Visual Recommendations for Scientiﬁc and Cultural Content

Eduardo Veas, Belgin Mutlu, Cecilia di Sciascio, Gerwald Tschinkel and Vedran Sabol

Knowledge Visualization Group, Know Center GmbH, Inffeldgasse 13, Graz, Austria

Keywords:

Visual Recommendation, Recommender Systems, Automated Visualization of Recommendations.

Abstract:

Supporting individuals who lack experience or competence to evaluate an overwhelming amout of information

such as from cultural, scientiﬁc and educational content makes recommender system invaluable to cope with

the information overload problem. However, even recommended information scales up and users still need

to consider large number of items. Visualization takes a foreground role, letting the user explore possibly

interesting results. It leverages the high bandwidth of the human visual system to convey massive amounts of

information. This paper argues the need to automate the creation of visualizations for unstructured data adapt-

ing it to the user’s preferences. We describe a prototype solution, taking a radical approach considering both

grounded visual perception guidelines and personalized recommendations to suggest the proper visualization.

1 INTRODUCTION

The knowledge hidden within huge number of cul-

tural, scientiﬁc and educational repositories in the

Web is usually neither easy to recognize by the gen-

eral public nor to utilize in scientiﬁc and educational

processes. People have to navigate a multitude of li-

braries, repositories and databases searching relevant

content for their tasks. Finding the intended infor-

mation in this way in a huge and continuously grow-

ing information space is a pressing challenge. This

tedious, time-consuming task is currently one of the

main research topics in the context of information re-

trieval for digital libraries, for open educational and

cultural repositories. Novel techniques are needed to

represent, organize, and provide the content to users

in a seamless way (Schl¨otterer et al., 2014).

Two distinguished techniques support users in

searching through massive amounts of information:

recommender systems (RS) and visualization. RSs

help users choose the right items by ﬁltering out ir-

relevant information and suggesting only the relevant

ones. The EU funded project EEXCESS

strives to

use the effectiveness of RSs to bring relevant cultural,

scientiﬁc and educationalcontent directly to the user’s

habitual environment (browser, content management

systems, mobile platforms) (Granitzer et al., 2013).

While recommending content, it is important to pro-

vide users with tools to effectively analyze the rec-

ommendation space. The usual recommendation list

EEXCESS project homepage: http://eexcess.eu/

quickly grows and becomes uncomprehensible with

many results. Our main goal is to augment user inter-

faces with visualizations, to assist user in analysis and

exploration of the recommendation space. Yet, de-

spite a broad understanding of visual perception and

how visual features are used for visual encoding, cre-

ating a visualization for arbitrary datasets remains an

expert task: choosing the right chart, choosing the

right data ﬁelds from the available ones for a meanig-

ful visualization require an understanding of the visu-

alization goal in terms of user needs and preferences.

This paper contends the need for adaptivemethods

that automatically create visualizations for unstruc-

tured datasets. We present a prototype visual recom-

mender with the distinguished feature, and contribu-

tion of the paper, that it automatically suggests ap-

propriate visualizations for the recommendations. As

recommendations are described with arbitrary ﬁelds,

proposing generic visualization is an inherent chal-

lenge. To deal with heterogeneous data formats we

rely on semantic technologies to: (a) deﬁne of data

models to semantically enrich an arbitrarily formated

recommendation list; (b) and to semantically describe

visualizations; (c) deﬁne mapping strategies between

those data models; and (d) algorithm to automatically

conduct such a mapping and suggest visualizations

for recommended items. Beyond a systematic ap-

proach to suggest appropriate charts and mappings we

propose methods to elicit and account for user prefer-

ences in the choice of the right visualization.

256

Veas E., Mutlu B., di Sciascio C., Tschinkel G. and Sabol V..

Visual Recommendations for Scientiﬁc and Cultural Content.

