Human Activity Recognition using Deep Learning Models on
Smartphones and Smartwatches Sensor Data
Bolu Oluwalade
1
, Sunil Neela
1
, Judy Wawira
2
, Tobiloba Adejumo
3
and Saptarshi Purkayastha
1
1
Department of BioHealth Informatics, Indiana University-Purdue University Indianapolis, U.S.A.
2
Department of Radiology, Imaging Sciences, Emory University, U.S.A.
3
Federal University of Agriculture, Abeokuta, Nigeria
Keywords: Human Activities Recognition (HAR), WISDM Dataset, Convolutional LSTM (ConvLSTM).
Abstract: In recent years, human activity recognition has garnered considerable attention both in industrial and academic
research because of the wide deployment of sensors, such as accelerometers and gyroscopes, in products such
as smartphones and smartwatches. Activity recognition is currently applied in various fields where valuable
information about an individual’s functional ability and lifestyle is needed. In this study, we used the popular
WISDM dataset for activity recognition. Using multivariate analysis of covariance (MANCOVA), we
established a statistically significant difference (p < 0.05) between the data generated from the sensors
embedded in smartphones and smartwatches. By doing this, we show that smartphones and smartwatches
don’t capture data in the same way due to the location where they are worn. We deployed several neural
network architectures to classify 15 different hand and non-hand oriented activities. These models include
Long short-term memory (LSTM), Bi-directional Long short-term memory (BiLSTM), Convolutional Neural
Network (CNN), and Convolutional LSTM (ConvLSTM). The developed models performed best with watch
accelerometer data. Also, we saw that the classification precision obtained with the convolutional input
classifiers (CNN and ConvLSTM) was higher than the end-to-end LSTM classifier in 12 of the 15 activities.
Additionally, the CNN model for the watch accelerometer was better able to classify non-hand oriented
activities when compared to hand-oriented activities.
1 INTRODUCTION
Over the past decade, smartphones have become an
important and indispensable aspect of human lives.
More recently, smartwatches have also been widely
accepted as an alternative to conventional watches,
which is referred to as the quantified self movement
(Swan, 2013). Even in low- and middle-income
countries, smartphones have been embedded in the
social fabric (Purkayastha et al., 2013), even though
smartwatches haven’t. Thus, working on both
smartphone and smartwatch sensor is still quite
relevant. Smartphones and smartwatches contain
sensors such as accelerometer, gyroscope, GPS, and
much more. These sensors help capture activities of
daily living such as walking, running, sitting, etc.
and is one of the main motivations for many to own
a smartwatch. Previous studies have shown
accelerometer and gyroscope to be very effective in
recognition of common human activity (Lockhart et
al., 2011). In this study, we classified common human
activities from the WISDM dataset through the use of
deep learning algorithms.
The WISDM (Wireless Sensor Data Mining) Lab
in the Department of Computer and Information
Science of Fordham University collected data from
the accelerometer and gyroscope sensors in the
smartphones and smartwatches of 51 subjects as they
performed 18 diverse activities of daily living (Weiss,
2019). The subjects were asked to perform 18
activities for 3 minutes each while keeping a
smartphone in the subject’s pocket and wearing the
smartwatch in the dominant hand. These activities
include basic ambulation related activities (e.g.,
walking, jogging, climbing stairs), hand-based daily
activities (e.g., brushing teeth, folding clothes), and
various eating activities (eating pasta, eating chips)
(Weiss, 2019). The smartphone and smartwatch
contain both accelerometer and gyroscope sensors,
yielding a total of four sensors. The sensor data was
collected at a rate of 20 Hz (i.e., every 50ms). Either
one of Samsung Galaxy S5 or Google Nexus 5/5X
smartphone running the Android 6.0 was used. The
LG G Watch was the smartwatch of choice. A total of
Oluwalade, B., Neela, S., Wawira, J., Adejumo, T. and Purkayastha, S.
Human Activity Recognition using Deep Learning Models on Smartphones and Smartwatches Sensor Data.
DOI: 10.5220/0010325906450650
In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 5: HEALTHINF, pages 645-650
ISBN: 978-989-758-490-9
Copyright
c
2021 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
645
15,630,426 raw measurements were collected.
