSs exploit social media, as explained in section 2,
personality and emotions, and more. but they are all focused on social networks established mainly for
Investigating book RSs is a worthwhile endeavour. They are readers such as LibraryThing. Our research investigates the use of a
useful in libraries, schools and e-learning portals, as well as book- general platform, Twitter, to make book recommendations. Here is
stores and e-commerce applications. They can help libraries with why: Twitter is not exclusive to bookworms, so it can help address
abundant unused resources—e.g., 75% of the books in the library of the issue of new users who have no reading profiles. Moreover, since
Changsha University of Science and Technology have never been its establishment, Twitter has been used to survey opinions, report
checked out [48]. The practice of reading for pleasure has declined news (more than 85% of the Twitter activities are related to news
events), raise awareness, create social and political movements and
in recent years, especially among children.1 This decline may affect
life quality: readers may be significantly more likely than non- more; topics discussed on this medium have wide diversity and are
readers to report better health/mental health, to volunteer and feel up-to-date [11]. This offers a chance to understand the reactions
strongly satisfied with life [18]. Exposure to fiction also correlates and opinions of active users to their surroundings, e.g., the social
and political scene. It allows the capturing of broad topics that
1http://tinyurl.com/jgunwfx
the user cares about and may not read about them yet which may
help with the over-specialization issue, users receiving non-diverse
Copyright © 2017 by the paper’s authors. Copying permitted for private and academic recommendation lists.
purposes. The remainder of the paper is organized as follows. Section 2
In: J. Degenhardt, S. Kallumadi, M. de Rijke, L. Si, A. Trotman, Y. Xu (eds.): summarizes the related work especially in the domains of social RSs
Proceedings of the SIGIR 2017 eCom workshop, August 2017, Tokyo, Japan, published at for books and news articles. Section 3 gives a high-level description
http://ceur-ws.org
of the system and its components. Section 4 explains the details
SIGIR 2017 eCom, August 2017, Tokyo, Japan Alharthi et al.
of data collection and preprocessing, as well as the system imple- Another system that exploits LibraryThing, presented by [14],
mentation. Section 5 defines the experiment settings and illustrates addresses the new item issue. Each book is characterized by tf-idf
its results. Section 6 discusses the results. Section 7 concludes and vectors of social tags (extracted from LibraryThing) and book tags
suggests future work. (from the whole text of a book). For new books with no available
social tags, a relevance model (RM) is adopted to learn from a
2 RELATED WORK book’s tags to predict social tags. A pure RM gives results similar
to collaborative filtering. To our knowledge, no book RSs exploit
2.1 Topic Modelling of Text in RSs social networks other than book-cataloguing websites.
Topic models have helped estimate preferences in many RSs. To
name a few, recommendations were based on the topics extracted 2.3.2 Twitter-based news recommendations. To make news rec-
from movie plots [4], articles [33, 45], online courses syllabi [2] ommendations, [1] treat a user profile as a query; the k most sim-
and trending categories on e-commerce portals [19]. Unlike the ilar candidate news articles are recommended. User profiles are
previously mentioned work which mainly analyzes textual descrip- constructed from three elements: hashtags, entities and topics. A
tion of items, TMB model the topics discussed in a user profile to concept is weighted by counting the times a user mentions it (e.g.,
capture their interests and make recommendations accordingly. #technology = 5). A framework, OpenCalais, is used to spot names
Based on topics learned from users’ Twitter accounts, RSs could of people, places and other entities in addition to topics; there is a
suggest hashtags [16] and friends [36]. TMB, on the other hand, limitation to 18 different topics (e.g., politics or sports). Entity-based
addresses the new user issue by exploiting tweets to recommend user profiles scored the highest S@k (Success at rank k) at 0.20.
items that are not Twitter-relevant (e.g., not hashtags). [8] propose a Twitter-based URL recommender. Cosine similarity
is computed between user profiles and URL topics, and the system
2.2 Social media and the new user issue recommends URL items with the highest scores. For each user, self-
profile and followee-profile are constructed out of bag-of-words.
Social media have been a great resource to “warm up” the user cold
For a URL, a bag-of-words is also created out of terms occurring in
start. A users connections on social network were exploited in [7],
tweets which embed the URL. In a field experiment, 44 participants
[40], [17], [3] and [29]. In addition to using Facebook friends lists,
rated the recommended URLs. The best performance was 72.1%
[40] analyzed users’ demographics and pages liked by a user.
accuracy when the RS used self-profiles and candidate URLs from
[32] solved the new user issue by analyzing a target user ’s
FoF (followee-of-followees).
tweets and identifying which movie genres she likes. The cosine
Unlike work in [1, 11, 20] which looks for news-related and nar-
similarity between a tweet and a movie storyline is calculated. If the
row lists of entities and categories, TMB is dynamic and represents
similarity is higher than 0.5, the movie’s genre is added to the user’s
the dominant topics discussed by a user without searching for pre-
favourite genres. Later, movies from the most frequent genres are
defined concepts. Our system does not require entity recognition
recommended.
or ontology development.
