=Paper=
{{Paper
|id=Vol-1446/amadl_pap4
|storemode=property
|title=Analyzing Student Action Sequences and Affect While Playing Physics Playground
|pdfUrl=https://ceur-ws.org/Vol-1446/amadl_pap4.pdf
|volume=Vol-1446
|dblpUrl=https://dblp.org/rec/conf/edm/AndresR15
}}
==Analyzing Student Action Sequences and Affect While Playing Physics Playground==
https://ceur-ws.org/Vol-1446/amadl_pap4.pdf
Analyzing Student Action Sequences and Affect
While Playing Physics Playground
Juan Miguel L. Andres1, Ma. Mercedes T. Rodrigo1,
Ryan S. Baker2, Luc Paquette2, Valerie J. Shute3, Matthew Ventura3
1
Ateneo de Manila University, Quezon City, Philippines
2
Teachers College, Columbia University, New York, NY, USA
3
Florida State University, Tallahassee, FL, USA
{mandres, mrodrigo}@ateneo.edu,
baker2@exchange.tc.columbia.edu, luc.paquette@gmail.com,
{vshute, mventura}@fsu.edu
Abstract. Physics Playground is an educational game that supports physics
learning. It accepts multiple solutions to most problems and does not
impose a stepwise progression through the content. Assessing student
performance in an open-ended environment such as this is therefore
challenging. This study investigates the relationships between student
action sequences and affect among students using Physics Playground. The
study identified most frequently traversed student action sequences and
investigated whether these sequences were indicative of either boredom or
confusion. The study found that boredom relates to poor performance
outcomes, and confusion relates to sub-optimal performance, as evidenced
by the significant correlations between the respective affective states, and
the student action sequences.
Keywords: Affect modeling, action sequences, boredom, confusion,
Physics Playground
1 Introduction
Physics Playground (PP) is an educational game that immerses learners in a
choice-rich environment for developing intuitive knowledge about simple
machines. As the environment does not impose a stepwise sequence on the
learner, and because some problems can have multiple solutions, learners have the
freedom to explore, attempt to solve, or abort problems as they wish. The
challenge these types of environments impose on educators is that of assessment.
Within such an open-ended system, how do educators and researchers assess
learning as well as the quality of the learning process?
This study focuses its attention on two main phenomena: student learning and
student affect. Student learning within PP refers to how well a player can
understand the concepts surrounding four simple machines through their efficient
execution in attempting to solve levels, as evidenced by the badges they earn.
Student affect refers to experiences of feelings or emotions. In this study, the
affective states of interest are confusion and boredom, as prior studies have shown
them to relate significantly with learning [4, 10]. Confusion is uncertainty about
what to do next [5]. Confusion is scientifically interesting because it has a positive
and negative dimension, wherein it either spurs learners to exert effort deliberately
and purposefully to resolve cognitive conflict, or leads learners to become
frustrated or bored, and may lead to disengagement from the learning task
altogether [7].
Boredom, on the other hand, is an unpleasant, transient affective state in which
the individual feels a pervasive lack of interest in and difficulty concentrating on
the current activity [8]. Boredom has been a topic of interest because of the
negative effects usually associated with it, such as poor long-term learning
outcomes when students are not provided any scaffolding [10] and its being
characteristic of less successful students [11].
A study conducted by Biswas, Kinnebrew, and Segedy [2] investigated
frequently traversed sequences of student actions using bottom-up, data-driven
sequence mining, the results of which contributed to the development of
performance- and behavior-based learner models. The analyses in this paper seek
to perform similar sequence-mining methods in order to find student sequences
that inform either of the affective states of interest.
This study conducted data-driven sequence-mining analyses to answer the
following research questions:
1. What were the frequently traversed student action sequences among
students playing Physics Playground?
2. Are these action sequences indicative of either boredom or confusion?
The analyses in this study are limited to the data collected during gameplay of
Physics Playground from six data gathering sessions conducted at a public school
in Quezon City in 2013. Data is limited to the interaction logs generated by the
game as well as human observation of affect as logged by two coders trained in
the Baker-Rodrigo-Ocumpaugh Monitoring Protocol [9].