We answer three main research questions in the
analysis of this data. Firstly, are there significant
differences in how the two devices capture data, even
though they both contain similar accelerometer and
gyroscope sensors? Secondly, what is the best method
to recognize the activities that were performed by the
subjects through this sensor data? Thirdly, by
using an identified activity, how accurately can we
forecast or simulate the activities? We conducted a
MANCOVA analysis, which showed that there is a
statistically significant difference between the data
obtained from the accelerometer and gyroscope in the
smartphone and smartwatch. After separating the
smartphone and smartwatch data, we classified the
activities using neural network architectures,
including Long short-term memory (LSTM), Bi-
directional Long short-term memory (BiLSTM),
Convolutional Neural Network (CNN), and
Convolutional LSTM (ConvLSTM). Finally, we
developed a GRU model to forecast the last 30
seconds of the watch accelerometer’s raw values and
calculated the accuracy metrics of this forecasting
with the actual values.
2
RELATED WORKS
Due to the increase in the availability of several
sensors like accelerometer and gyroscope in various
consumer products, including wearables, there has
been a rise in the number of research studies on human
activity recognition (HAR) using sensor data. In one
of the earliest HAR studies, Kwapisz et al. used
phone accelerometers to classify 6 human activities,
including walking, jogging, climbing the stairs,
walking down the stairs, sitting, and standing using
machine learning models like logistic regression and
multilayer perceptron (Kwapisz et al., 2011). Their
models recognized most of the activities with an
accuracy of 90%. Esfahni et al. created the PAMS
dataset containing both smartphone’s gyroscope and
accelerometer data (Esfahani and Malazi, 2017).
Using the section of the data collected from holding
the phone with non-dominant hand, they developed
multiple machine learning models to identify the
same six activities as Kwapisz et al., and obtained a
precision of more than 96% for all the models.
Random forest and multilayer perceptron models
outperformed the rest, with a precision of 99.48% and
99.62% respectively. These results were better than
the ones obtained from data collected when the phone
was held in the dominant thigh (Esfahani and Malazi,
2017). Also, Schalk et al. obtained more than 94%
accuracy for same activities as above by developing a
LSTM RNN model (Pienaar and Malekian, 2019).
Agarwal et al. proposed a LSTM-CNN Architecture
for Human Activity Recognition learning model for
HAR. This model was developed by combining a
shallow RNN and LSTM algorithm, and its overall
accuracy on the WISDM dataset achieved 95.78%
accuracy (Agarwal and Alam, 2020). In addition,
previous studies like Walse et al. (Walse et al., 2016)
and Khin (Oo, 2019) have also used the WISDM
accelerometer data to classify a maximum of 6
activities in their work. Although the above models
could generally recognize human activities, they were
evaluated on their ability to recognize just six human
activities and therefore do not provide generalization.
Our study addresses these shortcomings by
developing several deep learning algorithms for 15
human activities recorded in the WISDM data. We
select the best model based on the F1 score, i.e.,
considering both precision and recall. Here, we
obtained an average classification accuracy of more
than 91% in our best performing model. Additionally,
we try to simulate the data 30 seconds into the future
and provide metrics that might be used by other
researchers to build more generalizable models.
3
METHODOLOGY
3.1
WISDM Dataset Description
The WISDM dataset contains raw time series data
from phone and watch’s accelerometer and gyroscope
(X-, Y-, and Z-axis). The raw accelerometer’s and
gyroscopes signals consist of a value related to each
of the three axes. The raw data was segmented into
10-second data without overlapping. Three minutes
(180 Seconds) of raw data was divided into eighteen,
10-seconds segment where each segment’s range was
calculated to obtain 18 values. This was further
divided into 10 equal-sized bins to give X 0-9; Y 0-9,
Z0-9. Features were generated based on these 10 bins
obtained the raw accelerometer and gyroscope
readings. Binned distribution, average, standard
deviation, variance, average absolute difference, and
time between the peaks for each axis were calculated.
Apart from these, other features, including Mel-
frequency cepstral coefficients, cosine distance,
correlation, and average resultant acceleration, were
calculated but are not used in this study. After this
preprocessing, we finally had the following entries for
each of the device sensors:
phone accelerometer: 23,173
HEALTHINF 2021 - 14th International Conference on Health Informatics
646
phone gyroscope: 17,380
watch accelerometer: 18,310
watch gyroscope: 16,632
Also, the activities are divided into two classes: non-
hand and hand-oriented activities.