DOI: 10.5220/0005352802560261

In Proceedings of the 6th International Conference on Information Visualization Theory and Applications (IVAPP-2015), pages 256-261

ISBN: 978-989-758-088-8

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: Visual recommender Workﬂow: workﬂow to generate visualizations from a recommendation list.

2 RELATED WORK

For a recommender system, presentation inﬂuences

the user experience (Konstan and Riedl, 2012), and

the overall effectiveness of a recommendation or a

list thereof (Swearingen and Sinha, 2001; Ricci et al.,

2011). The items in a large list of recommendations

will have a broadset of characteristics relating them to

the user model and expectations(Kagie et al., 2011).

Visualization provides a useful overview and interac-

tive method to explore the recommendation space.

Suggesting or recommending visualizations is a

relatively new research topic. Much as recom-

mender systems, visualization has grown as a re-

search ﬁeld addressing the information overload prob-

lem, grounded on semiotics (Bertin, 1983), visual per-

ception (Mackinlay, 1986) and general visual encod-

ing principles (Munzner, 2011). The guidelines con-

tained in this broad body of research are mostly tar-

geted towards a visualization expert. However, sev-

eral systems, such as Caleydo (Lex et al., 2012) in-

clude aids for users to choose visualizations for their

data. Our approach differs in that it automatically sug-

gest visualizations for heterogeneous datasets, even

for datasets such as a recommendation lists, that the

user can then personalize.

To aid users in choosing a visualization for their

data, Mackinlay et al. present ShowMe (Mackinlay

et al., 2007), an integrated set of user interface com-

mands and functions aiming to automatically present

visualizations in Tableau

. The basic idea behind

ShowMe is to support the user by searching for graph-

ical presentations that may address her task. To do so,

the system takes advantages of an algebraic speciﬁ-

cation language, VizQL, describing the structure of a

graphical representation and the queries to match the

data with this structure. The user interface commands

Tableau homepage:http://www.tableausoftware.com/

use VizQL to build views including small multiple

displays. The selection of the appropriate visualiza-

tion is based on the data properties, such as datatype

(text, date, time, numeric, boolean), data role (mea-

sure or dimension) and data interpretation (discrete or

continuous). Selected visualizations are ranked based

on conditions they met, regardingthe data models cat-

egorical and quantitative. Our method follows a sim-

ilar approach to automate the selection of visualiza-

tions based on data type and data role but, in contrast,

it does not rank selected visualizations in these terms.

Alternatively, we propose a method to rank visualiza-

tions based on user preferences and so generate per-

sonalized visualization recommendations.

Nazemi et al. (Nazemi et al., 2013) describe a sys-

tem which suggests and adapts a set of applicable vi-

sualizations types based on data type and user’s be-

havior. The user behavior is investigated in a canon-

ical user model deﬁned by analyzing of user’s inter-

actions with visualizations. The system includes a set

of seven visualization algorithm and the selection of

the appropriate algorithm is based on the user’s data.

In contrast, we integrate different visualization types,

from graph visualization to geographical visualiza-

tions, using our ontology, and utilize one particular

algorithm to select the appropriate ones. We currently

consider only the data properties to generate the sug-

gestions while Nazemi et al. take a bottom-up ap-

proach, analyzing user interaction with visualization

to describe user behavior. Instead, we propose to in-

vestigate top-down methods to elicit user preferences,

e.g., by rating or collecting items. These methods

are complementary and can be deployed together with

user behaviour analytics.

3 VISUAL RECOMMENDER

Figure 1 shows the recommender workﬂow with em-

VisualRecommendationsforScientificandCulturalContent

257

phasis on the visualization path. Based on a given

user query, a Federated Recommender (FR) compiles

recommendations from a number of associated ser-

vice providers(e.g., Mendeley, Europeana and ZBW).

