LINK BETWEEN SENTIMENT AND HUMAN ACTIVITY REPRESENTED BY FOOTSTEPS

Experiment Exploiting IoT Devices And Social Networks

ABSTRACT

One subject (the first author of the paper) participated in a pilot experiment, which lasted 14 days. His task was to measure the number of
footsteps using a professional pedometer and write short texts expressing his mood that were evenly distributed throughout the day. The original
goal was to walk at least 10.000 steps a day and provide at least 20 short text messages a day.

Experiment = 14 days {10.000 steps + 20 tweets per day}

Measured steps per day

DATA OVERVIEW
- During 14 days of experiment done
- 94.591 steps in total
- 280 tweets in total
- Which is
- 6.625 steps in average per day
- 20 tweets in average per day

SENTIMENT
ANALYSIS

- Small group of 7 people .

- Three classification values:
positive (1), negative (-1) and
neutral (0).
Evaluation of sentiment classes

=

Counts

neg (-1)
neu (0)
pos (1)
Total

52
21
207
280

= 0

HYPOTHESIS DEFINITION
H0: Movement does not affect mood.
0 : =
HA: Movement results in a positive mood.
: <
- Exploratory data analysis
- Statistical Inference (paired t-test)

CONCLUSIONS

- Every tweet represents a document

- List of the documents makes a corpus
- Due to different granularity of records,
steps are aggregated to tweets

Classes

HYPOTHESIS TESTING METHODS

- Texts were recorded using Twitter

- Steps were measured using wearable
IoT Jawbone UP24

DATA PROCESSING

Human eval. summary

HUMAN EVALUATION

HEALTHINF
2016

Recorded tweets per day

TOOLS

= 0

MACHINE LEARNING PREDICTION

- Test corpus is 259 docs (without neutral)
- Used two external corpora for training:
- Movie review corpus, 10662
documents: pos. and neg.
- Twitter sentiment corpus, 10786
documents: pos., neu., neg. and n/a.
- Evaluated 5 ML methods.
- Chosen Random Forrest (75 % accuracy)
trained with Twitter sentiment corpus.

Density plot of human

sentiment evaluation

Density plot of ML
sentiment prediction

Random Forrest cross table

predicted
neg (-1) pos (1)
evaluated

MEASUREMENTS

The Internet of Things world brings to our lives many opportunities to monitor our daily activities by collecting data from various devices.
Complementary to it, the data expressing opinions, suggestions, interpretations, contradictions, and uncertainties are more accessible within
variety of online resources. This paper deals with collection and analysis of hard data representing the number of steps and soft data representing
the mood of participants who underwent a pilot experiment. The paper defines outlines of the problem and presents possible sources of reliable
data, sentiment evaluation, sentiment extraction using machine learning methods, and links between the data collected from IoT devices and
mood expressed by the participant in a textual form. The results provided by using inferential statistics are presented.

EXPERIMENT

Jaromr Salamon, Roman Mouek

Department of Computer Science and Engineering, Faculty of Applied Sciences,
University of West Bohemia, Plze, Czech Republic

neg (-1)

22

30

52

pos (1)

35

172

207

57

202

259

Summary of t-test results

t
df
mean diff
SE difference
95% lower
C.I. upper
p

methods
machine
human
learning
-0.70
-1.34
71.47 182.37
-76.43 -92.86
96.44 -346.41
-295.73 -229.89
142.88
44.16
0.4894 0.1828

- BIAS We are fully aware that the experiment has several biases. At first, walking is related to need to move somewhere. Secondly, expression
of mood is subjective as well as its evaluation. Lastly, but not less importantly, we need to consider conditions for Tweet writing.
- CORPORA Selection of a right corpus plays a significant role in future analysis. Better results for the Twitter sentiment corpus might be given
by the fact that we also used tweets in our experiment.
- HYPOTHESIS Summarizing all the results from Statistical Inference we can see that in both cases we can accept the null hypothesis H0:
Movement does not affect mood and reject the alternative hypothesis HA: Movement results in a positive mood.
- RESULT It is obvious that the result we got it is not corresponding to what we expected, namely to confirm the alternative hypothesis.
- IMPROVEMENTS We believe that the heart rate measured independently of current human activity, but together with steps and in the same
timeline with sentiment can give us better answers.

Pictograms credit: note by Ashley Ma, molecule by Zlatko Najdenovski, vitals monitor by Zlatko Najdenovski, smartwatch by Leonardo Schneider, bird by Thomas Le Bas, steps by Eugen Belyakoff, economy graph
by Vectors Market, diet and exercise by Jim Sullivan, funnel by Zlatko Najdenovski all from Noun Project.