HUMAN ACTIVITY

RECOGNITION
USING WISDM
DATASET
Deep learning approach for human activity
recognition
 HUMAN ACTIVITY RECOGNITION
➢ 1.Human Activity Recognition (HAR) is the process of
identifying and classifying physical activities performed by
individuals using data collected from sensors, such as
accelerometers and gyroscopes.

➢ 2.Sensor data is usually collected through:

➢ • Smartphones
➢ • Smartwatches
➢ • Wearable fitness trackers

➢ 3.Key Objectives of HAR:

➢ •Identify and classify human activities automatically
➢ •Enable context-aware computing
➢ •Enhance interaction between humans and smart systems
PURPOSE OF HAR
➢ The main purpose of HAR is to
automatically detect and understand
what people are doing, using data
from sensors, cameras, or other
devices.

Healthcare Monitoring
➢ Track elderly activities, fall
detection, rehabilitation support.
Fitness & Lifestyle
➢ Step counters, workout recognition,
calorie tracking.
Smart Homes
➢ Automate lighting, heating based on
activity.
Security & Surveillance
➢ Detect suspicious or unusual
behavior.
WISDM DATASET
➢ The WISDM (Wireless Sensor Data
Mining) dataset is a widely-used
benchmark dataset for Human Activity
Recognition (HAR) using data collected
from mobile sensors like accelerometers
and gyroscopes.

➢ Origin & Purpose

➢ Developed by the Wireless Sensor Data
Mining (WISDM) Lab at Fordham
University.
➢ Designed to help in training machine
learning and deep learning models for
recognizing physical activities based on
sensor data.
 WORKFLOW
• WISDM Dataset
• It has mobile sensor data (like motion) for different activities like walking, jogging, sitting, etc.
• Split into Training and Testing
• The dataset is split into two parts:
• Training Dataset: Used to teach the model.
• Testing Dataset: Used to check how well the model learned.
• Different Models Used for Training Classifiers Predict Activities
• Walking
• These models learn from the training data: • Walking Upstairs
• Dilated CNN • Walking Downstairs
• CNN • Sitting
• WaveNet • Standing
• RBFNN • Laying
• Jogging
• LSTM
• GRU
• Transformer
• Trained Model
• After learning, we get a trained model ready to predict activities.
 Dilated CNN Walking
Training
Dataset CNN Walking
Upstairs
WAVENET Walking
Downstairs
RBFNN
Different Sitting
WISDM LSTM
Classifiers
DATASET Standing
GRU
Laying
TRANSFORMER
Jogging

Testing
Dataset
TRAINED MODEL
 Activity Distribution Graph (Bar Plot)
➢ The activity distribution graph of the WISDM dataset is a bar chart that shows the number of sensor readings
recorded for each physical activity performed by users.
➢ Observations from the Graph
➢ Walking and Jogging have the tallest bars, meaning they have the highest number of samples:
•Walking ≈ 424,000 samples
•Jogging ≈ 344,000 samples
➤ These activities dominate the dataset.
➢ Upstairs and Downstairs have a moderate number of samples:
•Each has around 120,000 samples.
➢ Sitting and Standing have the fewest samples:
•Less than 60,000 each

X-AXIS Categories of physical activities:

Walking, Jogging, Upstairs, Downstairs, Sitting, Standing
Y-AXIS Number of recorded samples for each activity
BARS Height of each bar represents how many data samples (rows) exist for that
activity
 Number of Samples Per User
➢ This bar graph visualizes how many sensor data samples (rows) were recorded for each of the 36 users in the
WISDM v1.1 dataset.

➢ Total Users: Around 36 users in the WISDM DATASET

➢ Key Observations from the Graph:

User 20 contributed the highest number of samples — over 55,000 samples.
Users 4, 8, and 24 contributed fewer samples, with User 4 having the least (~12,000 samples).
Most users contributed between 25,000 to 40,000 samples,

GRAPH DESCRIPTION:

Represents User IDs from 1 to 36. Each user in the WISDM dataset
X-AXIS
is identified with a unique ID.
Represents the number of samples (rows of data) collected from
Y-AXIS
each user
Show the total sample count contributed by each user. Taller bars =
BARS
more samples.
 DEEP LEARNING
➢ Deep Learning is a subset of machine learning
that uses artificial neural networks with
multiple layers (deep networks) to
automatically learn hierarchical features from
data.
➢ It is especially powerful for:
➢ Image and video analysis
➢ Speech recognition
➢ Natural language processing
➢ Time-series data analysis (like HAR)
➢ Deep learning models like CNN, LSTM, GRU,
and Transformers can discover complex
patterns in data without needing manual feature
extraction.
 CNN
DILATION CNN
DEEP
LEARNING LSTM

