MUVTIME: A Multivariate Time Series Visualizer for

Behavioral Science

Emanuel Sousa

, Tiago Malheiro

, Estela Bicho

, Wolfram Erlhagen

, Jorge Santos

and Alfredo Pereira

Algoritmi Centre, University of Minho, Guimarães, Portugal

Center of Mathematics, Department of Mathematics and Applications, University of Minho, Guimarães, Portugal

CiPSI, School of Psychology, University of Minho, Braga, Portugal

Keywords: Multivariate Time Series, Visualization, Cognition.

Abstract: As behavioral science becomes progressively more data driven, the need is increasing for appropriate tools

for visual exploration and analysis of large datasets, often formed by multivariate time series. This paper

describes MUVTIME, a multimodal time series visualization tool, developed in Matlab that allows a user to

load a time series collection (a multivariate time series dataset) and an associated video. The user can plot

several time series on MUVTIME and use one of them to do brushing on the displayed data, i.e. select a

time range dynamically and have it updated on the display. The tool also features a categorical visualization

of two binary time series that works as a high-level descriptor of the coordination between two interacting

partners. The paper reports the successful use of MUVTIME under the scope of project TURNTAKE,

which was intended to contribute to the improvement of human-robot interaction systems by studying turn-

taking dynamics (role interchange) in parent-child dyads during joint action.

1 INTRODUCTION

Like many other research fields, behavioral science

is becoming increasingly more data intensive.

Behaviors can now be captured, analyzed, and

quantified semi-automatically and this is leading to a

change in the way psychological science is made

(Jaffe, 2014). A few trends are evident in the move

towards “big data” in psychological research. One

derives from the unexpected consequences of having

previously collected data made available in the

public domain for others to freely reuse and

reanalyze. This kind of aggregation has already

occurred for particular types of datasets and opened

the way for synthetic modeling approaches to future

studies – e.g. the NeuroSynth and NeuroVault

projects combine functional magnetic resonance

imaging (fMRI) data from multiple sources

(Gorgolewski et al., 2015; Yarkoni, 2012). Another

line of development is related to automatic

sampling, sometimes with residual human

investment in data acquisition, for instance

biological data obtained from fitness trackers, used

in studies of sleeping habits and how they interfere

in people’s general mood (Swan, 2013). Finally, a

type of studies with a long tradition in psychological

research, micro-analytic studies of social interaction

– highly informative but notoriously challenging and

labor-intensive, and thus limited in number,

(Bakeman and Gottman, 1997; Burgoon et al., 2007)

– are now at least more tractable. The term micro-

analytic here refers to situations where social

interaction is quantified, potentially across more

than one modality, on a time-scale of milliseconds.

Examples of this trend include the appearance of the

field of multimodal corpora studies (Kipp et al.,

2009), the first case of an almost exhaustive

sampling of a single child’s entire language input,

conducted in the first two years (Roy et al., 2015),

and novel developmental psychology studies of

infant’s sensorimotor dynamics and its effect on

learning, during mother-infant social play (Smith et

al., 2011; Yu and Smith, 2013; Pereira et al., 2014b).

In this type of studies, data is analyzed so that

complex relations between the variables coding

different dimensions of human activity (e.g. gaze

direction or body posture) can be determined.

Consider the case of a mother and an infant

Sousa, E., Malheiro, T., Bicho, E., Erlhagen, W., Santos, J. and Pereira, A.

MUVTIME: A Multivariate Time Series Visualizer for Behavioral Science.

DOI: 10.5220/0005725301650176

In Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2016) - Volume 2: IVAPP, pages 167-178

ISBN: 978-989-758-175-5

167

engaged in a joint task: Different dimensions of the

mother’s socially contingent activity, such as gaze

direction, body posture, speed of motions, and

frequency of vocalizations, among others, will affect

similar dimensions of the infant’s behavior and, in

turn, also be affected by the infant’s behavior in a

well-coordinated social exchange (De Barbaro et al.,

2013).

This recent trend of growing behavioral datasets

bears a great potential for information extraction.

This potential, however, is often limited by the tools

available to behavioral scientists. Often, the sheer

amount of data requires specific data visualization

tools, not only because of the challenges introduced

by dataset size but also by the very nature of the data

itself. For instance, time series resulting from motion

capture can contain both a macro and a micro-

structure: e.g. average velocity vs. velocity

variations produced by adapting in real-time to a

social partner. Typically, the dataset is a multivariate

time series and is difficult to visually explore and

identify associations between variables. This is

especially the case when one intends to interact with

non-stationary time series data that rapidly changes

depending on the time region of interest or when we

need contrasting visualizations.

In order to assist with some of these data analysis

tasks, we developed MUVTIME (MUltivariate

Visualization of TIME Series), an interactive tool

for multivariate time series visualization for social

interaction studies, intended to facilitate the process

of interleaving visualization and numerical analysis.

MUVTIME was developed in the context of project

TURNTAKE (Lisboa et al., 2014; Pereira et al.,

2014a), a project aimed at improving Human-Robot

Interaction design by studying developmentally

parent-child turn-taking dynamics with the ultimate

objective of improving robots’ ability to adjust to

individual rhythms in interaction, a factor known to

influence the quality of a social interaction (Jaffe et

al., 2001). Data analysis was both hypothesis-driven

– motivated by studies of mother-infant attachment

style (Jaffe et al., 2001) – and exploratory – the

relationship between vocal and motor coordination

is unclear.