2.3 Recommendations of textual items
This section covers social RSs dedicated to recommending textual
3 TOPIC-MODEL-BASED BOOK
items, including books and news articles. First, we need to differen- RECOMMENDER SYSTEM
tiate the characteristics of the book and the news recommendation This section explains the TMB components, book and user profiles,
tasks. News has a short lifetime and may become irrelevant within and formally defines the recommendation process.
days or even hours. On the other hand, many books have survived
hundreds of years and are still widely read and recommended. Fur-
3.1 Book and user profiles
thermore, news content is dynamic and changes rapidly / daily.
That requires the analysis of hashtags and entities such as names A book profile (BP) is represented as a vector of terms comprising
and places that may correspond with the news. However, books its description. We used short descriptions of books available online.
include broad aspects and are mostly unrelated to present names On the other hand, a user profile (UP) is a vector that consists of
and actions. Thus, unlike news, book social RSs need to look for terms extracted from the target user’s Twitter timeline. Terms are
users’ long-term interests. elicited from textual content of tweets and their embedded links.
Retweets and replies are included with tweets so as to avoid sparsity,
2.3.1 Book recommendations using social media. LibraryThing while hashtags are counted in if they are spelled correctly. User
is a social book-cataloguing Web site which allows users to form profiles are built automatically using topic modelling techniques
friendships and catalogue and tag books. [37] match books that without being mapped to an external ontology or to predefined
a target user likes with books that her friends like. Each book in categories. For topic modelling, we considered two techniques:
LibraryThing has a cloud of tags, and the system suggests the most Term Frequency - Inverse Document Frequency (tf-idf) and Non-
similar tag-represented books, using a word correlation matrix. Negative Matrix Factorization (NMF). We also experimented with
Another system also uses tags to find similar books in the user’s Latent Dirichlet Allocation (LDA), but did not report the results
friends list [38]. Books are considered similar when they share due to the low performance. This supports the finding of a previous
one or more tags with friends or are highly rated by a user’s most study [39] that NMF performs better than LDA when dealing with
reliable friends. tweets.
Unsupervised Topic Modelling in a Book Recommender System for New Users SIGIR 2017 eCom, August 2017, Tokyo, Japan
3.1.1 Term Frequency - Inverse Document Frequency. The tf-idf of words w 1 , w 2 , . . . wh found in b j ’s description. To recommend
weighting approach is widely used in information retrieval. Term books to ui , TMB calculates the cosine similarity (Equation 3) be-
frequency (t ft,d ) of a term t is the number of times it occurs in tween U Pui and BP for every book in Bui , and suggests the books
document d. A document in this context is all tweets and/or links in with k most similar BP.
one user timeline. Inverse document frequency (Equation 1) helps
distinguish the terms that are specific to a user/document. U Pui · BP j
similarity = (3)
N U Pui ∗ BP j
id ft = loд (1)
d ft If terms are replaced by their word embeddings, an average
N is the number of users and d ft is the number of documents vector is created for word vectors in U Pui and another for BP j .
where term t occurs. Equation 2 defines the tf-idf weight of term t Then, cosine similarity is performed between the resulting average
in document d. vectors.
t f -id ft,d = t ft,d ∗ id ft (2) 4 DATA PREPARATION AND SYSTEM
The terms with highest weights are considered the tf-idf topic IMPLEMENTATION
model [26]. This section describes how the dataset was collected and prepro-
cessed. It also presents the implementation of the system, unfolding
3.1.2 Non-Negative Matrix Factorization. This dimensionality
technical details of the creation of book and user profiles.
reduction and topic modelling technique has been found to work
well with short text [9, 15, 47]. For a user, a term-document matrix
is created; a document here is one tweet or link. NMF factorizes the
4.1 Data collection
m × n term-document matrix A into two non-negative matrices W We collected user data from Goodreads and Twitter, because there
and H . The former represent the term-topic matrix m × k, whereas are no datasets with both users social profiles and their reading
the latter is the topic-document matrix k × n. The number of NMF lists. The Twitter API was queried to retrieve any review shared
topics k should be defined ahead of decomposition. The matrix WH by Goodreads users, and more than 1000 tweets were found, from
approximates the original matrix A. Every document in WH repre- which we accessed their authors and IDs. Twitter API allows the
sents a linear combination of k topic vectors in W with coefficients collection of a maximum of 3500 tweets per user. We gathered text,
given by H [15]. ID and date of creations of tweets for user with Goodreads review.