MODELS
GRU
USED
TRANSFORMER

RBFNN

WAVENET
 CONVOLUTIONAL
NEURAL NETWORK
 CONVOLUTIONAL NEURAL NETWORK
➢ A Convolutional Neural
Network (CNN) is a type of deep
learning model primarily used for
image and video recognition,
classification, and pattern
detection
• CNN reduces the number of
parameters compared to fully
connected networks, making it
efficient for image data.
• It uses filters (kernels) to extract
features from the image in a layer-
wise manner.
• CNN is used in face recognition,
self-driving cars, medical image
analysis, and more.
CONFUSION MATRIX-CNN
➢ Misclassifications (non-
➢ Graph Description: diagonal values):
➢ X-axis: Predicted Activities ➢ Here’s where the model got
confused:
➢ Y-axis: Actual Activities
➢ Downstairs predicted as Upstairs:
17
➢ Downstairs → Walking: 5
➢ Jogging → Downstairs: 1
➢ Correct Predictions ➢ Jogging → Upstairs: 1
(diagonal values):
➢ Jogging → Walking: 3
➢ Downstairs → Downstairs:
377 ➢ Sitting → Upstairs: 1
➢ Jogging → Jogging: 1338 ➢ Standing → Sitting: 1
➢ Sitting → Sitting: 236 ➢ Upstairs → Downstairs: 5
➢ Standing → Standing: 189 ➢ Upstairs → Jogging: 4
➢ Upstairs → Upstairs: 476 ➢ Upstairs → Sitting: 1
➢ Walking → Walking: 1664 ➢ Upstairs → Walking: 4
➢ Walking → Sitting: 1
Training vs Validation Accuracy and Loss Curves
for CNN Model
Per-Class Metrics – CNN
Activity Precision Recall F1-Score Support

Downstairs 0.98 0.94 0.96 399

Dataset Accuracy

Jogging 1.00 1.00 1.00 1343

Train 99.58%
Sitting 0.99 1.00 0.99 237
Validation 98.98%
Standing 1.00 0.99 1.00 190
Test 98.98%
Upstairs 0.96 0.97 0.97 490

Walking 0.99 1.00 1.00 1665

 ROC Curve-CNN

• The ROC curve shows how well the CNN

distinguishes between different activity
classes by visualizing prediction confidence
across thresholds.
• It helps compare class-wise performance
and identify which activities the CNN
predicts accurately and which are often
confused.
 LONG SHORT
TERM MEMORY
 ➢ LSTM is a special type of Recurrent Neural
Network (RNN) designed to learn long-term
dependencies in sequential data. It overcomes
the vanishing gradient problem found in
traditional RNNs.
➢ Main components of LSTM
➢ Forget Gate
Decides what information to throw away from
the cell state.
➢ Input Gate
Decides what new information to store in the
cell.
➢ Cell State
The memory of the network; carries relevant
information across time steps.

LSTM ➢ Output Gate

Decides what to output based on the cell state.
CONFUSION MATRIX - LSTM
Misclassifications (non-diagonal
➢ Graph Description: values):
➢ X-axis: Predicted Activities ➢ Here’s where the model got
➢ Y-axis: Actual Activities confused:
➢ Downstairs → Upstairs: 33
➢ Downstairs → Walking: 8
➢ Downstairs → Jogging: 2
➢ Downstairs → Standing: 1
Correct Predictions (diagonal ➢ Jogging → Downstairs: 2
values):
➢ Jogging → Upstairs: 5
➢ Downstairs → Downstairs:
➢ Jogging → Walking: 1
355
➢ Sitting → Downstairs: 1
➢ Sitting → Upstairs: 1
➢ Jogging → Jogging: 1335 ➢ Standing → Sitting: 2
➢ Upstairs → Downstairs: 9
➢ Sitting → Sitting: 235
➢ Upstairs → Jogging: 2
➢ Upstairs → Sitting: 1
➢ Standing → Standing: 188
➢ Upstairs → Walking: 7
➢ Walking → Downstairs: 3
➢ Upstairs → Upstairs: 471
➢ Walking → Jogging: 3
➢ Walking → Sitting: 1
➢ Walking → Walking: 1651
➢ Walking → Upstairs: 7
Training vs Validation Accuracy and Loss Curves
for LSTM Model
Per-Class Metrics – LSTM
Activity Precision Recall F1-Score Support