Non-hand-oriented activities: walking, jogging,
stairs, standing, kicking, and sitting
Hand-oriented activities: dribbling, playing
catch, typing, writing, clapping, brushing teeth,
folding clothes, drinking and eating
3.2
MANCOVA Analysis
Multivariate Analysis of (Co)Variance (MANCOVA)
explores the relationship between multiple dependent
variables, and one or more categorical and/or
continuous outcome variables. To perform the
MANCOVA analysis on our data, we used the X-, Y-
, Z-axis data of the accelerometer and gyroscope at
each time epoch as the dependent variables. The
categorical variables were the phone and watch. We
defined our null hypothesis that there is no
statistically significant difference between the phone
and watch data. The alternate hypothesis was that
there is a statistically significant difference between
the phone and watch data. The null hypothesis was
rejected as we obtained a statistically significant
difference between the phone and the watch data (p
less than 0.05).
3.3
Classification Models
The MANCOVA analysis showed a significant
difference between the phone and watch data.
Therefore, we created separate classification models
for the phone and watch. We selected 44 features from
the watch and phone data with an activity label
attached to each row in the dataset. The activities
include walking, jogging, walking up the stairs,
sitting, standing, typing, brushing teeth, eating soup,
eating chips, eating pasta, drinking, eating
sandwiches, kicking, playing catch, dribbling a ball,
writing, and clapping. Thus, a total of 18 activities. We
combined all the different eating activities to form a
combined ”eating” activity. This reduced the number
of activities to 15. We used the Keras package in
Python for the experiments. The architecture of the
watch accelerometer classification models is
described below. In each case of training the models,
we stopped the epochs when the training loss became
equal to the validation loss to prevent overfitting:
3.3.1
Long Short-term Memory (LSTM)
Networks
Hochreiter and Schmidhuber originally introduced
LSTMs (Hochreiter and Schmidhuber, 1997), and
were refined and popularized later (Sherstinsky, 2018)
(Zhou et al., 2016). LSTMs are a special kind of
recurrent neural networks (RNNs) capable of learning
long-term dependencies. This quality in the network
architecture helps to remember certain useful parts of
the sequence and helps in learning parameters more
efficiently. The scaled dataset was input into the
LSTM model containing 128 LSTM units, followed
by a dropout layer (0.3 units), dense layer (64 Units),
dropout layer (0.2 units), dense layer (64 units), dense
layer (32 units) and a last dense layer with 15 units
(for the number of classes). We used Softmax as the
activation function in the last layer, ReLU for the
previous layers, and the Adam optimizer. The loss was
calculated in categorical cross-entropy. The model
was trained for 226 epochs with a batch size of 32.
3.3.2
Bi-directional Long Short-term
Memory (BILSTM) Networks
We also implemented a BiLSTM model to observe
the effects of either direction on performance. In the
BiLSTM, we fed the data once from the beginning to
the end and once from the end to the beginning. By
using two hidden states in BiLSTM, we can preserve
information from both the past and the future at any
point in time. The parameters of our BiLSTM model
is similar to the LSTM model.
3.3.3
Convolutional LSTM
Xingjian introduced convolutional LSTM’s (Shi et
al., 2015) in 2015. Convolutional LSTMs
(ConvLSTM) are created by extending the fully
connected LSTM to have a convolutional structure in
both the input-to-state and state-to-state transitions.
ConvLSTM network captures spatiotemporal
correlations better and outperforms Fully Connected
LSTM networks.
The scaled data is reshaped and inputted to a 1-
dimensional convolution layer with 128 filters of
kernel size 4 followed by a dropout layer (0.4), LSTM
layer with 128 units, Dense layer with 100 units,
Dense layer with 64 units, Dropout layer with 0.2
dropout rate, Dense layer with 32 units, and finally a
Dense layer with 15 units for classification. We used
Softmax as the activation function in the last layer,
ReLU for the previous layers, and Adam optimizer.
The loss was calculated in categorical cross-entropy.
Human Activity Recognition using Deep Learning Models on Smartphones and Smartwatches Sensor Data
647
The model was trained for 95 epochs with a batch size
of 32.
3.3.4
Convolutional Neural Network (CNN)
The data is reshaped and inputted to a 1-dimensional
convolution layer with 128 filters of kernel size 10
followed by a dropout layer (0.4), 1-dimensional
convolution layer with 128 units and 10 kernel size,
dropout layer with 0.2 dropout rate, 1-dimensional
max-pooling layer with 0.2 pool size, flatten layer,
dense layer with 64 units and finally a Dense layer
with 15 units for classification. We used Softmax as
the activation function in the last layer, ReLU for the
previous layers, and Adam optimizer. The loss was
calculated in categorical cross-entropy. The model
was trained for 148 epochs with a batch size of 32.