Links were extracted from user timelines and their textual contents
3.1.3 Topic embeddings. Word embeddings have gained a lot of (if any) were collected. This was achieved by applying an efficient
attention lately thanks to the revival of neural networks. They are Python library called Newspaper, which obtains a clean tag-free
word vectors with fixed dimensions. We use the word2vec model, text from a given Web page. Once the Twitter user profiles were
proposed by Mikolov et al. [30]. Terms in book and user profiles are complete, we collected data from Goodreads for the book profiles.
mapped to word embeddings produced by the word2vec model. The User Goodreads IDs were obtained from the tweets of default re-
model is trained on a very large amount of text and can predict the views. Next, a scrapper was developed to retrieve all review IDs and
context of a given word. It represents words in a space where two dates from users’ “read books” lists, which contain only completed
words occurring in similar contexts are neighbours. We used pre- books. The Goodreads API was consulted to extract information
trained word embeddings developed on the Google News dataset of about all books read by a user, including book metadata, text re-
around 100 billion words. It comprises vectors of 300 dimensions for views, ratings, read date and added date. The book metadata, which
3 million words and phrases.2 Other available pre-trained models can be used to build content-based recommender systems, include
(e.g., Global Vectors for Word Representation3 ) have been built using ISBN, ISBN13, title, authors, language, the average rating of all
text from Twitter and Wikipedia, but the Google news embeddings reader, the number of pages, publisher, publication date, text review
are more relevant to both books and tweets. While books have count and book description. The read date indicates the time of
formal descriptions, tweets are casual, with hashtags that require a completion of a read book, while the added date is the time when a
model which encompasses abbreviations. book was catalogued.
When users insert new books into their lists, they may discuss
3.2 The recommendation procedure them on their social media. Therefore, in TMB, the recommendation
Let U = u 1 , u 2 , . . . un be a set of Twitter users. For user ui , a timeframe RT considers added dates instead of read dates. The
time threshold Tui is established to avoid the overlap in learn- rating scale, according to Goodreads, treats 1-2 stars as “dislike”,
ing and prediction times. The learning timeframe LTui involves and 3-5 stars as “like”; books rated 3-5 will be called relevant in the
all tweets and links created by ui before Tui , whereas the recom- remainder of the paper. The number of users shrunk to 69 after the
mendations timeframe RTui contains books read by ui after Tui . deletion of non-English users, inactive users and those with private
For user ui ∈ U , a user profile U Pui is a vector comprising terms Goodreads accounts. Even though many datasets with large number
w 1 , w 2 , . . . wm extracted from tweets or links shared by ui during of users exist, some recommendation methodologies such as TMB
LTui . Let Bui = b1 , b2 , . . . bl be the set of books read by user ui require personal information about users. This makes it hard to
during RTui . For book b j ∈ Bui , the book profile BP j is a vector experiment on large datasets. Examples of such work include [41]
2 https://code.google.com/archive/p/word2vec/ which used a dataset of 52 users to test affective-based RS,and [34]
3 http://nlp.stanford.edu/projects/glove/ which tested a context-aware RS on an 89-user dataset.
SIGIR 2017 eCom, August 2017, Tokyo, Japan Alharthi et al.
Figure 1: Dataset collection and statistics.
4.2 Data preprocessing numbers are 121 and 13 respectively. The lowest number of ratings
Before topic extraction, text of tweets, link, and book descriptions needed to develop CB with quality recommendations is 10. This
must be cleaned. The tweets were tokenized using Tweet Tokenizer threshold is adopted by many researchers, including [46].
from NLTK [5], which is Twitter-conscious, and tagged using the We developed twelve variations of user profiles. They differ in
GATE Twitter part of speech tagger [12]. Hashtags were checked the topic modelling technique (NMF or tf-idf), in the source of data
using aspell.4 Misspelled words were excluded because they are not (tweets alone, links alone, or tweets and links) and in the word
useful: the goal is to match them with book descriptions which are representations (embeddings [emb] or none). The NMF algorithm
spelling-error-free. For links and book descriptions, regular NLTK was implemented using the scikit-learn Python package [6]. After
Word tokenizer [5] and Stanford part-of-speech tagger [42] were conducting many trials, the number of NMF topics was set to five,
applied. Only nouns (singular or plural) were kept, then lemmatized with six words in each, because topics became redundant afterward.
by NLTK WordNet Lemmatizer [5]. A noun, according to Merriam- The number of tf-idf topics was set to 100. To calculate cosine
Webster, represents an “entity, quality, state, action, or concept”. similarity between words vectors, we used genism, a Python library.
Nouns, then, can capture the interests of users more than any Not all topics have corresponding word vectors, and a reduction in
other part of speech. In fact, to model user interests based on their the number of topics is expected.
social media accounts, other researchers also considered only nouns
[10][43]. 5 EXPERIMENTS AND RESULTS
After building topic models from tweets and links, we noticed We measure the predictive power of the system using off-line evalu-
that unimportant (generic) terms such as “website” are dominant. ation, which is appropriate for obtaining the accuracy of an RS. The
Therefore, we went further by excluding NLTK stop words, 100 on-line appraisal would provide more performance insights, but it
most common English nouns,5 and words of fewer than 4 letters. For is an expensive option that requires the deployment of a real-time
tweets, we also filtered out the 200 words with lowest idf weights.6 version of TMB. A user study is another option; it was avoided
Repeated content of links is deleted, and so are Web-related terms, because it usually includes a limited number of users.
e.g., “website” and “Facebook”.