The query comprises the current interests of the user

(e.g., actual query, requests) as well as a user context

collected by the UI (e.g., visited page, interests). The

FR reverts with a list of items relevant to the query, al-

beit without details as to how items relate to the query

or amonst themselves. In part, visualizations should

help users establish this connection. A ﬁrst step is

to perform data analysis on the recommendation list,

collecting statistics and extracting data attributes. To

automatically suggest the right visualization, the vi-

sual recommender, based on visual perception and vi-

sual encoding guidelines, matches data attributes to

visual components of a visualization.

Each stage depends on one or more data models to

summarize, transform or enrich the recommendation

list and to create alternative (visual) representations

of it. We employ semantic technologies to describe

expressive data models, and build upon them the pro-

cesses to suggest visualizations.

Recommended Items. Associated providers of sci-

entiﬁc, educational and cultural content collect and

index various kinds of documents, such as confer-

ence publications, books, journals, lectures, images,

and events. Each provider deﬁnes and organizes its

repositories according to a (often closed) proprietary

data model. The data model describes attributes that

are used to match user queries. They also recommend

items in terms of the attributes in their data model.

The FR establishes a uniﬁed representation for rec-

ommended items in an extended version of the public

Europeana metamodel. In spite of the rich represen-

tation the metamodel supports, the FR only shares the

minimal relevant attributes common to all involved

repositories. Such minimalistic approach, presents

the client with a list of items and mostly categorical

ﬁelds describing properties and no numerical values.

RDF Data Cube Vocabulary. A recommended item

describes a single resource according to the plain data

model described previously. A recommendation list

is nothing else than a sequence (possibly ordered by

ranking) of such items. To ﬁgure out how to graph-

ically represent the recommendation list and items

thereof, we need a model that describes the properties

of the data they contain that we can later match to ap-

propriate visual components. Hereby, a recommenda-

tion list needs to be represented as an n-dimensional

data cube, identifying data dimensions and their se-

mantics. For this purpose, we chose the RDF Data

Cube Vocabulary (RDF-DCV

) semantic web stan-

RDF Data Cube Vocabulary: www.w3.org/TR/vocab-data-cube/

dard, developed by the W3C to deﬁne concrete data

models for arbitrary measurements (e.g., statistical

data). The RDF-DCV deﬁnes a collection of so called

observations, each consisting of a set of dimensions

and measures. Dimensions identify the observations

as categorical data, whereas measures are related to

concrete values. Both types of components are de-

ﬁned as generic elements of RDF so that complex

structures can be constructed out of primitive RDF

data types. A preprocessing stage analyses the rec-

ommendation list and generates a description of data

attributes in RDF-DCV.

Visual Analytics Vocabulary. We developed a Vi-

sual Analytics Vocabulary (Mutlu et al., 2014) to se-

mantically describe visualizations. We used RDF to

this end because: (1) it provides a common persis-

tence model for representing visualizations that can

be used by various visualization technologies, and

(2) it allows to query existing visualizations enriched

with data. The model strictly focuses on describing

the visual encoding process. Hence we represent vi-

sualizations in terms of their visual channels (Bertin,

1983). However, instead of pursuing a thorough spec-

iﬁcation encompassing all known facts about visual

perception (as in (Voigt et al., 2012)), we concen-

trate on pragmatic, simple facts that will aid the sen-

sible mapping (e.g. (Mackinlay, 1986)), extending

the description to many different types of visualiza-

tions. The VA Vocabulary consists of two parts: (1)

the model of an abstract visualization capturing com-

monalities shared between all concrete visualizations,

and (2) the model of a concrete visualization captur-

ing speciﬁc information. The abstract visualization

model speciﬁes structural components that any con-

crete visualization may have, such as: (a) name (b)

visual channels (encoding some attributes of the data,

e.g axes of a visualization), (c) description. Concrete

visualizations reﬁne the abstract visualization model

depending on their type by reiﬁcation of the visual

channels. Hence, visual channels are characterized

by: (i) datatype: set of primitive datatypes that a vi-

sual channel supports, (ii) occurrence: cardinality,

i.e. how many instances are allowed for the visual

channel, (3) persistence: whether a visual channel is

mandatory for the concrete visualization.