2 Methodology
2.1 Participant Profile
Data were gathered from 60 eighth grade public school students in Quezon City,
Philippines. Students ranged in age from 13 to 16. Of the participants, 31% were
male and 69% were female. As of 2011, the school had 1,976 students,
predominantly Filipino, and 66 teachers. Participants had an average grade on
assignments of B (on a scale from A to F).
2.2 Physics Playground
Physics Playground (PP) is an open-ended learning environment for physics that
was designed to help secondary school students understand qualitative physics.
Qualitative physics is a nonverbal, conceptual understanding of how the physical
world operates [12].
PP has 74 levels that require the player to guide a green ball to a red balloon.
An example level is shown in Fig. 1. The player achieves this goal by drawing
agents (ramps, pendulums, springboards, or levers) or by nudging the ball to the
left or right by clicking on it. The moment the objects are drawn, they behave
according to the law of gravity and Newton’s 3 laws of motion [12].
Fig. 1. Example PP level.
Performance Metrics. Gold and silver badges are awarded to students who
manage to solve a level. A gold badge is given to a student who is able to solve
the level by drawing a number of objects equal to the particular level’s par value
(i.e., the minimum number of objects needed to be drawn to solve the level). A
student who solves a level using more objects will earn a silver badge. A student
earns no badge if he was not able to solve the level. Many levels in PP have
multiple solutions, meaning a player can solve the level using different agents.
2.3 Interaction Logs
During gameplay, PP automatically generates interaction log files. Each level a
student plays creates a corresponding log file, which tracks every event that
occurs as the student interacts with the game. Per level attempt, PP tracks begin
and end times, the agents used, and the badges awarded upon level completion. PP
also logs the Freeform Objects that player draw, or objects that cannot be
classified as any of the four agents. The physics agents within PP are as follows:
– Ramp, any line drawn that helps to guide a ball in motion,
– Lever, an agent that rotates around a fixed point, usually called a fulcrum,
– Pendulum, an agent that directs an impulse tangent to its direction of motion,
– Springboard, an agent that stores elastic potential energy provided by a
falling weight.
2.4 The Observation Protocol
The Baker-Rodrigo-Ocumpaugh Monitoring Protocol (BROMP) is a protocol for
quantitative field observations of student affect and engagement-related behavior,
described in detail in [9]. The affective states observed within Physics Playground
in this study were engaged concentration, confusion, frustration, boredom,
happiness, delight, and curiosity. The affective categories were drawn from [6].
BROMP guides observers in coding affect through different utterances, body
language, and interaction with the software specific to each affective state. A total
of seven affective states were coded, however, this study focuses on three:
concentration, confusion, and boredom. These were identified as follows:
1. Concentration — immersion and focus on the task at hand, leaning toward
the computer and attempting to solve the level, a subset of the flow
experience described in [5].
2. Confusion — scratching his head, repeatedly attempting to solve the same
level, statements such as “I don’t understand?” and “Why didn’t it work?”
3. Boredom — slouching, sitting back and looking around the classroom for
prolonged periods of time, statements such as “Can we do something
else?” and “This is boring!”
Following BROMP, two trained observers observed ten students per session,
coding students in a round-robin manner, in 20-second intervals throughout the
entire observation period of 2 hours. During each 20-second window, both
BROMP observers code the current student’s affect independently. If the student
exhibited two or more distinct states during a 20-second observation window, the
observers only coded the first state. The inter-coder reliability for affect for the
two observers in the study was acceptably high with a Cohen’s Kappa [3] of 0.67.
The typical threshold for certifying a coder in the use of BROMP is 0.6, a
standard previously used in certifying 71 coders in the use of BROMP (e.g., [9]).
The observers recorded their observations using HART, or the Human Affect
Recording Tool. HART is an Android application developed to guide researchers
in conducting quantitative field observations according to BROMP, and facilitate
synchronization of BROMP data with educational software log data.