5.1 Experiment settings
4.3 System implementation
Strategies of top k recommendations are tested in a similar fashion
A user Twitter timeline was divided in half, and the date of the to the leave-one-out evaluation applied in [13, 21, 23]; it splits the
middle tweet was considered a time threshold that differentiates dataset into a training set and a one-item test set, then generates
learning and recommendations periods. To ensure that tweets do a list of the top N recommendations from the training set. In our
not address the predicted books, a one-month difference was set setting, however, the training set is made up of all books not in-
between the timeframes. The average numbers of tweets and books cluded in the recommendation timeframe, so we cannot use it for
included in the learning period are 758 and 802, while the minimum prediction. This is why we followed a slightly different assessment
4 http://aspell.net/
methodology, used in ranking-based RSs adopted by [24] and [25].
5 http://www.linguasorb.com/english/most-common-nouns/ We created one set of 1000 random books that are unique and
6 The idf weights for tweets from all users after the deletion of stop words. not rated by any user. For each user, we randomly selected one
Unsupervised Topic Modelling in a Book Recommender System for New Users SIGIR 2017 eCom, August 2017, Tokyo, Japan
relevant book from the recommendation timeframe, added it to allows the capturing of distinctive topics frequently discussed by
the 1000 books and asked our system to perform ranking. If the one user in contrast with those discussed by her community. To
rank of the relevant book is f, the RS should have the lowest f eliminate noise, we only kept the top tf-idf words. Otherwise, the
value (preferably 1). If f ≤ k, it is a hit, otherwise it is a miss. average word embedding of all terms in Twitter time-line would be
Similarly to many related projects, we set k to 10. Metrics adopted skewed towards less significant terms. We think that this method
are hit-rate (Equation 4), sometimes called recall, and the average obtains fine-grained interests not extensively shared among users.
reciprocal hit-rank (Equation 5) [13]. To avoid a bias, five trials For example, a term that is not as popular, like “mythology”, may
were conducted, and the reported results averaged. We measured have a high idf value and be in the top tf-idf list.
the statistical difference in results using the t-test at a maximum of All variations of TMB could identify books that interest the
p-value = 0.05. user out of a thousand other books, with the link-based tf-idf-emb
#hits retrieving the highest number of books. To illustrate how word
HR = (4)
#users embeddings contribute to the recommendations made by link-based
#hit s tf-idf-emb, we plotted (Figure 3) the word embeddings of one user
1 Õ 1
ARHR = (5) profile (b) and his two book profiles (c, d). Section (a) of Figure 3
#users i=1 fi shows the closeness of word vectors found the UP and BPs. One can
Our approach was compared to a content-based RS and to a notice the variety of topics in the user profile. User interests might
random system. CB was implemented using the default settings of be broad and not only related to the books they already preferred.
Graphlab, a well-established framework for RSs. Books in the CB The textual content of the links can be longer than that of the
training and test sets were represented with book metadata (see tweets, and so possibly capturing a wider range of interests. In fact,
section 4.1). Although we considered a comparison with collabora- there is an evident difference in their effect on pure NMF and tf-idf.
tive filtering (CF), the rating matrix is highly sparse, which means Thanks to word embeddings, however, the performance of models
that the results would not reflect a typical CF. that adopt tweets increased dramatically. Word vectors could enrich
Some of the randomly chosen 1000 books might have topics the topics by including the context of terms. Their improvement of
similar to a user profile. On the other hand, they could share similar tweet-based algorithms could be due to the presence of hashtags,
content, e.g., author or description with a user’s read books. How- which summarize a whole subject or event.
ever, we did not filter out such books because this would introduce We conducted error analysis to investigate the differences in
a bias and favour one system over the other. In addition, for each performance between CB and TMB (tf-idf based on links and word
of the 69 users, we examined five books, so the overall number of embeddings). In a leave-one-out evaluation, we tested five books
tested books is 345. If there is a bias with a few books, it should not for each user. The two systems retrieved the same number of rele-
affect the majority of test cases. vant books when giving recommendations to 31 users. CB could
retrieve more relevant books than TMB for 17 users, whereas TMB
5.1.1 Results. Figure 2 shows the HR and ARHR scores of four- surpassed CB when dealing with 21 users. For better understanding,
teen recommendation techniques. The best-performing methods we analyzed each system’s best recommendations.
are CB and tf-idf-emb built with links; it achieved the highest HR CB retrieved three out of five books relevant for users A and B,
results, while CB reached the best ARHR. Tweet-based tf-idf-emb while TMB suggested only one book to user A and none to B. User
has similar results to CB. The results of these two methods are not A had 512 books in the CB training set, while user B had 542. From
statistically different. In general, tf-idf gives better results than NMF. the three books recommended to A, only one had the same author
This is expected due to the difference in numbers of topic terms. as a book in the training set; that is to say, CB relied on book de-
Comparing the algorithms with and without word embedding vec- scriptions to make the recommendation. The one book which TMB
tors, the addition of word embedding enhances the performance. recommended to user A had cosine similarity of 0.69 and shared
The results are statistically different except for the tf-idf-emb of words that were semantically close to the user topics (e.g., drawing
tweets and the tf-idf-emb of tweets and links. There is no consis- vs. illustrator). Like for user A, only one book recommended by CB
tency in the effect of using tweets or links. For example, using links to user B shared the same author with a book in the CB training
with tf-idf gives the highest score but with NMF-emb the score is set. The possible reason why TMB could not suggest any book to
the lowest among all data categories. The random system could not user B is that the user’s topics were related to political issues (e.g.,
bring any relevant book to the top k. abortion, immigration), while the user’s readings were diverse. For
example, user B’s five relevant books addressed history, romance,
6 DISCUSSION philosophy and education. The user’s interests were broad, while
The field of RSs is active. Many state-of-the-art recommendation his discussed topics on Twitter were narrow and related to current
methods have been proposed in the recent years. However, we only issues.