Downstairs 0.96 0.89 0.92 399

Jogging 0.99 0.99 0.99 1343

Stage Accuracy
Train 98.82% Sitting 0.98 0.99 0.99 237
Validation 97.50%
Test 97.94% Standing 0.99 0.99 0.99 190

Upstairs 0.91 0.96 0.94 490

Walking 0.99 0.99 0.99 1665

ROC Curve-LSTM
• The ROC curve for an LSTM model
demonstrates its strength in
modeling long-term temporal
dependencies, which improves class
distinction in time-series data like
HAR.

• A high AUC in the curve indicates

the LSTM's ability to consistently
detect complex patterns across time,
especially for activities with gradual
transitions.
GATED
RECURRENT UNIT
 GRU

➢ GRU is a type of Recurrent Neural Network (RNN)

designed to handle sequence data and solve the vanishing
gradient problem. It is simpler and faster than LSTM while
achieving comparable performance.

➢ Update Gate:
o Decides what to keep from the past.
➢ Reset Gate:
o Decides what to forget from the past.
➢ Candidate State:
o A possible new memory made from the current input
and past state.
➢ Final Output (New State):
o Mix of the old memory and the new one based on the
update gate.
CONFUSION MATRIX-GRU
➢ Graph Description: ➢ Misclassifications
➢ X-axis: Predicted Activities (non-diagonal values):
➢ Y-axis: Actual Activities ➢ Here’s where the model got
confused:
➢ Downstairs → Standing: 1
➢ Downstairs → Upstairs: 13
➢ Correct Predictions ➢ Downstairs → Walking: 2
(diagonal values): ➢ Jogging → Downstairs: 2
➢ Jogging → Upstairs: 4
➢ Downstairs → Downstairs: ➢ Jogging → Walking: 1
383 ➢ Standing → Sitting: 4
➢ Sitting → Upstairs: 1
➢ Jogging → Jogging: 1336 ➢ Upstairs → Downstairs: 11
➢ Upstairs → Jogging: 3
➢ Sitting → Sitting: 236 ➢ Upstairs → Standing: 1
➢ Upstairs → Walking: 5
➢ Standing → Standing: ➢ Walking → Downstairs: 4
186 ➢ Walking → Jogging: 4
➢ Walking → Sitting: 1
➢ Upstairs → Upstairs: 470 ➢ Walking → Upstairs: 5

➢ Walking → Walking: 1651

Training vs Validation Accuracy and Loss Curves for
GRU Model
Per-Class Metrics –
GRU

Activity Precision Recall F1-Score Support

Downstairs 0.96 0.96 0.96 399
Metric Accuracy (%)
Jogging 0.99 0.99 0.99 1343
Train Accuracy 99.48%
Sitting 0.98 1.00 0.99 237
Validation Accuracy 98.66%
Standing 0.99 0.98 0.98 190
Test Accuracy 98.57%
Upstairs 0.95 0.96 0.96 490
Walking 1.00 0.99 0.99 1665
ROC CURVE-GRU

• The ROC curve for a GRU

model shows how well it
captures temporal dependencies
in sequential data, helping
differentiate between activity
classes over time.