4
RESULTS
4.1
Classification Results
To classify the activities, we utilized four classifiers,
namely Long short-term memory (LSTM), Bi-
directional Long short-term memory (LSTM),
Convolutional Neural Network (CNN), and
Convolutional LSTM (ConvLSTM). The Precision
and F1 scores were used as evaluation metrics to
analyze the performance of the classifiers.
Table 1: The Macro-F1 values of different classifiers for
watch sensor data.
Models Acceleromete
r
Gyroscope
Both
CNN 0.849
0.687
0.774
BiLSTM 0.848
0.617
0.721
ConvLSTM 0.843
0.658
0.754
LSTM 0.825
0.627
0.743
Table 2: The Macro-F1 values of different classifiers for
phone sensor data.
Models Accelerometer Gyroscope
Both
CNN 0.796
0.387
0.631
BiLSTM 0.773
0.429
0.611
ConvLSTM 0.814
0.432
0.638
LSTM 0.756
0.395
0.743
Tables 1 and 2 demonstrate the Macro-F1 values
of the different classifiers for the watch and phone
data. From the above tables, the classifiers performed
better with the watch data. Also, the classifier
performed better with accelerometer than gyroscope
data. Thus, the classifiers performed best on the watch
accelerometer data with a Macro-F1 measure of
0.849, 0.848 and 0.843 and 0.825 for the CNN,
BiLSTM, ConvLSTM, and LSTM models,
respectively. These classifiers’ performance was not
greatly different from each other, with all the
classifiers having a Macro-F1 measure of more than
0.80 (80%) with the CNN marginally performing the
best with a Macro-F1 measure of 0.849 (84.9%).
The watch accelerometer data were divided into
training, testing, and validation data, respectively, in
0.8, 0.1, and 0.1 ratios. These data contain 14648,
1831, and 1831 records for training, testing, and
validation, respectively. The classifiers’ precision
values on the watch accelerometer data separated into
hand oriented and non-hand oriented classes are
presented in Tables 3 and 4 below.
Table 3: The precision values of different classifiers for
watch accelerometer data (Non hand-oriented activities).
Activities CNN BiLSTM ConvLSTM
LSTM
Mean
Walking 0.861 0.869
0.925
0.846
0.875
Jogging 0.833 0.859
0.865
0.872
0.857
Stairs 0.909 0.939
0.900
0.862
0.903
Sitting 0.918 0.836
0.844
0.851
0.862
Standing 0.917 0.818
0.669
0.760
0.791
Kicking
0.926 0.857 0.867
0.857
0.877
Mean
0.894 0.863 0.845
0.841
0.861
Table 4: The precision values of different classifiers for
watch accelerometer data (Hand-oriented activities).
Activities CNN BiLSTM ConvLSTM LSTM
Mean
Typing 0.862 0.773 0.838
0.671
0.786
Brushing 0.981 0.955 0.963
0.972
0.968
Drinking 0.867 0.808 0.801
0.768
0.811
Eating 0.757 0.884 0.677
0.779
0.774
Catch 0.850 0.871 0.870
0.906
0.874
Dribbling 0.844 0.802 0.797
0.768
0.802
Writing 0.807 0.816 0.906
0.836
0.841
Clapping 0.785 0.902 0.758
0.769
0.804
Folding 0.821 0.830 0.871 0.794
0.829
Mean 0.842 0.849 0.831 0.807
0.832
Tables 3 and 4 show that the precision of the
convolutional input classifiers (CNN and
ConvLSTM) was higher than the end-to-end LSTM
classifiers. The convolutional classifiers gave the best
classifications in 12 of 15 activities irrespective of the
class of the activities (i.e., non-hand-oriented and
hand-oriented activities). The convolution input
layer has also been shown to outperform conventional
fully connected LSTM in capturing spatiotemporal
correlations in data (Shi et al., 2015). Another fact
that can be inferred from the table 3 above is that the
classifiers perform better in the non-hand-oriented
activities like standing, stairs, etc. than the hand-
oriented activities like eating, clapping, etc. with the
HEALTHINF 2021 - 14th International Conference on Health Informatics
648
exception of brushing teeth, which has the highest
precision values of all the classifiers.
4.2 Forecasting Results
In addition to the classification models, we also
predicted the last 30 seconds of the raw data for all the
eating activities. To achieve this, we utilized a
gated recurrent unit (GRU) model. We used the Root
Mean Square Error (RMSE), Mean Square Error,
Mean Absolute Percentage Error (MAPE), and
symmetric Mean Absolute Percentage Error (sMAPE)
as evaluation metrics.