compared TMB results with CB which has been around for a long Users C, D and E received five, three and two recommended
while. TMB gives similar performance to a traditional system, CB books by TMB, respectively, while CB could recommend two books
without the need for user rating history. Nevertheless, we do not for user C and none for user D and E. User C had 140 books in
claim that the system works independently. To verify this, more the CB training set. Most of his readings were related to religious
comprehensive experiments are required. matters. The user topic profile reflected these interests: the top tf-idf
One suggested method, which gives the best results, is to use words were glory, theology and gospel. The lowest cosine similarity
word embeddings of top tf-idf terms. The use of tf-idf weighting
SIGIR 2017 eCom, August 2017, Tokyo, Japan Alharthi et al.
Figure 2: The comparison of TMB approaches, CB and the random system.
Figure 3: Word embeddings of terms in a user profile and two book descriptions.
Unsupervised Topic Modelling in a Book Recommender System for New Users SIGIR 2017 eCom, August 2017, Tokyo, Japan
between the user topic profile and the five retrieved books was 0.72. [7] Eduardo Castillejo, Aitor Almeida, and Diego López-De-Ipiña. 2012. Alleviating
User D had 13 books in the CB training set. All books in the training cold-user start problem with users’ social network data in recommendation
systems. In Workshop on Problems and Applications in AI, ECAI-12.
and test set were written by distinct authors. The user topics were [8] Jilin Chen, Rowan Nairn, Les Nelson, Michael Bernstein, and Ed Chi. 2010. Short
also related to philosophy and religion, as well as the user readings and Tweet: Experiments on Recommending Content from Information Streams.
In Proc. SIGCHI Conference on Human Factors in Computing Systems (CHI ’10).
(see Figure 3). User E had 528 books in the CB training set. Her topic ACM, 1185–1194. DOI:https://doi.org/10.1145/1753326.1753503
profile covered wide interests (e.g., courtroom, femininity, mutiny, [9] Xueqi Cheng, Jiafeng Guo, Shenghua Liu, Yanfeng Wang, and Xiaohui Yan. 2013.
and heroin) and the two recommended books were slightly similar. Learning Topics in Short Texts by Non-negative Matrix Factorization on Term
Correlation Matrix.. In Proc. SIAM International Conference on Data Mining. SIAM,
One of them, titled “Against the Country”, was described with words 749–757. http://dblp.uni-trier.de/db/conf/sdm/sdm2013.html#ChengGLWY13
such as offender, antihero and blast. The other book was described [10] D. Choi, J. Kim, E. Lee, C. Choi, J. Hong, and P. Kim. 2014. Research for the
with such words as assassination and murder. Pattern Analysis of Individual Interest Using SNS Data: Focusing on Facebook.
In 2014 Eighth International Conference on Innovative Mobile and Internet Services
in Ubiquitous Computing. 36–40. DOI:https://doi.org/10.1109/IMIS.2014.94
7 CONCLUSION [11] Gianmarco De Francisci Morales, Aristides Gionis, and Claudio Lucchese. 2012.
From Chatter to Headlines: Harnessing the Real-time Web for Personalized News
This paper proposes TMB, a system that builds a topic model for a Recommendation. In Proc. Fifth ACM International Conference on Web Search and
Data Mining (WSDM ’12). ACM, 153–162. DOI:https://doi.org/10.1145/2124295.
user from textual content shared voluntarily on her social media, 2124315
and recommends the books most related to these topics. We ac- [12] Leon Derczynski, Alan Ritter, Sam Clark, and Kalina Bontcheva. 2013. Twitter
quired a user’s distinctive topics by tf-idf weighting and represent Part-of-Speech Tagging for All: Overcoming Sparse and Noisy Data. In Proc.