Our main goal was applying data visualization

techniques to the specific question of understanding

real-time social coordination as it happens in

mother-child interactions. Functional requirements

and visualization design rationale for the tool

derived from the research questions, our previous

experience in using behavior coding and annotation

tools (the authors include developers and end-users),

and the specific requirements of visualizing turn-

taking dynamics. This produced a minimum set of

functional requirements we selected for

development.

MUVTIME implements a time brushing tool

allowing a user to focus on specific time periods,

analyzing the data with different visualizations. It

also integrates a video-playing tool where current

time is marked in a time cursor and data views are

all linked by time. This is particularly critical in

behavioral research since it allows immediately

contrasting the actual data of an interaction episode

with the judgment made by the researchers when

visualizing behavioral events. The data visualization

techniques include a minimum core set of time

series visualizations plus a novel visualization of

turn-taking, a high-level categorical visualization,

automatically extracted from the data, and that

works as a visual descriptor of the coordination

between two interacting partners, following the work

of (Jaffe et al., 2001).

The paper describes the MUVTIME most

important features and then reports and discusses its

utilization on the TURNTAKE project.

2 RELATED WORK

In this section we review some computer

applications for interactive exploration of time

oriented data. A more comprehensive survey can be

found in Aigner et al., (2011)

One domain with an intensive use of this tool is

medical care and research where it is used for

medical diagnosis, treatment management, and data

exploration. (Catarci et al., 2003) presented one of

the first interfaces for dealing with biomedical data

that also included some limited capabilities of time

series visualization. A more time oriented graphic

tool was proposed by Bade et al., (2004) for

comparing data of patients in treatment with

corresponding medical guidelines It featured

methods for displaying qualitative and quantitative

temporal information in the same graph. Another

project, the CareGiver (Brodbeck et al., 2005),

displayed both categorical and numerical data in

different bands while maintaining a unique

timeframe.

Other tools, more focused on data exploration,

were KNAVE-II (Shahar et al., 2006) which allowed

to visualize and explore both raw data and temporal

abstractions derived from it (e.g. episodes of high

blood pressure). PatternFinder (Fails et al., 2006)

featured a graphical interface for performing queries

(formulated as sequences of events) on a database of

IVAPP 2016 - International Conference on Information Visualization Theory and Applications

168

patient histories and produce visualizations.

Lifelines 2 (Wang et al., 2009) allows temporal

alignment of event-based time series data of

different patients according to the occurrence of a

specific event (e.g. the application of a therapy) and

compare posterior events. Similan

(Wongsuphasawat and Shneiderman, 2009), was

designed to query and visualize medical records,

when searching for specific sequences of categorical

events. CareCruiser (Gschwandtner et al., 2011)

organizes multiple treatment plan algorithms in a

tree like or flow chart view. The user can then chose

one plan and visualize the history of the patients that

received that particular line of treatment.

VisuExplore (Rind et al., 2011) uses some of the

techniques from Bade et al., (2004) but more

specifically turned to data exploration.

More general tools for interactive explorations of

time series data also exist. The multiple iterations of

TimeSearcher (Buono et al., 2007; Hochheiser and

Shneiderman, 2004) were built around the concept

of time box, a graphical object that a user can draw

with the mouse on a two-dimensional plot of time

series data (where the horizontal and vertical axis

represent time and the measured dimension,

respectively). This acts as a filter by defining a

period of interest (the time interval that fits in the

box) and a range of values (values that range

between the vertical limits of the box). The result of

the filter is the set of data items whose values fit

inside those ranges. GeoTime (Kapler and Wright,

2005) is a 3D visualization tool for space-time data

exploration where a ground plane represents spatial

information while variation in time is displayed in

the third dimension. VIS-STAMP (Guo et al., 2006)

also allows exploration of geo-temporal data using

self-organizing maps (SOM) for data clustering.

EventViewer (Beard et al., 2007) is a framework for

visualization of sensor-based data acquired on

multiple locations during long time spans.

Exploration of both temporal and spatial patterns is

supported. FacetZoom (Dachselt et al., 2008) allows

a user to navigate time oriented data using different

time granularities.

In the behavioral science domain, the analysis of

categorical time series is also common. These time

series are typically obtained through annotations of

events occurred in observation sessions recorded on

video and normally performed by trained researchers

in an exploratory fashion or using structured coding

schemes. One highly influential computer tool,

MacSHAPA (Sanderson et al., 1994), was developed

for fast annotation of these recordings and geared

towards exploratory sequential data analysis

(ESDA). Annotations were entered as values within

the cells of a spreadsheet, where the categorical

variables were represented as columns. It included

some graphic visualization capabilities, but most

importantly it offered digital frame-by-frame video

playback in VCR tape recordings. As computers

evolved, similar tools, free and commercial, became

available. Some are concentrated on speech analysis,

for example ELAN (Wittenburg et al., 2006) and

EXMARaLDA (Schmidt and Wörner, 2009) or

multimodal corpora like ANVIL (Kipp, 2012),

TASX-annotator (Milde et al., 2001) or MacVisTa

(Young and Bann, 1996). Recently, Yu and

colleagues (Yu et al., 2012; Yu et al., 2009)

proposed a tool for interactive exploration of time

series with video playback. The tool includes several

timeline visualizations methods for visual

exploration of both numerical and categorical time

series. Two other tools, ChronoViz (Fouse et al.,

2011) and BEDA (Kim et al., 2013) were also

developed for visualization and annotation of time

series data with video playback. Notables (Lee et al.,

2013) is an online platform featuring a visualization

called “plexlines” where categorical events occurred

during an interaction session between clinicians and

children are displayed as circles in a timeline.