• A smoother and higher ROC-

AUC indicates that the GRU is
effectively learning time-based
patterns, especially in activities
with clear sequential transitions.
TRANSFORMER
 TRANSFORMER
➢ A Transformer is a smart model that understands
or generates language (like this chat!). Feed-Forward
It reads the whole sentence at once instead of
word-by-word like RNN or LSTM.
➢ Key Components:
➢ Self-Attention: Encoder
Encoder-Decoder
Each word looks at all other words to Attention
understand the full meaning.
Feed-Forward
➢ Feed Forward
A small brain that helps the model process and
think more after attention.
➢ Encoder
Self-Attention
Reads and understands the input sentence using
attention and feed forward. Self-Attention
Example: Reads “Translate this sentence.”
➢ Decoder:
Generates the output (like a translation) using
the encoder’s info.
CONFUSION MATRIX-TRANSFORMER
➢ Graph Description: ➢ Misclassifications (non-
➢ X-axis: Predicted Activities diagonal values):
➢ Y-axis: Actual Activities ➢ Downstairs → Jogging: 31
➢ Downstairs → Sitting: 2
➢ Downstairs → Standing: 8
➢ Downstairs → Upstairs: 93
➢ Correct Predictions ➢ Downstairs → Walking: 265
(diagonal values): ➢ Jogging → Walking: 32
➢ Sitting → Jogging: 8
➢ Sitting → Standing: 6
➢ Downstairs → Downstairs: 0 ➢ Sitting → Upstairs: 6
➢ Sitting → Walking: 17
➢ Jogging → Jogging: 1311 ➢ Standing → Jogging: 1
➢ Standing → Sitting: 2
➢ Sitting → Sitting: 200 ➢ Standing → Upstairs: 3
➢ Standing → Walking: 1
➢ Upstairs → Jogging: 46
➢ Standing → Standing: 183 ➢ Upstairs → Sitting: 4
➢ Upstairs → Standing: 16
➢ Upstairs → Upstairs: 154 ➢ Upstairs → Walking: 270
➢ Walking → Jogging: 5
➢ Walking → Walking: 1616 ➢ Walking → Sitting: 5
➢ Walking → Standing: 1
➢ Walking → Upstairs: 38
Training vs Validation Accuracy and Loss Curves
for TRANSFORMER Model
PER-CLASS METRICS-
TRANSFORMER
Activity Precision Recall F1-Score Support

Downstairs 0.00 0.00 0.00 399

Dataset Accuracy (%)
Train 80.00 Jogging 0.94 0.98 0.96 1343

Validation 80.74 Sitting 0.94 0.84 0.89 237

Test 80.11
Standing 0.86 0.96 0.91 190

Upstairs 0.52 0.31 0.39 490

Walking 0.73 0.97 0.84 1665

 ROC CURVE-
TRANSFORMER
➢ The Transformer’s ROC curve shows
strong class-wise separation, leveraging
self-attention to effectively model
complex temporal patterns in human
activity recognition.
DILATION CNN
DILATED CNN

➢ Dilated CNN uses “dilated” or “atrous” convolutions,

which allow the model to look at a wider context without
increasing the number of parameters.
➢ Common Layer Structure in Dilated CNN:
➢ Input Layer
➢ Dilated Conv1D/Conv2D Layer
➢ Activation (ReLU)
➢ Dropout (optional)
➢ Flatten/Global Pooling
➢ Dense (output) layer
CONFUSION MATRIX-DILATION CNN
➢ Graph ➢ Misclassifications (non-
Description:
diagonal values):
➢ X-axis: Predicted Activities
➢ Y-axis: Actual Activities ➢ Here’s where the model got
confused:
➢ Correct Predictions
➢ Downstairs → Jogging: 3
(diagonal values): ➢ Downstairs → Upstairs: 14
➢ Downstairs → Downstairs:
➢ Jogging → Upstairs: 3
382 ➢ Sitting → Standing: 1
➢ Standing → Sitting: 1
➢ Jogging → Jogging: 1340 ➢ Upstairs → Downstairs: 3
➢ Upstairs → Jogging: 3
➢ Sitting → Sitting: 236 ➢ Upstairs → Walking: 1
➢ Walking → Downstairs: 3
➢ Standing → Standing: 189
➢ Walking → Jogging: 3
➢ Upstairs → Upstairs: 483
➢ Walking → Sitting: 1
➢ Walking → Upstairs: 5
➢ Walking → Walking: 1653
Training vs Validation Accuracy and Loss Curves for
DILATION CNN Model
PER-CLASS METRICS –
DILATION CNN
F1-
Activity Precision Recall Support
Score
Downstairs 0.98 0.96 0.97 399
Dataset Accuracy (%)
Jogging 0.99 1.00 1.00 1343
Train 99.47
Validation 98.75 Sitting 0.99 1.00 0.99 237
Test 99.05 Standing 1.00 0.99 1.00 190
Upstairs 0.95 0.99 0.97 490
Walking 1.00 0.99 1.00 1665
 ROC CURVE-
DILATION CNN

• The ROC curve for a Dilated CNN

reflects its ability to capture long-
range temporal patterns in data (like
sensor signals in HAR) and classify
them effectively across various
thresholds.