International Conference on Recent Advances in Natural Language Processing.
them as word embeddings in order to capture their context. TMB Association for Computational Linguistics.
achieves a recommendation accuracy similar to CB, a commonly [13] Mukund Deshpande and George Karypis. 2004. Item-based top-N Recommen-
used book RS, particularly when word embeddings are deployed. dation Algorithms. ACM Trans. Inf. Syst. 22, 1 (Jan. 2004), 143–177. DOI:
https://doi.org/10.1145/963770.963776
Thus, TMB can aid current RSs in suggesting books to new users [14] Sharon Givon and Victor Lavrenko. 2009. Predicting Social-tags for Cold Start
without major loss in performance. Book Recommendations. In Proc. Third ACM Conference on Recommender Systems
(RecSys ’09). ACM, 333–336. DOI:https://doi.org/10.1145/1639714.1639781
For future improvement, we plan to study the temporal effect on [15] Daniel Godfrey, Caley Johns, Carl Dean Meyer, Shaina Race, and Carol Sadek.
topic models, as well as the relationship between the level of user 2014. A Case Study in Text Mining: Interpreting Twitter Data From World Cup
activity and the accuracy of the recommendations. Since hashtags Tweets. CoRR abs/1408.5427 (2014). http://arxiv.org/abs/1408.5427
[16] Fréderic Godin, Viktor Slavkovikj, Wesley De Neve, Benjamin Schrauwen, and
carry more meaning than other terms on Twitter, an interesting Rik Van de Walle. 2013. Using Topic Models for Twitter Hashtag Recommen-
approach would be to create hashtag-based profiles that are en- dation. In Proc. 22Nd International Conference on World Wide Web (WWW ’13
riched with word vectors. Also, user profiles could include other Companion). ACM, 593–596. DOI:https://doi.org/10.1145/2487788.2488002
[17] Ido Guy, Naama Zwerdling, Inbal Ronen, David Carmel, and Erel Uziel. 2010. So-
parts of speech, especially verbs and adjectives. To enhance the cial Media Recommendation Based on People and Tags. In Proc. 33rd International
performance of the system, we will train embeddings for tweets ACM SIGIR Conference on Research and Development in Information Retrieval
(SIGIR ’10). ACM, 194–201. DOI:https://doi.org/10.1145/1835449.1835484
and books. [18] Kelly Hill. 2013. The Arts and Individual Well-Being in Canada. (February 2013).
http://www.hillstrategies.com/content/arts-and-individual-well-being-canada
ACKNOWLEDGMENT [Online; posted 13 February 2013].
[19] Diane J. Hu, Rob Hall, and Josh Attenberg. 2014. Style in the Long Tail: Dis-
Support for this work has come from the Natural Sciences and covering Unique Interests with Latent Variable Models in Large Scale Social
E-commerce. In Proc. 20th ACM SIGKDD International Conference on Knowledge
Engineering Research Council of Canada. Discovery and Data Mining (KDD ’14). ACM, 1640–1649. DOI:https://doi.org/10.
1145/2623330.2623338
[20] Nirmal Jonnalagedda, Susan Gauch, Kevin Labille, and Sultan Alfarhood. 2016.
REFERENCES Incorporating popularity in a personalized news recommender system. PeerJ
[1] Fabian Abel, Qi Gao, Geert-Jan Houben, and Ke Tao. 2011. Analyzing User Computer Science 2 (2016), e63.
Modeling on Twitter for Personalized News Recommendations. In Proc. 19th In- [21] Zhao Kang, Chong Peng, and Qiang Cheng. 2016. Top-N Recommender System
ternational Conference on User Modeling, Adaption, and Personalization (UMAP’11). via Matrix Completion. In Proc. Thirtieth AAAI Conference on Artificial Intelligence
Springer-Verlag, Berlin, Heidelberg, 1–12. http://dl.acm.org/citation.cfm?id= (AAAI’16). AAAI Press, 179–184. http://dl.acm.org/citation.cfm?id=3015812.
2021855.2021857 3015839
[2] Rel Guzman Apaza, Elizabeth Vera Cervantes, Laura Cruz Quispe, and Jose Ochoa [22] Hikmet Kapusuzoglu and Sule Gunduz Öguducu. 2011. A Relational Rec-
Luna. 2014. Online Courses Recommendation based on LDA. (2014). ommender System Based on Domain Ontology. In Emerging Intelligent Data
[3] David Ben-Shimon, Alexander Tsikinovsky, Lior Rokach, Amnon Meisles, Guy and Web Technologies (EIDWT), 2011 International Conference on. 36–41. DOI:
Shani, and Lihi Naamani. 2007. Recommender System from Personal Social Net- https://doi.org/10.1109/EIDWT.2011.15
works. In Advances in Intelligent Web Mastering: Proc. 5th Atlantic Web Intelligence [23] George Karypis. 2001. Evaluation of Item-Based Top-N Recommendation Algo-
Conference – AWIC’2007, Fontainbleau, France, June 25 – 27, 2007, Katarzyna M. rithms. In Proc. Tenth International Conference on Information and Knowledge Man-
Wegrzyn-Wolska and Piotr S. Szczepaniak (Eds.). Springer Berlin Heidelberg, agement (CIKM ’01). ACM, 247–254. DOI:https://doi.org/10.1145/502585.502627
Berlin, Heidelberg, 47–55. DOI:https://doi.org/10.1007/978-3-540-72575-6_8 [24] Yehuda Koren. 2008. Factorization Meets the Neighborhood: A Multifaceted
[4] Sonia Bergamaschi and Laura Po. 2015. Comparing LDA and LSA Topic Models Collaborative Filtering Model. In Proc. 14th ACM SIGKDD International Conference
for Content-Based Movie Recommendation Systems. In Web Information Systems on Knowledge Discovery and Data Mining (KDD ’08). ACM, 426–434. DOI:https:
and Technologies: 10th International Conference, WEBIST 2014, Barcelona, Spain, //doi.org/10.1145/1401890.1401944
April 3-5, 2014, Revised Selected Papers, Valérie Monfort and Karl-Heinz Krempels [25] Qiuxia Lu, Tianqi Chen, Weinan Zhang, Diyi Yang, and Yong Yu. 2012. Serendip-
(Eds.). Springer International Publishing, Cham, 247–263. DOI:https://doi.org/ itous Personalized Ranking for Top-N Recommendation. In Proc. The 2012
10.1007/978-3-319-27030-2_16 IEEE/WIC/ACM International Joint Conferences on Web Intelligence and Intelligent
[5] Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural Language Processing Agent Technology - Volume 01 (WI-IAT ’12). IEEE Computer Society, Washington,
with Python (1st ed.). O’Reilly Media, Inc. DC, USA, 258–265. http://dl.acm.org/citation.cfm?id=2457524.2457692
[6] Lars Buitinck, Gilles Louppe, Mathieu Blondel, Fabian Pedregosa, Andreas [26] Christopher D. Manning, Prabhakar Raghavan, and Hinrich Schütze. 2008. Intro-
Mueller, Olivier Grisel, Vlad Niculae, Peter Prettenhofer, Alexandre Gramfort, duction to Information Retrieval. Cambridge University Press.