3 MUVTIME

MUVTIME was designed to assist on the process of

multivariate time series data analysis, both for direct

data exploration and in the process of prototyping

processing algorithms. It was implemented in

Matlab because this computational platform is our

primary tool for complex analysis. Thus, we can

apply processing methods on multivariate time

series and rapidly visualizing them without having to

export the data to csv or other format and import it

on other tool such as TimeSearcher or ChronoViz.

MUVTIME is built around two Matlab objects

designed to work with time series data: timeseries

and a collection of timeseries (tscollection).

Timeseries objects contain a univariate time vector

and a multivariate data vector. They also contain

metadata, e.g.: units, starting and ending time, data

quality measures, as well as a structure for events

description. The tscollection is formed by a set of

timeseries objects that share a common time vector

but may refer to different data types. When

performing time series processing in Matlab it is

useful to use these objects since they offer a set of

standard methods for manipulation (e.g.

interpolation, concatenations) that ensure data

MUVTIME: A Multivariate Time Series Visualizer for Behavioral Science

169

Figure 1: Depiction of MUVTime main interface. a) Timeline navigator. The yellow panel defines the time period under

visualization. b) Visualization panels. Each panel displays one or more time series with a specific visualization method. c)

Video player and control. Video time is represented by the vertical blue line on the visualization panels. d) Point light

display. 3D visualization of the motion capture data represented as points in a 3D space.

consistency and they also allow basic feature

extraction like averages and standard deviation.

MUVTIME can be called from the Matlab command

prompt directly with a tscollection object as input

argument. It can also be called with no input

arguments and a collection stored on a “.mat” file

can be loaded through the menu options. A depiction

of the application’s main interface is presented in

Figure 1, where its main elements are visible: The

timeline navigator, the visualization panels, the

video window and control and the Point light

walking display. Next we describe each of the

elements in detail.

3.1 Main Interface

3.1.1 Timeline Navigator

The timeline navigator controls the time window of

the data currently in visualization in the band graphs

(Figure 1.a). Graphically, it consists of a standard

time series line plot (with data referring to one of the

time series of the collection), where time runs

horizontally from left to right. An overlaid semi-

opaque yellow panel defines the time period under

visualization. Initially, when a collection is loaded,

the time window covers the complete time series but

its width can be adjusted with the computer mouse

by click-and-drag of the cursors on the right and left

bottom corners of the window. The complete

window can also be displaced through click-and-

drag of the yellow region. The time series plotted on

the background (the user can control which one it is)

serves as cue for the adjustment of the brushing

window on a particular time region (Yu et al., 2012).

3.1.2 Visualization Panels

The visualization panels (Figure 1.b) are the main

component of MUVTIME where the data is

visualized. The total number of panels can be

defined on the edit box on top of the visualizer.

Within each panel, the visualization can be changed

by clicking on the “Edit button” (top left corner of

each band). This opens an options window where a

particular visualization can be selected and its

parameters adjusted. The time interval visualized is

defined by the sliding panel on the timeline

navigator. When this time panel is adjusted the

visualization panels are automatically updated so

that the time window under visualization is the same

on all of them. The graphs are always kept aligned

so that each horizontal coordinate is vertically

aligned with the same coordinate in the other graphs.

This allows visual comparison of different

IVAPP 2016 - International Conference on Information Visualization Theory and Applications

170

visualizations, and different time series, for the same

time region.

3.1.3 Video Player and Control

MUVTIME can optionally load and play a video file

on a separate visualization window on the screen.

The underlying video player is the Microsoft

windows media player that is part of the free set of

tools available for Microsoft windows. This

“outsourcing” of the video playing decreases the

computational effort for Matlab, allowing it to run

fluidly even with the video playing. The player can

be controlled either through a standard video

interface (Figure 1.c) or through keyboard shortcuts.

Two different reproduction modes are available. In

the first one, the video loops continuously in the

time window of visualization. In the second the

video controller automatically shifts the time

window, running the video until its end. A vertical

blue line is always drawn on both the time brushing

panel and in the visualization panes, marking the

current video time. A double click on any location of

a visualization panel will make the video current

time jump to the instant defined by horizontal

position of the mouse cursor when clicked.

3.1.4 Point Light Display

The point light display tool (Figure 1.d) is a

visualization window that can display motion

capture data represented as point lights if the time

series collection includes time series of 3D data

point. Point lights refers to using a small sphere to

represent a motion capture marker and is a type of

motion visualization originally derived from studies

of biological motion perception, in particular point

light walkers (Johansson, 1973; Johnson and

Shiffrar, 2013). In MUVTime it allows contrasting

live video with motion data, mostly for controlling

data reliability. The tool includes a menu for

choosing which points the user wants to visualize

(MUVTIME will automatically search time series

with three dimensional data). Point motion is

synchronized with the current time.