• It typically shows high AUC values

when the dilation allows the model
to detect subtle variations between
similar activities, improving class
separation.
 REDIAL BASIS
FUNCTION NEURAL
NETWORK
RADIAL BASIS FUNCTION NEURAL NETWORK
• RBFNN is a type of neural network that uses radial basis
functions as activation functions — mainly used for
pattern recognition, function approximation, and
classification tasks.
• Main Layers of RBFNN
• Input Layer
Just passes the features (like sensor values or pixels) to
the next layer without processing.
• Hidden Layer
Uses Radial Basis Functions (typically Gaussian).
• Output Layer
Usually linear.
Combines outputs from the hidden layer to make
predictions or classifications.
CONFUSION MATRIX - RBFNN Misclassifications
• Downstairs → Jogging: 62
• Downstairs → Sitting: 1
• Downstairs → Standing: 10
➢ Graph Description: • Downstairs → Upstairs: 4
• Downstairs → Walking: 317
➢ X-axis: Predicted Activities • Jogging → Downstairs: 1
• Jogging → Sitting: 4
➢ Y-axis: Actual Activities • Jogging → Standing: 1
• Jogging → Upstairs: 2
• Jogging → Walking: 269
➢ Correct Predictions • Sitting → Jogging: 7
• Sitting → Standing: 5
➢ Downstairs → Downstairs:
• Sitting → Walking: 5
5 • Standing → Jogging: 2
• Standing → Sitting: 3
➢ Jogging → Jogging: 1066 • Standing → Upstairs: 2
• Standing → Walking: 26
➢ Sitting → Sitting: 220 • Upstairs → Downstairs: 2
• Upstairs → Jogging: 108
• Upstairs → Sitting: 2
➢ Standing → Standing: 157
• Upstairs → Standing: 6
• Upstairs → Walking: 367
➢ Upstairs → Upstairs: 367 • Walking → Downstairs: 1
• Walking → Jogging: 231
➢ Walking → Walking: 1410 • Walking → Sitting: 5
• Walking → Standing: 16
• Walking → Upstairs: 5
Training vs Validation Accuracy and Loss Curves for
RBFNN Model
PER-CLASS
METICS-RBFNN
F1-
ACTIVITY PRECISION RECALL SCORE SUPPORT

Downstairs 0.56 0.01 0.02 399

Metric Accuracy
Jogging 0.72 0.79 0.76 1343
Train 67.45%
Sitting 0.94 0.93 0.93 237
Validation 65.26%
Test 66.21% Standing 0.81 0.83 0.82 190

Upstairs 0.33 0.01 0.02 490

Walking 0.59 0.85 0.69 1665

ROC CURVE-RBFNN

• The ROC curve for RBFNN

(Radial Basis Function Neural
Network) helps visualize how
well the model separates
different classes using its radial-
based decision boundaries.

• It highlights the sensitivity of the

RBFNN to class overlap—
showing stronger performance
when classes are well-separated
and weaker when features are
closely clustered.
WAVENET
WAVENET
➢ WaveNet is a deep generative model developed by
DeepMind, originally used for audio generation, but
also effective for time series prediction and human
activity recognition.
➢ Causal Convolution
➢ Ensures no future information is used for predicting
the current output.
➢ Dilated Convolution
➢ Expands the receptive field exponentially, allowing
the network to see far back in the sequence
efficiently.
➢ Gated Activation Units
➢ Combines tanh and sigmoid outputs to control the
flow of information (like in GRUs).
CONFUSION MATRIX-WAVENET
➢ Graph
Description:
➢ Misclassifications:
➢ X-axis: Predicted Activities
➢ Downstairs → Upstairs: 6
➢ Y-axis: Actual Activities ➢ Downstairs → Walking: 1
➢ Jogging → Downstairs: 2
➢ Correct Predictions: ➢ Jogging → Upstairs: 1
➢ Sitting → Standing: 1
➢ Downstairs → Downstairs: ➢ Sitting → Upstairs: 1
392 ➢ Standing → Sitting: 1
➢ Upstairs → Downstairs: 10
➢ Jogging → Jogging: 1340 ➢ Upstairs → Jogging: 2
➢ Upstairs → Sitting: 1
➢ Sitting → Sitting: 236
➢ Upstairs → Walking: 5
➢ Standing → Standing: 189
➢ Walking → Downstairs: 3
➢ Walking → Jogging: 2
➢ Upstairs → Upstairs: 472 ➢ Walking → Sitting: 1
➢ Walking → Upstairs: 1
➢ Walking → Walking: 1658
Training vs Validation Accuracy and Loss Curves for
WAVENET Model
PER-CLASS METRICS
WAVENET
Activity Precision Recall F1-Score Support