Jaques Grobler, Robert Layton, Jake VanderPlas, Arnaud Joly, Brian Holt, and [27] Raymond A. Mar and Keith Oatley. 2008. The Function of Fiction is the Abstrac-
Gaël Varoquaux. 2013. API design for machine learning software: experiences tion and Simulation of Social Experience. Perspectives on Psychological Science 3,
from the scikit-learn project. In ECML PKDD Workshop: Languages for Data 3 (01 May 2008), 173–192. DOI:https://doi.org/10.1111/j.1745-6924.2008.00073.x
Mining and Machine Learning. 108–122.
SIGIR 2017 eCom, August 2017, Tokyo, Japan Alharthi et al.
[28] Raymond A. Mar, Keith Oatley, and Jordan B. Peterson. 2009. Exploring the [45] Chong Wang and David M. Blei. 2011. Collaborative Topic Modeling for Rec-
link between reading fiction and empathy: Ruling out individual differences ommending Scientific Articles. In Proc. 17th ACM SIGKDD International Confer-
and examining outcomes. Communications 34, 4 (1 Dec. 2009), 407–428. DOI: ence on Knowledge Discovery and Data Mining (KDD ’11). ACM, 448–456. DOI:
https://doi.org/10.1515/comm.2009.025 https://doi.org/10.1145/2020408.2020480
[29] Daniel Mican, Loredana Mocean, and Nicolae Tomai. 2012. Building a Social [46] Yiwen Wang, Natalia Stash, Lora Aroyo, Laura Hollink, and Guus Schreiber.
Recommender System by Harvesting Social Relationships and Trust Scores 2009. Using Semantic Relations for Content-based Recommender Systems in
between Users. In Business Information Systems Workshops: BIS 2012 International Cultural Heritage. In Proceedings of the 2009 International Conference on Ontology
Workshops and Future Internet Symposium, Vilnius, Lithuania, May 21-23, 2012 Patterns - Volume 516 (WOP’09). CEUR-WS.org, Aachen, Germany, Germany,
Revised Papers, Witold Abramowicz, John Domingue, and Krzysztof Węcel (Eds.). 16–28. http://dl.acm.org/citation.cfm?id=2889761.2889763
Springer Berlin Heidelberg, Berlin, Heidelberg, 1–12. DOI:https://doi.org/10. [47] Xiaohui Yan, Jiafeng Guo, Shenghua Liu, Xue-qi Cheng, and Yanfeng Wang. 2012.
1007/978-3-642-34228-8_1 Clustering Short Text Using Ncut-weighted Non-negative Matrix Factorization.
[30] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient In Proc. 21st ACM International Conference on Information and Knowledge Manage-
Estimation of Word Representations in Vector Space. CoRR abs/1301.3781 (2013). ment (CIKM ’12). ACM, 2259–2262. DOI:https://doi.org/10.1145/2396761.2398615
http://arxiv.org/abs/1301.3781 [48] Xuejun Yang, Hongchun Zeng, and Weihong Huang. 2009. ARTMAP-Based
[31] Raymond J. Mooney and Loriene Roy. 2000. Content-based Book Recommending Data Mining Approach and Its Application to Library Book Recommendation.
Using Learning for Text Categorization. In Proc. Fifth ACM Conference on Digital In Intelligent Ubiquitous Computing and Education, 2009 International Symposium
Libraries (DL ’00). ACM, 195–204. DOI:https://doi.org/10.1145/336597.336662 on. 26–29. DOI:https://doi.org/10.1109/IUCE.2009.43
[32] P. Nair, M. Moh, and T. S. Moh. 2016. Using Social Media Presence for Alleviating
Cold Start Problems in Privacy Protection. In 2016 International Conference on
Collaboration Technologies and Systems (CTS). 11–17. DOI:https://doi.org/10.