Table 5: The different performance metrics of different
activities for phone data.
Activities
RMSE
MSE
MAPE
sMAPE
H (eating soup)
0.078
0.00603
126.580
40.57
I (eating chips)
0.070
0.00494
11.43
10.59
J (eating pasta)
0.039
0.00152
7.263
7.25
K (drinking
from cup)
0.083
0.00687
13.080
13.296
L (eating
sandwich)
0.050
0.00247
4.45
4.54
Mean
0.064
0.0044
32.56
15.25
Table 5 shows that GRU gave the best forecast for
eating sandwiches when compared to other activities
like drinking from a cup, eating pasta, eating chips,
eating soup, etc. Also, it can be inferred from Table 5
that the MAPE overstated the error found in activity
H because of the presence of outliers when compared
to the values of other activities with their respective
sMAPE.
5
DISCUSSION
The statistically significant difference (p < 0.05)
between the same kind of sensors in smartphones and
smartwatches using a MANCOVA analysis points to
some interesting observations. This is likely due to the
location of the pocket in comparison to the hand.
However, there is more to it than just the differences
in the X-, Y-, and Z-axis values due to the height of
the sensors from the ground. Our analysis showed that
the difference between the peaks and the throughs was
also larger in the smartwatch. Furthermore,
identifying a constant that differentiated the X-, Y-, or
Z-axis between the two devices was practically
impossible. This means that the difference between
the sensors is not purely due to the height from the
ground. Following this distinction, we created various
deep learning models to classify 15 different activities
based on the smartwatch accelerometer data only. Our
findings suggest that non-hand-oriented activities like
standing, stairs, etc. are better generalized and
better classified than hand-oriented activities like
eating, clapping, etc. with the exception of brushing
teeth as the classifiers performed better and had
higher precision values. This implies that the
smartwatch accelerometer data can better classify
non-hand oriented activities even if the smartwatch is
located on the dominant hand. A recent study by
Agarwal et al. (Agarwal and Alam, 2020), used a
lightweight RNN-LSTM architecture to classify 6
different non-hand oriented activities based on
smartphone accelerometer data. They obtained an
average of 0.9581 (95.81%) precision. Our best
model obtained a precision to 92.6% for kicking
and 98.1% for brushing teeth. We obtained an
average of 89% precision for 6 non-hand oriented
activities recognition and 84% for 9 hand oriented
activities. However, these are not directly
comparable, as we also calculated the Macro-F1
values, which consider both the precision and recall
i.e., when our model accurately classifies the activity
but also fails at classifying the activity. This should
be considered as a better measure for HAR.
Moreover, we also did better at HAR compared to
other papers that used the WISDM data and used
LSTM-CNN, and shallow RNN. Accurate HAR is
clinically relevant, not only because individuals have
started using wearables widely, but also because there
are many clinically-relevant sensors such as the
BioStamp MC10, fall detection devices, and
telemedicine sensors. These sensors might also be
useful to identify mental health issues (Addepally and
Purkayastha, 2017) or motor function disorders (Ellis
et al., 2015) using mHealth apps. All these devices
have accelerometer and gyroscope sensors to
understand patient’s gait, posture, and stability for
accurate measurement of other clinical features. As
home healthcare, senior home care and health care
outside the hospital settings become more common,
the application of HAR is becoming more relevant.
We also created a GRU model to forecast the last
30 seconds of raw data generated by the watch
accelerometer for 4 eating activities based on the
previous 210 seconds data. We obtained an average
RMSE of 0.064, which implies a minimal difference
from the actual values. Thus, we can say that
generating such sensor data for generalizing models
might also be a feasible approach for retraining
models or transfer learning of models to similar
sensors of other devices. This is a future direction that
we are pursuing.
Human Activity Recognition using Deep Learning Models on Smartphones and Smartwatches Sensor Data
649
6
LIMITATIONS
Although the use of classification models such as
Long short-term memory (LSTM), Bi-directional
Long short-term memory, Convolutional Neural
Network (CNN) and Convolutional LSTM, etc., is a
fairly common approach to predicting the movement
of a person, this study does not provide the needed
generalization with the hand-oriented activities. We
have not, for example, examined differences in the
performance metrics of the eating activity when
forecasting the raw values of the last 30 seconds of
the watch accelerometer.