3.2 Data Visualizations

For each visualization panel, the user can decide

which data and type of visualization to use. Next we

describe the four types of visualization available.

3.2.1 Overlapping Line and Categorical

Graphs

Line graphs are probably the most common type of

visualization for time series and are also included in

MUVTIME. The interface allows multiple time

series to be plotted as lines in a single visualization

panel (figure 2a). For each line that is appended to

the visualization a new vertical axis, colored the

same as the corresponding line, is added on the left

side of the pane. The scale on the axis is

automatically set to fit the range of the data, within

the time range of the current time interval of

visualization. With this visualization of multiple

axes, data of different dimensions can be compared

in terms of their trends (e.g. one rises when other

falls). The color of each line and corresponding axis

is automatically set when the line is added but can

be altered by the user. When the mouse cursor

hovers the graph, a vertical line appears also

displaying the values of the variables on the instant

defined by the mouse horizontal position on the

graph.

Overlapped on the line plots, the user can add a

categorical visualization that will be displayed as a

set of shaded areas. This allows inspecting the

relations between numerical and categorical

variables. MUVTIME allows that any time series

can be defined as categorical, as long as its values

are all integers. By default, the color map for the

categorical time series is automatically defined

depending on the total number of categories.

MUVTIME can generate two types of color maps:

one defines the colors in order to facilitate visual

differentiation, maximizing their distance in the

CIELAB color space (Schanda, 2007); the other

assumes that the categorical space is scaled. Thus

the color map is defined as a heat map of the

categories, based on their numerical value. The user

can also define a color map by loading .xls file

where each line is formed by four numbers, the first

one referring to the category numerical identifier and

the other ones describing the colors in the RGB

space.

The choice of using colors to represent

categories raises the question of scalability

considering that color differentiation becomes

increasingly difficult as the number of colors

increases. However, considering the application

domain it is unlikely that a large number of

categories will be required for same plot.

MUVTIME: A Multivariate Time Series Visualizer for Behavioral Science

171

Figure 2: Available visualization types: (a) Overlapping line and categorical graphs. (b) Stacked categorical map. (c)

Horizon graph. (d) AVTA diagram.

3.2.2 Stacked Categorical Maps

This type of visualization (figure 2b) allows the

comparison of multiple categorical time series as a

set of stacked colored bands. Like in the previously

described visualization, any time series whose

values are exclusively integers can be chosen as

categorical and displayed in this form. The color

map can also be automatically defined to maximize

differentiation or to act as categorical heat map.

3.2.3 Horizon Graphs

Horizon graphs (figure 2c) (Few, 2008; Saito et al.,

2005) are a type of plot idealized to minimize the

space necessary to display a time series. They are

based on area plots but employ a few techniques to

minimize space while facilitating perception. First of

all the negative part of the graph is drawn mirrored

to the x-axis, but with a different color to allow

distinction. Second, the vertical direction of the area

plot is divided into bands where the color tone is

darker as the values fit into increasingly higher

ranges. Thirdly, instead of being stacked, the bands

are overlaid, reducing the necessary space for

display. This type of visualization reduces the height

of the graph compared to a normal line graph by

relation of 1: 



. We choose 



 3 thus

reducing the height 6 times without compromising

the ability of the users to distinguish the bands (Heer

et al., 2009).

3.2.4 AVTA Diagram

The AVTA diagram (figure 2d) is a data abstraction

that can be obtained from two binary time series that

quantify the activity of the two agents by some

measure. It is inspired by the Automated Vocal

Transaction Analysis system (Cassotta et al., 1964;

Jaffe et al., 2001), initially proposed to represent all

possible dyadic states of a speech interaction

between two partners. It defines a behavioral

dimension called a turn and a turn rule, such that

each turn is unambiguously attributed to one of the

IVAPP 2016 - International Conference on Information Visualization Theory and Applications

172

partners (i.e. similar to our informal notion of who is

“holding the floor” in conversation). A turn begins at

the instant that any participant vocalizes alone, and it

is held until the other vocalizes alone, at which point

the turn is exchanged. The turn itself is at a different

conceptual level than the vocal states, as it can be a

composite of all of them. Beyond the turn, the

system also defines states of conversation in terms

of absence/presence of vocalization, duration of

pauses and activity. In the TURNTAKE project, we

generalized this system assuming a broader

definition of activity that also includes body motion

of any kind. While in the initial works the diagram

as used for schematic demonstrations, MUVTime is

able to automatically generate this visualization

based on input data permitting its use for interaction

analysis. It receives as input two time series

representing either vocal activity (e.g. speech

transcriptions) or motion (e.g. some measure of the

physical motion of a tracking marker). The input

signals are binarized to zero or one by thresholding.

The turn attribution rule is similar to the one of Jaffe

et al. (2001). The possible activity states attained by

each of the partners are the following ones:

1. Active (A) – A continuous interval of time where

one individual is the turn holder and is active

alone, containing no period of inactivity greater

than x ms where x is defined by the user.