3.1 Suggesting Visualizations

There are two stages to generically suggesting visual-

izations for a recommendation list: extraction of data

attributes (Block 2 in Figure 1), and the actual map-

ping of those attributes to visual channels of a visual-

ization (Block 3 in Figure 1).

Extraction of data attributes involves expressing

IVAPP2015-InternationalConferenceonInformationVisualizationTheoryandApplications

258

recommended items in a form of quantitatively ob-

servable data. To do so, recommended items are col-

lected and organized according to the attributes they

expose. Some data attributes can be inferred, e.g.

from a metamodel. For others, we build a distribu-

tion of items over attributes, i.e., data attributes and

items are put in relation so as to derive the various

options to express the recommendation space graph-

ically. For example, we count recommended items

for a particular attribute such as provider (Mendeley,

Econbiz, etc.), or determine if a ﬁeld is a category

and how many dimensions it has. This new repre-

sentation, meets two important objectives. First, the

aforementioneddistributioncan be presented visually.

Second, its attributes can be used as a means to an-

alyze results – to ﬁlter or rank items. We use the

RDF-DCV to represent all these elements: measures

express the quantity of items, or quantity of attribute

values. Dimensions represent categorical values only

(strings, numbers, date, etc.).

The next stage consists of mapping dimensions

and measures of RDF-DCV observations to the visual

channels of visualizations, in semantic terms, a rela-

tion from dimensions and measures to visual chan-

nels. However, there are caveats to building such a

relation. For example, dimensions and measures are

strongly typed in the RDF-DCV. Thus, their datatypes

must comply with those of the related visual chan-

nels. Also instances of RDF-DCV may consist of

arbitrary number of dimensions (recommended items

could have arbitrary number of ﬁelds). But its struc-

ture must comply with the structure of the visualiza-

tion. A prerequisite that is complicated because some

visualizations have optional visual channels, so that

many mapping combinations are possible. To iden-

tify appropriate visualizations for a semantically en-

riched recommendation list, our mapping algorithm

inspects the vocabularies describing data and visual-

izations according to two criteria: (a) data types of

involved dimensions, measures and visual channels,

and (b) the structure of instances in both vocabular-

ies (the concrete recommended items, and concrete

visualizations such as time visualization, bar chart,

etc.). The semantic representation of both recom-

mended items and visualizations has the beneﬁt that

powerful queries can be easily built to ﬁnd particular

instances and compare their semantics.

The result of the mapping is a list of visualiza-

tions including all possible mapping combinations for

their visual channels that have satisﬁed both criteria

above. Visualizations are then reiﬁed in the user in-

terface for visual analysis of recommendations (e.g.

using Google Charts, D3).

3.2 Usage Scenario

Consider the following scenario: Jane studies an ar-

ticle in Wikipedia about women in the workforce.

When highlighting a sentence in the article the re-

comender delivers relevant documents represented in

a list. The list contains documents of different kinds

(scientiﬁc papers, pictures, articles etc.). To navigate

the results and determine the relevant content for her

study, the user starts the visual recommender(see Fig-

ure 2).

Jane wants to see scientiﬁc publications from the

last three years written in English. She selects a time-

line, which allows her to ﬁlter recommendations by

creation year, provider, and language. The semantic

description of the timeline has three visual channels,

x-axis

y-axis

, and

color

. Data analysis identi-

ﬁed that both language and provider are categorical.

Hence, the mapping algorithm identiﬁes the follow-

ing mapping combinations: year to

x-axis

, language

and provider to either

y-axis

color

. Choosing the

ﬁrst mapping, the

y-axis

is divided in two for {en,

de}, and color encodes the discrete set {mendeley,

ZBW, Europeana}. The mapping can be used to high-

light results: clicking on a provider now dims all non-

related items both in the visualization and in the list.

To see which provider has more documents on the

topic in the last three years, she swaps the

y-axis

provider, the visual recommender adjusts accordingly

mapping language to color. The visualization clearly

shows the contributions of provider each year. Jane

can now highlight items by language clicking on the

corresponding color label.