7
CONCLUSION
In this study, we classified smartphone and
smartwatch accelerometer and gyroscope data. We
classified the majority of the activities using artificial
neural network algorithms, including Long short-term
memory (LSTM), Bi-directional Long short-term
memory, Convolutional Neural Network (CNN), and
Convolutional LSTM. Our classification analysis on
15 different activities resulted in an average
classification accuracy of more than 91% in our best
performing model. Although previous findings
indicated that 6 human activities were used during the
analysis, our study followed several 15 human
activities, which are better generalized than those in
major studies conducted previously. It is possible that
outcomes would vary if over 20 or 25 human
activities are used. Future researchers should
consider investigating the impact of more human
activities. Nonetheless, our results provide the needed
generalization for non-hand oriented activities
recognition cases only.
REFERENCES
Addepally, S. A. and Purkayastha, S. (2017). Mobile-
application based cognitive behavior therapy (cbt) for
identifying and managing depression and anxiety. In
International Conference on Digital Human Modeling
and Applications in Health, Safety, Ergonomics and
Risk Management, pages 3–12. Springer.
Agarwal, P. and Alam, M. (2020). A lightweight deep
learning model for human activity recognition on edge
devices. Procedia Computer Science, 167:2364–2373.
Ellis, R. J., Ng, Y. S., Zhu, S., Tan, D. M., Anderson, B.,
Schlaug, G., and Wang, Y. (2015). A validated
smartphone-based assessment of gait and gait
variability in parkinson’s disease. PLoS one,
10(10):e0141694.
Esfahani, P. and Malazi, H. T. (2017). Pams: A new
position-aware multi-sensor dataset for human activity
recognition using smartphones. In 2017 19th
International Symposium on Computer Architecture and
Digital Systems (CADS), pages 1–7. IEEE.
Hochreiter, S. and Schmidhuber, J. (1997). Long short-term
memory. Neural Comput., 9(8):1735–1780.
Kwapisz, J. R., Weiss, G. M., and Moore, S. A. (2011).
Activity recognition using cell phone accelerometers.
ACM SigKDD Explorations Newsletter, 12(2):74–82.
Lockhart, J. W., Weiss, G. M., Xue, J. C., Gallagher, S. T.,
Grosner, A. B., and Pulickal, T. T. (2011). Design
considerations for the wisdm smart phone-based sensor
mining architecture. In Proceedings of the Fifth
International Workshop on Knowledge Discovery from
Sensor Data, pages 25–33.
Oo, K. K. (2019). Daily human activity recognition using
adaboost classifiers on wisdm dataset.
Pienaar, S. W. and Malekian, R. (2019). Human activity
recognition using lstm-rnn deep neural network
architecture. In 2019 IEEE 2nd Wireless Africa
Conference (WAC), pages 1–5. IEEE.
Purkayastha, S., Manda, T. D., and Sanner, T. A. (2013). A
post-development perspective on mhealth–an
implementation initiative in malawi. In 2013 46th
Hawaii International Conference on System Sciences,
pages 4217–4225. IEEE.
Sherstinsky, A. (2018). Fundamentals of recurrent neural
network (RNN) and long short-term memory (LSTM)
network. CoRR, abs/1808.03314.
Shi, X., Chen, Z., Wang, H., Yeung, D., Wong, W., and Woo,
W. (2015). Convolutional LSTM network: A machine
learning approach for precipitation nowcasting. CoRR,
abs/1506.04214.
Swan, M. (2013). The quantified self: Fundamental
disruption in big data science and biological discovery.
Big data, 1(2):85–99.
Walse, K., Dharaskar, R., and Thakare, V. (2016).
Performance evaluation of classifiers on wisdm dataset
for human activity recognition. In Proceedings of the
Second International Conference on Information and
Communication Technology for Competitive Strategies,
pages 1–7.
Weiss, G. M. (2019). Wisdm smartphone and smartwatch
activity and biometrics dataset. UCI Machine Learning
Repository: WISDM Smartphone and Smartwatch
Activity and Biometrics Dataset Data Set.
Zhou, P., Qi, Z., Zheng, S., Xu, J., Bao, H., and Xu,
B. (2016). Text classification improved by integrating
bidirectional LSTM with two-dimensional max pooling.
In Proceedings of COLING 2016, the 26th
International Conference on Computational
Linguistics: Technical Papers, pages 3485–3495,
Osaka, Japan. The COLING 2016 Organizing
Committee.
HEALTHINF 2021 - 14th International Conference on Health Informatics
650