2. Pause (P) – A joint period of no activity greater

than or equal to x ms bounded by the Active

periods of the turn holder.

3. Switching Pause (SP) – A joint silence greater

or equal to x ms, initiated by the turn holder but

terminated by unilateral activity of the partner,

that gains the turn.

4. Interruptive Simultaneous Activity (ISA) – It

is a period of time that begins with the activity of

the partner that does not hold the turn, while the

turn holder is still active, and ends when the turn

holder stops being active, and at that point the

partner that initiated the interruption gains the

turn.

5. Noninterruptive Simultaneous Activity (NSA)

– It begins with the activity of the partner that

does not hold the turn and ends when the partner

who holds the turn is active continuously.

MUVTIME performs the binarization of the two

time series according with a user-defined threshold.

It then runs the time series to determine the turns

and the dyadic states according with the rules

defined. The diagram itself is formed by two

horizontal axes, each one describing the state of a

partner. The turn holder is marked by a thin

horizontal gray bar drawn on its timeline. The other

states are represented by colored boxes stacked on

the gray bar representing the turn (A - green, P and

SP - white, ISA - yellow, NSA - blue). Vertical blue

dashed lines linking the timelines mark the turn

changes and blue arrows distinguish their direction

of change. While this visualization is not as

compressed as, for instance the stacked categorical

plots, the fact that it displays the activity of the

individuals in two separate streams facilitates the

understanding of who has the turn and the

identification of dyadic states in which the

individuals have simultaneous activity (e.g. NSA).

4 TURN-TAKING DYNAMICS

DURING JOINT PLAY: A CASE

STUDY OF MUVTIME

Next, we summarize a few key findings from using

MUVTIME with the multivariate time series dataset

generated in the TURNTAKE project. A main goal

was to study turn-taking dynamics, when children

are engaged in open-ended joint play with a parent.

Visualization was critical since we needed a tool that

allowed the researcher to replay time series data

(e.g. who is active/inactive in voice or motion

signals) synchronized with the video of the

interaction.

We conducted a set of experimental studies

where mother-child dyads had to engage in a joint

action task that varied in level of difficulty; we

captured each dyad’s: vocalizations, head and wrist

movements using a motion capture system, and

recorded a video of the interaction. Each interaction

section lasted approximately 8 minutes resulting in a

(multimodal) multivariate time series dataset ranging

from 16x10

to 26x10

data samples (movements

were captured at 200Hz).

4.1 Methods

4.1.1 Participants

Ten adult-child dyads participated in the study;

children’s age ranged from 52.1 months to 78.6

months and the adult was always either the child’s

mother or father.

4.1.2 Procedure

Dyads were tested in two tasks: one was more

demanding and it required the parent to teach the

MUVTIME: A Multivariate Time Series Visualizer for Behavioral Science

173

child how to build an object (the object was too

complex for the child alone to succeed) and the rules

forced both partners to engage in the construction; a

second task still included a joint goal – the task was

to build the tallest tower possible with blocks – but

placed no restrictions on the dyad in terms of how

they could achieve the task. Thus, the second task of

constructing a tower was not free play but was

clearly less constrained and demanding than the first

task.

Participants sat across each other on a small table

and wore a sports headband and one wristband on

both wrists. Reflective markers, 14 mm in size, were

then attached to the bands using Velcro placed.

There were four markers in the head and one in each

wrist. A video camera pointed at the table workspace

area recorded video and audio of the interaction.

4.1.3 Data Coding and Processing

Motion data was coded and processed using Vicon

Nexus software package (Vicon, 2015); speech data

was transcribed and time-coded using ELAN

(Wittenburg et al., 2006). Transformation of coded

motion and speech transcription data into a dataset

of suitable for time series analysis involved five

main steps: data reduction of the motion data, using

the MSV (mean-square velocity) approach (Gray et

al., 2005) – this computes a unidimensional signal

from 3D point data; calculation of vocalization

on/off binary time series; registering motion and

speech data; calculation of the AVTA model of turn-

taking for all possible pairs of binary time series

(speech or markers); and exporting datasets to a .mat

file.

The final processed dataset, one per dyad and

task, is a tscollection with a large number of time

series: vocalization on/off per partner; head and

wrists’ MSV per partner; individual AVTA states

per pair of signals (e.g. adult’s vocalization with

child’s vocalization; adult’s binarized head MSV

with child’s binarized head MSV). Each AVTA pair

generates 12 binary time series.

4.2 Data Visualization Findings

Figure 3 shows a set of time series plots produced

with MUVTIME. The plots are from a single time

period of one dyad, while they were engaged in the

most difficult task – building an object by copying it

from an image. This example is representative of the

entire sample; the equivalent plots in the majority of

dyads were similar to this one.

The first two panels contain a visually detailed

version of an AVTA diagram; the plot depicts the

dyadic turn-taking state, labels each one and shows

two running tracks, one per participant, and it is

possible to see who is the turn holder at each

moment. The first panel, a), shows an AVTA

diagram for adult’s vocalization paired with child’s

vocalizations; the second panel, b), shows an AVTA

diagram for adult’s right hand movements paired

with child’s head movements – this is a

characteristic pair since head stabilization is often

associated with paying attention to the social partner

in social play (Yu and Smith, 2013). In panel b), the

diagram was calculated on the fly by specifying a

threshold on the MSV signals (any value above the

threshold counts as partner active) and we used a

conservative value. Both plots immediately show the

coordinated nature of these social exchanges:

interruptions are infrequent and partners smoothly

alternate turns. It is also clear that speech modality

and movements have different time scales.