4 USER PREFERENCES

As described up-to this stage, the visual recommender

suggests visualizations solely on general encoding

guidelines for data attributes. But, often users would

rather use a particular chart to analyze a certain as-

pect of the data. The question in general is how to

accommodate for these kinds of user preferences. In

this context we investigate two top-down approaches,

designed to apply collaborative ﬁltering to reﬁne the

suggestion of charts coming from the mapping algo-

rithm and also to enrich the recommendation list.

As the mapping algorithm produces a list of us-

able visualizations and for each a list of possible map-

pings, we apply collaborative ﬁltering to reﬁne those

lists giving a rating to the mappings and suggesting

ﬁrst the highly ranked visualizations. The algorighm

to do so requires ratings from users for the visual-

izations. As the literature on recommender systems

VisualRecommendationsforScientificandCulturalContent

259

Figure 2: Visual recommender tool: exemplary visualization and mapping suggestions for recommendations.

points out, ratings can be multidimensional. We have

designed a rating based on scales pertaining the per-

ceived usefulness and also visual organization of a

visualization. The usefulness of a visualization de-

pends, of course, on the task and information needs

of the user. The challenge is to elicit preferences on

these terms: which visualization is useful for particu-

lar information needs. To do so, we are investigating

through a crowdsourced study, what information peo-

ple extract from visualizations and their rating. The

future version of the visual recommender will inte-

grate ratings and preferences elicited from the study

as an optimization measure.

The second top-down approach to elicit prefer-

ences is to analyse the items that users collect. One

feature of the proposed visual recommender is that it

allows users to create collections of items. A collec-

tion consists of a topic title and a list of items and

an optional description. We are currently investigat-

ing methods to rank the items in a collection using

lightweight information extraction methods to match

keywords to the topic or description. The ranking

of items as well as their belonging to particular col-

lections will be used to enrich the recommendation

list. To reﬁne the visual recommendation, we in-

tend to store meaningful visualizations in collections,

thus complementing the preferences for a visualiza-

tion with collections associated.

5 DISCUSSION

One strength of the proposed method is that the visual

channels are used to directly inﬂuence the interaction.

The front-end deﬁnes an interactivesystem based on a

model-view controller API, which allows connecting

the output of the interaction with charts to different

views on the data. For example, when ﬁltering items

in a chart, a list is automatically updated (dimming

unselected items). Similarly, the list can be used to

deﬁne the focus and context in the chart.

The workﬂow and use-case involves several

choices of charts to explore the data. One shortcom-

ing of the current approach is that chart suggestions

are based on the mappings which, although concep-

tually correct, at times show junk charts. We found

that this is often due to the generalization leading to

poor information about data attributes (e.g.,what are

the intervals and distributions). We are investigating

optimizations based on user preferences, for example

following methods in Section 4. The methods for elic-

itation of preferences require a critical mass of users

and ratings for visualizations and items. Our cur-

rent work investigates methods to acquire these initial

preferences.

6 CONCLUSION

This paper introduces a visual recommender tool that

suggests visualizations for recommendations in the

scientiﬁc, cultural and educational domain. The main

power of these generic visualizations is that only

those that can actually represent the data are sug-

gested. If the items in the recommendation list do

not contain dates, timelines cannot be reiﬁed and will

not be suggested. The current implementation of the

visual recommender tools support nine conventional

charts. Our tool has been deployed in the frame of

EEXCESS, but it is not constrained to the cultural

domain. It can actually suggest visualizations for

any recommendation list expressed in the suggested

format. This is because we deﬁned semantically-

enriched data models for visualizations and recom-

mendations. Thereby, powerful queries quickly ex-

plore a semantic space consisting of a huge number

of recommendations and link them to visualizations.