2.6 Data Collection Process
Before playing PP, students answered a 16-item multiple-choice pretest for 20
minutes. Students then played the game for 2 hours, during which time two
trained observers used BROMP to code student affect and behavior on the HART
application. A total of 4,320 observations were collected (i.e., 36 observations per
participant per each of the two observers). After completing gameplay,
participants answered a 16-item multiple-choice posttest for 20 minutes. The
pretest and posttest were designed to assess knowledge of physics concepts, and
have been used in previous studies involving PP [12].
To investigate how students interacted with PP, the study made use of the
interaction logs recorded during gameplay to analyze student performance. Of the
60 participants, data from 11 students were lost because of faulty data capture and
corrupted log files. Only 49 students had complete observations and logs. As a
result, the analysis in this paper is limited to these students, and the 3,528
remaining affect observations. Engaged concentration was observed 72% of the
time, confusion was observed 8% of the time, and boredom and frustration were
observed 7% of the time. Happiness, delight, and curiosity comprise the remaining
6% of the observation time.
3 Analyses and Results
3.1 Agent Sequences
All PP-generated logs were parsed and filtered to produce a list containing only
the events relevant to the study. Sequences were then separated into one of two
categories: 1) silver sequences, or the sequences that ultimately led to a silver
badge, which comprised 44% of all level attempts, and 2) unsolved sequences, or
the sequences that led to the student quitting the level without finding a solution,
which comprised 39% of all level attempts. Sequences that ended in gold badges
were dropped from the analysis because they only comprised 17% of all level
attempts.
Every time a student earns a badge after solving a level, the badge is awarded
for one of the four agents (e.g., a player is awarded a silver ramp badge for
solving the level using a ramp, and another player is awarded a gold pendulum
badge for solving another level using a pendulum). We tracked the agents the
badges were awarded for per level, and used this list of badges to relabel the
sequences based on correctness. If the level awarded a badge for an agent, that
agent was labeled as correct for that level; if not, the agent was labeled as wrong
for the level. For example, on a level that awarded badges for springboards and
levers, a sequence of Lever+ >+ Ramp+ >+ Springboard+ >+ Level+ End+ (silver9
springboard) would be relabeled as correct+ >+ wrong+ >+ correct+ >+ Level+
End+(silver).
The relabeling was done because most of the sequences were level-dependent,
that is, a majority of some sequences appeared on only one or two levels. By
relabeling based on correctness, we were able to ensure level-independence
among sequences. Sequences were tabulated and their frequencies calculated (i.e.,
how many times each of the 49 students traversed each of the sequences). We
calculated for distribution of sequence frequencies, and the sequences we found to
occur rarely (i.e., less than 30% of the population traversed them) were dropped
from the analysis. We found that the gold sequences occurred rarely, which was
another reason they were dropped from the analysis. The resulting silver and
unsolved sequences can be found in Tables 1 and 2, respectively, along with the
frequency means and standard deviations.
Table 1 lists the top 7 silver sequences within PP, which were traversed by
more than 30% of the study’s population. The Sequences column shows what the
respective sequences look like, and the Frequency column shows the average
number of times the 49 students traversed them and the standard deviations.
Highlighted sequences showed significant correlations with either boredom or
confusion, as discussed further in Section 3.2. Table 2 is presented in the same
manner.
Table 1. Top 7 silver sequences, their traversal frequency means, and standard deviations.
