Spotify Song’s Emotion Analysis

Deep Learning
Classification
Published

June 21, 2023

  1. Home
  2. Artificial Intelligence Projects
  3. Spotify Song’s Emotion Analysis

Overview

This notebook investigates the implementation of artificial neural network (ANN) models to classify spotify song’s label using the following dataset.

Download Dataset (.csv)

The analysis covers these main steps:

  1. Data preprocessing and Exploratory Data Analysis (EDA)
  2. Model training
  3. Model evaluation

Additionally, hyperparameter tuning via Keras-Tuner is systematically applied across some model configurations to determine optimal model parameters and prevent overfitting.

The objective of this notebook is to perform an end-to-end machine learning process, implement a Multi Layer Perceptron (MLP) model, and its hyperparameter tuning.

Install and Import Packages

%pip install --upgrade -q keras-tuner matplotlib numpy pandas scikit-learn tensorboard tensorflow
import keras_tuner as kt
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import tensorflow as tf

Load Dataset

df = pd.read_csv('data/spotify.csv')
df.head()
duration (ms) danceability energy loudness speechiness acousticness instrumentalness liveness valence tempo spec_rate labels
0 195000.0 0.611 0.614 -8.815 0.0672 0.0169 0.000794 0.7530 0.520 128.050 3.446154e-07 2
1 194641.0 0.638 0.781 -6.848 0.0285 0.0118 0.009530 0.3490 0.250 122.985 1.464234e-07 1
2 217573.0 0.560 0.810 -8.029 0.0872 0.0071 0.000008 0.2410 0.247 170.044 4.007850e-07 1
3 443478.0 0.525 0.699 -4.571 0.0353 0.0178 0.000088 0.0888 0.199 92.011 7.959809e-08 0
4 225862.0 0.367 0.771 -5.863 0.1060 0.3650 0.000001 0.0965 0.163 115.917 4.693131e-07 1

Data Preprocessing and EDA

Let’s take a look at the overview of the dataset.

df.info()
<class 'pandas.DataFrame'>
RangeIndex: 277938 entries, 0 to 277937
Data columns (total 12 columns):
 #   Column            Non-Null Count   Dtype  
---  ------            --------------   -----  
 0   duration (ms)     277938 non-null  float64
 1   danceability      277938 non-null  float64
 2   energy            277938 non-null  float64
 3   loudness          277938 non-null  float64
 4   speechiness       277938 non-null  float64
 5   acousticness      277938 non-null  float64
 6   instrumentalness  277938 non-null  float64
 7   liveness          277938 non-null  float64
 8   valence           277938 non-null  float64
 9   tempo             277938 non-null  float64
 10  spec_rate         277938 non-null  float64
 11  labels            277938 non-null  int64  
dtypes: float64(11), int64(1)
memory usage: 25.4 MB

Based on the output above, here are some informations about the dataset. - The dataset has 277938 rows and 12 columns. - The target variable is ‘labels’. - All features are already in numerical.

Let’s also see the distribution of the target variable (labels).

df['labels'].value_counts().plot(kind='bar')

Notice that the labels distribution is imbalance. However, we have massive amount of data (totaling over 270,000 samples). Even the smallest class (Class 3) contains around 40,000 sample which means more than sufficient for modern machine learning algorithms to learn the patterns and features of the class without suffering from a lack of data variety.

Check and Drop Duplicated Rows (If Exist)

# check and drop duplicated rows (if exist)
total_duplicates = df.duplicated().sum()
print(f'Total duplicated rows: {total_duplicates}')
if total_duplicates:
    df = df.drop_duplicates().reset_index(drop=True)
    print('All duplicated rows has been removed.')
else:
    print('No duplicated rows need to be dropped.')
Total duplicated rows: 1678
All duplicated rows has been removed.

Check Missing Values

df.isna().sum()
duration (ms)       0
danceability        0
energy              0
loudness            0
speechiness         0
acousticness        0
instrumentalness    0
liveness            0
valence             0
tempo               0
spec_rate           0
labels              0
dtype: int64

Numerical Features Scaling

Based on the displayed df before, we can see that the numerical features exhibit vastly different scales and measurement ranges. MinMaxScaler is applied to transform all numerical features to transform the data into range \([0,1]\).

Before that, split the the data into train, validation, and test set. Here, we are using 80:10:10 ratio for the train, validation, and test set, respectively.

We must split our data first, to avoid data leakage

from sklearn.model_selection import train_test_split

target = 'labels'
y = df[target].values

X_df = df.drop(columns=[target])

X_train_df, X_valtest_df, y_train, y_valtest = train_test_split(X_df, y, test_size=0.2, random_state=42)
X_val_df, X_test_df, y_val, y_test = train_test_split(X_valtest_df, y_valtest, test_size=0.5, random_state=42)
print(f'X_train_df shape: {X_train_df.shape}')
print(f'X_val_df shape: {X_val_df.shape}')
print(f'X_test_df shape: {X_test_df.shape}')
print(f'y_train shape: {y_train.shape}')
print(f'y_val shape: {y_val.shape}')
print(f'y_test shape: {y_test.shape}')
X_train_df shape: (221008, 11)
X_val_df shape: (27626, 11)
X_test_df shape: (27626, 11)
y_train shape: (221008,)
y_val shape: (27626,)
y_test shape: (27626,)
from sklearn.preprocessing import MinMaxScaler

scaler = MinMaxScaler()