The current visualization suggestion consists of a

list of different visualizations with a lot of possible

mapping combinations. Our future goal is to explore

IVAPP2015-InternationalConferenceonInformationVisualizationTheoryandApplications

260

ways to narrow down the choices relying on user be-

havior, context building towards content-based rec-

ommendation of visualizations. We have brieﬂy de-

scribed two methods that are currently under investi-

gation to elicit user preferences and context for visu-

alizations and that will help suggest meaningful visu-

alizations for the information needs of the user.

ACKNOWLEDGEMENTS

This work is funded by the EC 7th Framework

project EEXCESS (grant 600601). The Know-Center

GmbH is funded within the Austrian COMET Pro-

gram - managed by the Austrian Research Promotion

Agency (FFG).

REFERENCES

Bertin, J. (1983). Semiology of graphics. University of Wis-

consin Press.

Granitzer, M., Seifert, C., Russegger, S., and Tochtermann,

K. (2013). Unfolding cultural, educational and sci-

entiﬁc long-tail content in the web. In UMAP Work-

shops.

Kagie, M., van Wezel, M. C., and Groenen, P. J. F. (2011).

Map based visualization of product catalogs. In Ricci,

F., Rokach, L., Shapira, B., and Kantor, P. B., ed-

itors, Recommender Systems Handbook, pages 547–

576. Springer.

Konstan, J. A. and Riedl, J. (2012). Recommender systems:

From algorithms to user experience. User Modeling

and User-Adapted Interaction, 22(1-2):101–123.

Lex, A., Streit, M., Schulz, H., Partl, C., Schmalstieg, D.,

Park, P. J., and Gehlenborg, N. (2012). StratomeX: vi-

sual analysis of Large-Scale heterogeneous genomics

data for cancer subtype characterization. Computer

Graphics Forum (EuroVis ’12), 31(3):1175–1184.

Mackinlay, J. (1986). Automating the design of graphical

presentations of relational information. ACM Trans.

Graph., 5(2):110–141.

Mackinlay, J., Hanrahan, P., and Stolte, C. (2007). Show

me: Automatic presentation for visual analysis. IEEE

Transactions on Visualization and Computer Graph-

ics, 13(6):1137–1144.

Munzner, T. (2011). Applying information visualization

principles to biological network displays. volume

7865, pages 78650D–78650D–13.

Mutlu, B., Hoeﬂer, P., Tschinkel, G., Veas, E., Sabol, V.,

Stegmaier, F., and Granitzer, M. (2014). Suggesting

visualizations for published data. In Proceedings of

the 5th International Conference on Information Visu-

alization Theory and Applications, IVAPP.

Nazemi, K., Retz, R., Bernard, J., Kohlhammer, J., and

Fellner, D. (2013). Adaptive semantic visualization

for bibliographic entries. In Bebis, G., Boyle, R.,

Parvin, B., Koracin, D., Li, B., Porikli, F., Zordan,

V., Klosowski, J., Coquillart, S., Luo, X., Chen, M.,

and Gotz, D., editors, Advances in Visual Computing,

volume 8034 of Lecture Notes in Computer Science,

pages 13–24. Springer Berlin Heidelberg.

Ricci, F., Rokach, L., Shapira, B., and Kantor, P. B.

(2011). Introduction to Recommender Systems Hand-

book, pages 1–35. Springer US.

Schl¨otterer, J., Seifert, C., and Granitzer, M. (2014). Web-

based just-in-time retrieval for cultural content. In

PATCH ’14: Proceedings of the 7th International

ACM Workshop on Personalized Access to Cultural

Heritage.

Swearingen, K. and Sinha, R. (2001). Beyond algorithms:

An hci perspective on recommender systems. In ACM

SIGIR. Workshop on Recommender Systems, volume

Vol. 13, Numbers 5-6, pages 393–408.

Voigt, M. et al. (2012). Context-aware recommendation of

visulization components. In The Fourth International

Conference on Information, Process, and Knowldege

Management.

VisualRecommendationsforScientificandCulturalContent

261