Frequency
Sequences
Mean SD
1 correct>Level+End+(silver)+ 3.53 2.34
2 Level+End+(silver)+ 2.61 2.33
3 wrong>Level+End+(silver)+ 1.90 1.37
4 correct>correct>Level+End+(silver)+ 1.61 1.15
5 wrong>correct>Level+End+(silver)+ 0.90 1.01
6 correct>correct>correct>Level+End+(silver)+ 0.80 1.00
7 wrong>correct>correct>Level+End+(silver)+ 0.69 0.77
The silver sequences in Table 1 show signs of experimentation, with students
playing around with the correct and incorrect agents to solve the levels, as seen in
sequences 5 and 7. Sequences 1, 4, and 6 show students using the correct agents,
but are unable to earn gold badges. This suggests that students, while knowing
which agents to use, do not have a full grasp of the physics concepts surrounding
the agents’ execution. Sequence 3 shows students using wrong objects to solve the
levels. While this may suggest that students are still struggling to understand how
the agents work and which agent would best solve a level given the ball and the
balloon’s positions, this may have also been caused by the PP logger labeling the
objects they drew as freeform objects, and not one of the correct agents.
Sequence 1 shows the students drawing only the correct agent, but are still
unable to earn a gold badge. The sequence-mining algorithm only pulled events
related to drawing any of the four main agents, which are enumerated in Section
2.3. Drawing a lever or a springboard, for example, would require drawing more
than one component. A lever requires the fulcrum, the board, and the object
dropped on the board to project the ball upwards. In order for the agent to work, it
has to be executed correctly (i.e., the board must be long enough, with the fulcrum
in the right position, and the object dropped on the board must be heavy enough to
propel the ball into the air). Sequence 1 may have been caused by students
drawing the correct agent, but improperly executing it. For example, the student
may not have drawn the right-sized weight to drop on the lever, and thus had to
draw another. While drawing another weight to drop on the lever counts towards
the level’s object count, it was not logged as a separate event by the sequence
mining analysis because the player did not draw another agent, only a component
of it. Sequence 2, on the other hand, is suspect because despite the student
drawing no objects to solve a level, he ends up with only a silver badge. This was
most likely caused by the improper logging of the game. The top 7 most
frequently traversed silver sequences account for 58% of the total number of silver
sequences.
Table 2. Top 6 unsolved sequences, their traversal frequency means, and standard
deviations.
Frequency
Sequences
Mean SD
1 Level+End+(none)+ 10.69 8.17
2 wrong>Level+End+(none)+ 1.55 1.65
3 correct>Level+End+(none)+ 1.29 1.50
4 wrong>wrong>Level+End+(none)+ 0.45 0.65
5 correct>correct>Level+End+(none)+ 0.41 0.73
6 wrong>correct>Level+End+(none)+ 0.39 0.57
Table 2, which shows the top 6 unsolved sequences, shows signs of students
giving up. Sequence 1 shows students giving up without even drawing a single
object, which could have been caused by one of two things: 1) the student saw the
level and decided to quit without attempting to solve it, or 2) again, the logger did
not log the objects correctly. This sequence is similar to one of the silver
sequences in that no objects were drawn. What makes them different, however, is
what the sequences ultimately led to. The silver sequences ended in a silver badge,
and the unsolved sequences ended in the student earning no badge. The majority
of the sequences listed in Table 2 show students experimenting mainly with wrong
objects, whether agents or freeform objects. This implies that the students are
lacking in the understanding of how to solve the levels. Sequences 3 and 5 are
interesting because it is unclear whether or not the students understood the
concepts of the agents. That is, students were drawing the correct agents, but
could not get the ball to reach the balloon. Despite drawing one or two correct
agents, the students decided to give up and quit. The top 6 unsolved sequences
account for 81% of the total number of unsolved sequences.
3.2 Relationship with Affect
We computed frequencies for each of the 13 sequences that the 49 students
traversed. Correlations were then run between each of the 13 arrays and the
incidences of confusion and boredom. Because the number of tests introduces the
possibility of false discoveries, Storey’s adjustment [13] was used as a post-hoc
control, which provides a q-value, representing the probability that the finding
was a false discovery. Tables 3 and 4 show the results. Highlights and asterisks (*)
were used on significant findings (q ≤ 0.05).
Table 3 lists the top 7 most frequently traversed silver sequences, from left to
right. The sequences these header numbers represent can be found in Table 1. The
table shows the correlation between each of the top 7 silver sequences using a
metric that represents the percentage of all attempts that match each of the
sequences, the percentage of time the students were observed to be confused (r,
con), and the percentage of time the students were observed to be bored (r, bor).