1109/CTS.2016.0022
[33] Sergey Nikolenko. 2015. SVD-LDA: Topic Modeling for Full-Text Recommender
Systems. In Advances in Artificial Intelligence and Its Applications: 14th Mexican In-
ternational Conference on Artificial Intelligence, MICAI 2015, Cuernavaca, Morelos,
Mexico, October 25-31, 2015, Proceedings, Part II, Obdulia Pichardo Lagunas, Oscar
Herrera Alcántara, and Gustavo Arroyo Figueroa (Eds.). Springer International
Publishing, Cham, 67–79. DOI:https://doi.org/10.1007/978-3-319-27101-9_5
[34] Ante Odić, Marko Tkalčič, Andrej Košir, and Jurij F. Tasič. 2011. A.: Relevant
context in a movie recommender system: Users’ opinion vs. statistical detection.
In In: Proc. of the 4th Workshop on Context-Aware Recommender Systems (2011.
[35] Dharmendra Pathak, Sandeep Matharia, and C. N. S. Murthy. 2013. NOVA:
Hybrid book recommendation engine. In Advance Computing Conference (IACC),
2013 IEEE 3rd International. 977–982. DOI:https://doi.org/10.1109/IAdCC.2013.
6514359
[36] Marco Pennacchiotti and Siva Gurumurthy. 2011. Investigating Topic Models
for Social Media User Recommendation. In Proc. 20th International Conference
Companion on World Wide Web (WWW ’11). ACM, 101–102. DOI:https://doi.
org/10.1145/1963192.1963244
[37] Maria Soledad Pera, Nicole Condie, and Yiu-Kai Ng. 2011. Personalized Book
Recommendations Created by Using Social Media Data. In Proc. 2010 International
Conference on Web Information Systems Engineering (WISS’10). Springer-Verlag,
Berlin, Heidelberg, 390–403. http://dl.acm.org/citation.cfm?id=2044492.2044531
[38] Maria Soledad Pera and Yiu-Kai Ng. 2011. With a Little Help from My Friends:
Generating Personalized Book Recommendations Using Data Extracted from a
Social Website. In Proc. 2011 IEEE/WIC/ACM International Conferences on Web
Intelligence and Intelligent Agent Technology - Volume 01 (WI-IAT ’11). IEEE
Computer Society, Washington, DC, USA, 96–99. DOI:https://doi.org/10.1109/
WI-IAT.2011.9
[39] Ankan Saha and Vikas Sindhwani. 2012. Learning Evolving and Emerging Topics
in Social Media: A Dynamic Nmf Approach with Temporal Regularization. In
Proc. Fifth ACM International Conference on Web Search and Data Mining (WSDM
’12). ACM, 693–702. DOI:https://doi.org/10.1145/2124295.2124376
[40] Suvash Sedhain, Scott Sanner, Darius Braziunas, Lexing Xie, and Jordan Chris-
tensen. 2014. Social Collaborative Filtering for Cold-start Recommendations. In
Proc. 8th ACM Conference on Recommender Systems (RecSys ’14). ACM, 345–348.
DOI:https://doi.org/10.1145/2645710.2645772
[41] Marko Tkalčič, Andrej Košir, and Jurij Tasič. 2013. The LDOS-PerAff-1 corpus
of facial-expression video clips with affective, personality and user-interaction
metadata. Journal on Multimodal User Interfaces 7, 1 (2013), 143–155. DOI:
https://doi.org/10.1007/s12193-012-0107-7
[42] Kristina Toutanova, Dan Klein, Christopher D. Manning, and Yoram Singer.
2003. Feature-rich Part-of-speech Tagging with a Cyclic Dependency Network.
In Proc. 2003 Conference of the North American Chapter of the Association for
Computational Linguistics on Human Language Technology - Volume 1 (NAACL
’03). Association for Computational Linguistics, Stroudsburg, PA, USA, 173–180.
DOI:https://doi.org/10.3115/1073445.1073478
[43] Keita Tsuji, Nobuya Takizawa, Sho Sato, Ui Ikeuchi, Atsushi Ikeuchi, Fuyuki
Yoshikane, and Hiroshi Itsumura. 2014. Book Recommendation Based on Library
Loan Records and Bibliographic Information. Procedia - Social and Behavioral
Sciences (2014), 478–486. DOI:https://doi.org/10.1016/j.sbspro.2014.07.142 3rd
International Conference on Integrated Information (IC-ININFO).
[44] Paula Cristina Vaz, Ricardo Ribeiro, and David Martins de Matos. 2013. Book Rec-
ommender Prototype Based on Author’s Writing Style. In Proc. 10th Conference on
Open Research Areas in Information Retrieval (OAIR ’13). LE CENTRE DE HAUTES
ETUDES INTERNATIONALES D’INFORMATIQUE DOCUMENTAIRE, Paris,
France, France, 227–228. http://dl.acm.org/citation.cfm?id=2491748.2491800