Metric Accuracy (%) Downstairs 0.96 0.98 0.97 399

Train Accuracy 99.38% Jogging 1.00 1.00 1.00 1343

Validation
98.57% Sitting 0.99 1.00 0.99 237
Accuracy
Test Accuracy 99.14% Standing 1.00 0.99 1.00 190

Upstairs 0.98 0.96 0.97 490

Walking 1.00 1.00 1.00 1665

 ROC CURVE-
WAVENET

• WaveNet’s ROC curve demonstrates

excellent classification performance,
with high true positive rates and low
false positive rates across most
activity classes, indicating its
effectiveness in capturing temporal
dependencies in human activity
data.
MODEL
COMPARISION
 MODEL TEST ACCURACY MSE OBSERVATIONS
Best accuracy and lowest
WaveNet 99.14% 0.1020
MSE overall
Excellent performance,
Dilation CNN 99.05% 0.1092
very close to WaveNet
Strong performance,
CNN 98.98% 0.1371
slightly higher MSE
Performs well, slightly
GRU 98.57% 0.1656
behind CNN models
Acceptable accuracy,
LSTM 97.94% 0.2599
higher MSE than GRU
Underperformed, likely
TRANSFORMER 80.11% 2.2717
needs tuning
Lowest accuracy and
RBFNN 66.21% 4.1321
highest MSE

➢ BEST PERFORMER: WaveNet is the best-performing model.

➢ WEAK PERMORERS: Transformer and RBFNN
MODEL
PERFORMANCE
 F1- F1-
TEST TRAIN BEST
MODEL MSE SCOR WORST SCOR
ACC. ACC. CLASS
E CLASS E

WaveNet 99.14% 99.38% 0.1020 Jogging 1.00 Upstairs 0.97

Dilation CNN 99.05% 99.47% 0.1092 Jogging 1.00 Downstairs 0.97

CNN 98.98% 99.58% 0.1371 Jogging 1.00 Downstairs 0.96

GRU 98.57% 99.48% 0.1656 Jogging 0.99 Downstairs 0.96

LSTM 97.94% 98.82% 0.2599 Jogging 0.99 Downstairs 0.92

Transformer 80.11% 80.00% 2.2717 Jogging 0.96 Downstairs 0.00

RBFNN 66.21% 67.45% 4.1321 Sitting 0.93 Downstairs 0.02

➢ Jogging is the best recognized activity by nearly all models.

➢ Downstairs is the worst classified activity across all models for:
COMPARISION OF SINGLE
MODELS
 MODEL TEST ACCURACY TRAIN ACCURACY VALIDATIOB ACCURACY

WaveNet 99.14% 99.38% 98.57%

Dilation CNN 99.05% 99.47% 98.75%

CNN 98.98% 99.58% 98.98%

GRU 98.57% 99.48% 98.66%

LSTM 97.94% 98.82% 97.50%

Transformer 80.11% 80.00% 80.74%

RBFNN 66.21% 67.45% 65.26%

GRAPHS
GRAPHS COMPARING TEST AND VALIDATION
ACCURACY OF ALL MODELS
GRAPH COMPARING TRAIN AND VALIDATION
ACCURACY
GRAPH COMPARING TRAIN LOSS AND
VALIDATION LOSS
 MODEL BEHAVIOUR

Test
Specificity Sensitivity Training
Model Accuracy Test Loss MSE
(%) (%) Time (sec)
(%)

WaveNet 99.14 0.0303 99.82 98.83 0.1020 774.10

Dilation CNN 99.05 0.0332 99.81 98.74 0.1092 317.70

CNN 98.98 0.0339 99.78 98.37 0.1371 464.67

GRU 98.57 0.0440 99.71 98.00 0.1656 1391.64

LSTM 97.94 0.0685 99.58 96.96 0.2599 1602.29

Transformer 80.11 0.5857 95.04 67.80 2.2717 2470.54

RBFNN 66.21 0.9352 91.27 56.97 4.1321 21.24

VISUALIZATION OF
MODELS
THANK YOU