Table 4 is presented in the same manner, with sequence information in Table 2 for
the top 6 unsolved sequences.
Table 3. Correlations between top 7 silver sequences, confusion, and boredom.
Top 7 silver sequences
1 2 3 4 5 6 7
r, con -0.33 0.23 0.41* 0.03 0.17 0.54* 0.28
r, bor -0.20 -0.17 -0.19 -0.05 0.14 -0.19 -0.20
Table 3 shows two significant positive correlations between confusion and the
silver sequences. The two sequences showed signs of lesser understanding of the
agents. Sequence 3 shows students using only a wrong object to solve a level,
which may have been caused either by incorrect object labeling (e.g., PP logged a
ramp as a Freeform Object), or the student found a different way of solving the
level. Like in most learning environments, players are able to game the system –
or systematically misuse the game’s features to solve a level [1] – within PP
through stacking. Stacking is done when players draw freeform objects to either
prop the ball forward or upward, which may have been the case in sequence 3.
Sequence 6 shows students drawing only correct agents. These sequences having
significant correlations with confusion may imply lesser understanding among
confused students as the they are not only dealing with proper agent execution, but
also with deciding which agent would best solve the level. Despite the challenges
faced by these students, however, they still managed to find a solution to the level.
Our findings suggest that the inability to grasp the physics concepts surrounding
the agents is a sign of confusion.
Table 4. Correlations between top 6 unsolved sequences, confusion, and boredom.
Top 6 unsolved learning sequences
1 2 3 4 5 6
r, con -0.17 0.00 -0.12 -0.01 -0.06 0.04
r, bor -0.12 0.13 0.12 -0.03 0.48* 0.06
Table 4 shows that one of the most frequently traversed unsolved sequences
has a significant positive correlation with boredom. This sequence shows students
using only correct agents, but ultimately deciding to give up. This may have been
caused by the inability to execute the agents correctly, which may imply that,
unlike confused students, bored students were not likely to exert additional effort
to try to solve the level or understand proper agent execution. As mentioned
previously, boredom has been found to have significant relationships with
negative performance outcomes. In this case, sequences all ultimately led to
disengagement: students quitting the level before finding a solution, showing
signs of giving up and lack of understanding of any of the four agents.
4 Conclusions and Future Work
This study sought to identify the most frequently traversed student action
sequences among eighth grade students while interacting with an education game
for physics called Physics Playground. Further, the study sought to investigate
how these sequences may be indicative of affective states, particularly boredom
and confusion, which have been found to significantly affect student learning.
Data-driven sequence mining techniques were conducted to identify most
frequently traversed actions sequences in two categories: the sequences that would
eventually lead the student to a silver badge, and the paths that would eventually
lead the student to not earning a badge.
In the silver sequences, students played around with freeform objects and some
of the four agents in attempting to solve the level. The study found confusion to
correlate significantly with two of the silver sequences, which supports previous
findings regarding the relationship between confusion and in-game achievement,
which suggest that because students are unable to grasp the concepts surrounding
the agents and their executions, students resort to finding other solutions.
In the unsolved sequences, students would give up and quit without finding a
solution, despite already using the correct agents to solve the level. The study
found boredom to correlate significantly with one of the unsolved sequences. This
finding supports the literature that has shown that boredom relates to poor learning
outcomes. This work provides further evidence that boredom and disengagement
from learning go hand-in-hand.
This study provides specific sequences of student actions that are indicative of
the boredom and confusion, which has implications on the design and further
development of Physics Playground. This study also contributes to the literature
by providing empirical support that boredom and confusion are affective states
that influence performance outcomes within open-ended learning environments,
and are thus affective states that learning environments must focus on detecting
and providing remediation to. We found that both bored and confused students
will tend to continuously use correct agents in attempting to solve levels, but
execute them incorrectly. The difference between the two, however, is that
confused students tend to end up solving the level, while bored students give up.