The information in panels a) and b) can be

compressed and depicted by stacked categorical

maps. Panel c) shows the same information of panel

a) but using this type of visualization and panel d)

shows the same information but with a highly

conservative activity threshold. This comparison

shows full detail of the AVTA diagrams is not

mandatory: panel c) still shows clearly smooth turn-

taking. Panel d) is similar to panel b) but with the

threshold change, any small movement counts as

activity and this affects the turn-taking state (e.g.

there are more interruptions visible for instance).

The motion data still shows evidence of turn-taking

but also shows how the decision of what counts as

making a movement or standing still is problematic

and changes the conclusions on the dynamics of the

interaction.

A third set of panels, e) to g), shows an effort to

further compress the full multidimensional dyadic

state. What is shown are stacked categorical maps of

binary variables namely: speech on/off of each

partner in panel e); head, right and left wrist of

adult’s binarized MSV in panel f); and head, right

and left wrist of child’s binarized MSV in panel f).

Again, the compression shows that for some signals,

the key information is preserved: in panel e) the

smooth, well-coordinated nature of turn-taking when

vocalizing is still evident. Panels f) and g) show

motion data as on/off. Although panel g) shows the

child as constantly moving, this visualization was

still highly informative: first, even with a strict

threshold, the adult has periods of complete stillness,

a signal known to be used by adults in

IVAPP 2016 - International Conference on Information Visualization Theory and Applications

174

Figure 3: Example of multivariate visualization from the same time period of a particular dyad while they were engaged in a

joint construction task. a) AVTA diagram for adult’s vocalization and child’s vocalizations; b) AVTA diagram for adult’s

right hand movements and child’s head movements; c) same information as (a) but as a stacked categorical view; d) same

information as (b) but with a different threshold and with a stacked categorical view; e-g) stacked categorical maps of

binary variables, namely (e) speech on/off of each partner in panel (f) binarization of head, right and left wrist of adult’s

MSV and (g) binarization of head, right and left wrist of child’s MSV; h) adult’s switching pause (categorical) and MSV

value for the child’s head (line); i) reverse of (h).

infant-directed demonstrations of novel actions, e.g.

(Rolf et al., 2009); second the constant activity of

the child calls into the question if this corresponds to

constant activity, natural differences in terms of

motor stability with adults or the threshold decision

itself – an example of visualization assisting

quantitative modeling as a preliminary phase.

Finally, panels h) and i) show an example of

contrasting speech with motion data, in this case an

important specific combination. A critical aspect of

turn-taking is the switching pause (the moment of

silence when the two partners switch) – it is a

marker of coordination strength (Jaffe et al., 2001).

Panel h) shows the adult’s switching pause as a

categorical map and overlaid, the quantitative MSV

value for the child’s head of coordination; panel i)

shows the reverse.

5 CONCLUSIONS

In recent years, behavioral science has followed a

general trend in science and is becoming more data

driven. As a consequence, a greater emphasis is now

placed on tools and processes for data exploration

and visualization of large datasets. This paper

described MUVTIME, an interactive graphical tool

for multivariate time series visualization, developed

in the context of TURNTAKE, a social interaction

research project whose purpose is to study the

interaction dynamics of parent-infant dyadic pairs

and apply the resulting conclusions to Human-Robot

MUVTIME: A Multivariate Time Series Visualizer for Behavioral Science

175

interaction design.

MUVTIME was developed to allow fluent

interaction of the user with the data. The user can

control the time window of visualization, contrast

different visualizations of the same or different time

series, and also compare the video recordings of the

experiments with the data. Also, it features an

abstract visualization named AVTA, automatically

extracted from the data and that works as high-level

descriptor of the interaction between the dyadic

pairs.

During the course of the project we have applied

MUVTIME extensively to visualize the time series

resulting from the interaction studies. The tool

allowed us to make fast visual evaluations of

interaction performance of the dyads, determining

key moments of the interaction and contrasting the

time series with the real action recorded in video.

The overlapping of line and categorical plots was

particularly useful in suggesting correlations

between numerical and categorical variables while

the strictly categorical visualizations (stacked

categorical maps and AVTA) provided insightful

visual footprints of the interactions. The horizon

graph, initially implemented due to its high

information/space ratio has not been particularly

explored in the context of the TURNTAKE analysis

because of our current focus on categorical

variables. Future studies will include examining

quantitative measures of motion like mean square

velocity and in this case the horizon plots might be

particularly useful.

ACKNOWLEDGEMENTS

This research was supported by: Marie Curie

International Incoming Fellowship PIIF-GA-2011-

301155; Portuguese Foundation for Science and

Technology (FCT) Strategic program FCT-

UID/EEA/00066/2013; FCT project PTDC/PSI-

PCO/121494/2010. AFP was also partially funded

by the FCT project (IF/00217/2013).