The analyses run in this paper were part of a bigger investigation, and as such,
there are several interesting ways forward in light of our findings. The paper aims
for its findings to contribute to the creation of a tool that can automatically detect
affect given a sequence of student interactions, and provide necessary remediation
in order to curb student experiences of boredom.
Relationship analyses run between student action sequences and incidences of
affect in this paper were done through correlations. However, findings were not
able to determine whether boredom or confusion occurred more frequently during
specific action sequences. We want to find out whether boredom or confusion
occurred before, during, or after the students’ execution of the action sequences,
and in doing so, see whether or not the affective states were causes or effects of
the action sequence executions. We are currently investigating this relationship in
a separate study.
Acknowledgements. We would like to thank the Ateneo Center for Educational
Development, Carmela C. Oracion, Christopher Ryan Adalem, and the officials at
Krus Na Ligas High School, Jessica O. Sugay, Dr. Matthew Small, and the Gates
Foundation Grant #OP106038 for collaborating with us.
References
1. Baker, R. S., Corbett, A. T., Koedinger, K. R., & Wagner, A. Z. (2004). Off-task
behavior in the cognitive tutor classroom: when students game the system. In
Proceedings of the SIGCHI conference on Human factors in computing systems (pp.
383-390). ACM.
2. Biswas, G., Kinnebrew, J. S., & Segedy, J. R. (2011). Using a Cognitive/Metacognitive
Task Model to analyze Students Learning Behaviors.
3. Cohen, J. (1960). A Coefficient of Agreement for Nominal Scales. Educational and
Psychological Measurement, 20 (1960), 37-46.
4. Craig, S., Graesser, A., Sullins, J., & Gholson, B. (2004). Affect and learning: an
exploratory look into the role of affect in learning with AutoTutor. Journal of
Educational Media, 29(3), 241-250.
5. Csikszentmihalyi, M. (1990). Flow: The psychology of optimal experience. New Y ork:
Harper Perennial.
6. D’Mello, S. K., Craig, S. D., Witherspoon, A., McDaniel, B., & Graesser, A. (2005).
Integrating affect sensors in an intelligent tutoring system. In Proceedings of the
Workshop on Affective Interactions: The computer in the affective loop workshop,
International conference on intelligent user interfaces (pp. 7- 13). New York:
Association for Computing Machinery.
7. D’Mello, S., Graesser, A. (2012). Dynamics of affective states during complex
learning. Learning and Instruction, 22(2): 145-157.
8. Fisherl, C. D. (1993). Boredom at work: A neglected concept. Human Relations, 46(3),
395-417.
9. Ocumpaugh, J., Baker, R.S., Rodrigo, M.M.T. (2015) Baker Rodrigo Ocumpaugh
Monitoring Protocol (BROMP) 2.0 Technical and Training Manual. Technical Report.
New York, NY: Teachers College, Columbia University. Manila, Philippines: Ateneo
Laboratory for the Learning Sciences.
10. Pardos, Z. A., Baker, R. S., San Pedro, M., Gowda, S. M., & Gowda, S. M. (2014).
Affective States and State Tests: Investigating How Affect and Engagement during the
School Year Predict End-of-Year Learning Outcomes. Journal of Learning Analytics,
1(1), 107-128.
11. San Pedro, M. O. Z., d Baker, R. S., Gowda, S. M., & Heffernan, N. T. (2013, January).
Towards an understanding of affect and knowledge from student interaction with an
Intelligent Tutoring System. In Artificial Intelligence in Education (pp. 41-50).
Springer Berlin Heidelberg.
12. Shute, V. J., Ventura, M., & Kim, Y. J. (2013). Assessment and Learning of Qualitative
Physics in Newton's Playground. The Journal of Educational Research, 106(6), 423-
430.
13. Storey, J.D. (2002). A direct approach to false discovery rates. Journal of the Royal
Statistical Society, Series B, 64: 479-498.