REFERENCES

Aigner W., Miksch S., Schumann H., et al. (2011)

Visualization of Time-Oriented Data. Human-

Computer Interaction: 69-103.

Bade R., Schlechtweg S. and Miksch S. (2004)

Connecting time-oriented data and information to a

coherent interactive visualization. Proceedings of the

2004 conference on Human factors in computing

systems CHI 04 6: 105-112.

Bakeman R and Gottman JM. (1997) Observing

interaction: An introduction to sequential analysis:

Cambridge university press.

Beard K., Deese H. and Pettigrew N. R. (2007) A

framework for visualization and exploration of events.

Information Visualization 7: 133-151.

Brodbeck D., Gasser R. and Degen M. (2005) Enabling

large-scale telemedical disease management through

interactive visualization. European Notes in Medical

Informatics 1: 1172-1177.

Buono P., Plaisant C., Simeone A., et al. (2007)

Similarity-Based Forecasting with Simultaneous

Previews: A River Plot Interface for Time Series

Forecasting. 2007 11th International Conference

Information Visualization (IV '07). IEEE, 191-196.

Burgoon J. K., Stern L. A. and Dillman L. (2007)

Interpersonal adaptation: Dyadic interaction patterns:

Cambridge University Press.

Cassotta L., Feldstein S. and Jaffe J. (1964) AVTA: a

Device for Automatic Vocal Transation Analysis.

Journal of the experimental analysis of behavior 7:

99-104.

Catarci T., Santucci G. and Silva S. F. (2003) An

Interactive Visual Exploration of Medical Data for

Evaluating Health Centres. Journal of research and

practice in information technology 35: 99-119.

Dachselt R., Frisch M. and Weiland M. (2008)

FacetZoom: A Continuous Multi-Scale Widget for

Navigating Hierarchical Metadata. In: ACM (ed)

Proceedings of the SIGCHI Conference on Human

Factors in Computing Systems. Florence, Italy: ACM

Press, 1353.

De Barbaro K., Johnson C. M., Forster D., et al. (2013)

Methodological considerations for investigating the

microdynamics of social interaction development.

IEEE Transactions on Autonomous Mental

Development 5: 258-270.

Fails J. A., Karlson A., Shahamat L., et al. (2006) A visual

interface for multivariate temporal data: Finding

patterns of events across multiple histories. IEEE

Symposium on Visual Analytics Science and

Technology 2006, VAST 2006 - Proceedings: 167-174.

Few S., (2008) Time on the Horizon. Visual Business

Intelligence Newsletter: 1-7.

Fouse A., Weibel N., Hutchins E., et al. (2011)

ChronoViz: a system for supporting navigation of

time-coded data. CHI'11 Extended Abstracts on

Human Factors in Computing Systems. ACM, 299-

304.

Gorgolewski K. J., Varoquaux G., Rivera G., et al. (2015)

NeuroVault. org: A web-based repository for

collecting and sharing unthresholded statistical maps

of the human brain. Frontiers in neuroinformatics 9.

Gray J., Breazeal C., Berlin M., et al. (2005) Action

parsing and goal inference using self as simulator.

Robot and Human Interactive Communication, 2005.

ROMAN 2005. IEEE International Workshop on.

IEEE, 202-209.

Gschwandtner T., Aigner W., Kaiser K., et al. (2011)

IVAPP 2016 - International Conference on Information Visualization Theory and Applications

176

CareCruiser: Exploring and visualizing plans, events,

and effects interactively. IEEE Pacific Visualization

Symposium 2011, PacificVis 2011 - Proceedings: 43-

50.

Guo D., Chen J., MacEachren A. M., et al. (2006) A

Visualization System for Space-Time and Multivariate

Patterns (VIS-STAMP). IEEE Transactions on

Visualization and Computer Graphics 12: 1461-1474.

Heer J., Kong N and Agrawala M. (2009) Sizing the

horizon: the effects of chart size and layering on the

graphical perception of time series visualizations. CHI

'09Proceedings of the SIGCHI Conference on Human

Factors in Computing Systems, April 4–9, 2009: 1303-

1312.

Hochheiser H. and Shneiderman B. (2004) Dynamic query

tools for time series data sets: Timebox widgets for

interactive exploration. Information Visualization 3: 1-

18.

Jaffe E., (2014) What Big Data Means For Psychological

Science. APS Observer.

Jaffe J., Beebe B., Feldstein S., et al. (2001) Rhythms of

dialogue in infancy: Coordinated timing in

development. Monographs of the society for research

in child development.

Johansson G. (1973) Visual perception of biological

motion and a model for its analysis. Attention,

Perception, & Psychophysics 14: 201-211.

Johnson K. and Shiffrar M. (2013) People watching:

Social, perceptual, and neurophysiological studies of

body perception: Oxford University Press.

Kapler T. and Wright W. (2005) GeoTime information

visualization. Information Visualization 4: 136-146.

Kim J. G., Snodgrass M., Pietrowicz M., et al. (2013)

Visual Analysis of Relationships between Behavioral

and Physiological Sensor Data. Workshop on Visual

Analytics in Healthcare. Washington, USA.

Kipp M. (2012) Multimedia Annotation, Querying, and

Analysis in Anvil. Multimedia Information Extraction:

Advances in Video, Audio, and Imagery Analysis for

Search, Data Mining, Surveillance, and Authoring:

351-367.

Kipp M., Martin J-C, Paggio P, et al. (2009) Multimodal

corpora: from models of natural interaction to systems

and applications: Springer.

Lee J., Lin S and Karahalios K. (2013) Visualizing

Patterns of Social and Communicative Behavior in

Children Using Plexlines. Workshop on Visual

Analytics in Healthcare.

Lisboa I. C., Sousa E., Santos J. A., et al. (2014) Parent-

child vocal coordination during joint construction

tasks: a replication of the AVTA model of turn-taking.

9th Meeting of the Portuguese Association of

Experimental Psychology APPE 2014. Covilhã,

Portugal.

Milde J-T, Milde J-T, Gut U., et al. (2001) The TASX-

environment: an XML-based corpus database for time

aligned language data. Proceedings of the IRCS

Workshop On Linguistic Databases, 11-13 December

2001: 174-180.

Pereira A. F., Lisboa I. C., Sousa E., et al. (2014a) Vocal

and motor coordination in joint-construction tasks

during parent-child social interactions. Workshop

Dynamic Interactions Between Visual Experiences,

Actions and Word Learning, part of the Fourth Joint

Conference on Development and Learning and on

Epigenetic Robotics. Genoa, Italy.

Pereira A. F., Smith L. B. and Yu C. (2014b) A bottom-up

view of toddler word learning. Psychonomic bulletin

& review 21: 178-185.

Rind A., Aigner W, Miksch S, et al. (2011) Visual

Exploration of Time-Oriented Patient Data for Chronic

Diseases: Design Study and Evaluation. Information

Quality in e-Health. Lecture Notes in Computer

Science. 301-320.

Rolf M., Hanheide M and Rohlfing KJ. (2009) Attention

via synchrony: Making use of multimodal cues in

social learning. Autonomous Mental Development,

IEEE Transactions on 1: 55-67.

Roy B. C., Frank M. C., DeCamp P., et al. (2015)

Predicting the birth of a spoken word. Proceedings of

the National Academy of Sciences: 201419773.

Saito T., Miyamura H. N., Yamamoto M., et al. (2005)

Two-tone pseudo coloring: Compact visualization for

one-dimensional data. Proceedings - IEEE Symposium

on Information Visualization, INFO VIS: 173-180.

Sanderson P., Scott J., Johnston T., et al. (1994)

MacSHAPA and the enterprise of exploratory

sequential data analysis (ESDA). International

Journal of Human-Computer Studies. 633-681.

Schanda J. (2007) CIE Colorimetry. Colorimetry.

Hoboken, NJ, USA: John Wiley & Sons, Inc., 25-78.

Schmidt T. and Wörner K. (2009) EXMARaLDA—

Creating, analyzing and sharing spoken language

corpora for pragmatics research. Pragmatics-

Quarterly Publication of the International Pragmatics

Association 19: 565.

Shahar Y., Goren-Bar D., Boaz D., et al. (2006)

Distributed, intelligent, interactive visualization and

exploration of time-oriented clinical data and their

abstractions. Artificial Intelligence in Medicine 38:

115-135.

Smith L. B., Yu C. and Pereira A. F. (2011) Not your

mother’s view: the dynamics of toddler visual

experience. Developmental Science 14: 9-17.

Swan M. (2013) The quantified self: Fundamental

disruption in big data science and biological discovery.

Big Data 1: 85-99.

Vicon. (2015) Nexus 2. Available at:

http://www.vicon.com/products/software/nexus.

Wang T. D, Plaisant C, Shneiderman B, et al. (2009)

Temporal summaries: Supporting temporal categorical

searching, aggregation and comparison. IEEE

Transactions on Visualization and Computer Graphics

15: 1049-1056.

Wittenburg P., Brugman H., Russel A., et al. (2006)

ELAN: a professional framework for multimodality

research. Proceedings of LREC. 5th.

Wongsuphasawat K. and Shneiderman B. (2009) Finding

comparable temporal categorical records: A similarity

measure with an interactive visualization. VAST 09 -

MUVTIME: A Multivariate Time Series Visualizer for Behavioral Science

177

IEEE Symposium on Visual Analytics Science and

Technology, Proceedings: 27-34.

Yarkoni T. (2012) Psychoinformatics new horizons at the

interface of the psychological and computing sciences.

Current Directions in Psychological Science 21: 391-

397.

Young F. W. and Bann C. M. (1996) ViSta: The visual

statistics system. Technical Report 94–1 (c).

Yu C and Smith L. B. (2013) Joint Attention without Gaze

Following: Human Infants and Their Parents

Coordinate Visual Attention to Objects through Eye-

Hand Coordination. PLoS ONE 8.

Yu C., Yurovsky D. and Xu T. L. (2012) Visual data

mining: An exploratory approach to analyzing

temporal patterns of eye movements. Infancy 17: 33-

60.

Yu C., Zhong Y., Smith T., et al., (2009) Visual data

mining of multimedia data for social and behavioral

studies. Information Visualization 8: 56-70.

IVAPP 2016 - International Conference on Information Visualization Theory and Applications

178