X_train = X_train_df.values
X_val = X_val_df.values
X_test = X_test_df.values

X_train_scaled = scaler.fit_transform(X_train)
X_val_scaled = scaler.transform(X_val)
X_test_scaled = scaler.transform(X_test)

X_train_scaled_df = pd.DataFrame(data=X_train_scaled, columns=X_df.columns)
X_val_scaled_df = pd.DataFrame(data=X_val_scaled, columns=X_df.columns)
X_test_scaled_df = pd.DataFrame(data=X_test_scaled, columns=X_df.columns)
X_train_scaled_df
duration (ms) danceability energy loudness speechiness acousticness instrumentalness liveness valence tempo spec_rate
0 0.063740 0.496964 0.7730 0.855589 0.033990 0.019980 0.000000 0.1390 0.2960 0.350949 0.002305
1 0.025156 0.533401 0.0896 0.632955 0.034404 0.995984 0.826000 0.1070 0.0509 0.620996 0.005683
2 0.087669 0.741903 0.6240 0.777582 0.051088 0.143574 0.016000 0.3630 0.9210 0.491596 0.002537
3 0.035692 0.142713 0.0193 0.414395 0.044767 0.952811 0.653000 0.1080 0.0642 0.262751 0.005312
4 0.042879 0.738866 0.3440 0.820073 0.065907 0.945783 0.000000 0.7070 0.7260 0.549274 0.006560
... ... ... ... ... ... ... ... ... ... ... ...
221003 0.046574 0.554656 0.9870 0.868472 0.081244 0.000134 0.000303 0.3430 0.3810 0.424925 0.007468
221004 0.033153 0.614372 0.5970 0.788105 0.032021 0.235944 0.005890 0.1450 0.4360 0.497577 0.004076
221005 0.100396 0.566802 0.1440 0.712119 0.044560 0.925703 0.122000 0.1040 0.3280 0.495401 0.001937
221006 0.092121 0.776316 0.7260 0.803610 0.043420 0.000100 0.924000 0.0983 0.0384 0.522603 0.002054
221007 0.040220 0.763158 0.6010 0.793119 0.073679 0.118474 0.000236 0.1220 0.9040 0.272965 0.007799

221008 rows × 11 columns

X_val_scaled_df
duration (ms) danceability energy loudness speechiness acousticness instrumentalness liveness valence tempo spec_rate
0 0.038265 0.820850 0.433 0.793921 0.125389 0.065161 0.000000 0.1040 0.4220 0.551462 0.013921
1 0.101077 0.386640 0.476 0.705547 0.090674 0.646586 0.000001 0.1190 0.6580 0.600142 0.003915
2 0.053011 0.812753 0.701 0.839173 0.031710 0.095984 0.000000 0.1150 0.5020 0.449020 0.002572
3 0.035242 0.400810 0.156 0.675970 0.030674 0.838353 0.000014 0.0791 0.5140 0.338461 0.003684
4 0.051663 0.466599 0.749 0.853506 0.039378 0.236948 0.000000 0.0998 0.5640 0.326683 0.003274
... ... ... ... ... ... ... ... ... ... ... ...
27621 0.055060 0.434211 0.754 0.880367 0.090777 0.019277 0.004810 0.3440 0.5600 0.611745 0.007097
27622 0.044716 0.605263 0.775 0.809843 0.087979 0.558233 0.703000 0.9000 0.1560 0.387921 0.008411
27623 0.077913 0.405870 0.330 0.726113 0.089430 0.914659 0.000000 0.1010 0.3350 0.576855 0.004985
27624 0.045081 0.387652 0.936 0.859708 0.194819 0.000748 0.012300 0.3400 0.0495 0.413796 0.018479
27625 0.060598 0.622470 0.718 0.820289 0.036269 0.040361 0.899000 0.1040 0.1740 0.408174 0.002584

27626 rows × 11 columns

X_test_scaled_df
duration (ms) danceability energy loudness speechiness acousticness instrumentalness liveness valence tempo spec_rate
0 0.067403 0.574899 0.0533 0.676880 0.042280 0.951807 0.002300 0.1010 0.212 0.469763 0.002715
1 0.059379 0.277328 0.8580 0.841364 0.032435 0.380522 0.208000 0.1720 0.488 0.630508 0.002357
2 0.051501 0.732794 0.6860 0.840145 0.218653 0.021888 0.000000 0.0735 0.451 0.547094 0.018237
3 0.041992 0.695344 0.5630 0.852364 0.039896 0.061948 0.000026 0.4800 0.542 0.312390 0.004052
4 0.065311 0.593117 0.6770 0.796991 0.030466 0.043173 0.093600 0.0884 0.562 0.587886 0.002018
... ... ... ... ... ... ... ... ... ... ... ...
27621 0.041942 0.527328 0.9780 0.904035 0.283938 0.209839 0.272000 0.4240 0.175 0.735547 0.028869
27622 0.056544 0.579960 0.5750 0.807498 0.036373 0.765060 0.569000 0.2990 0.199 0.285494 0.002771
27623 0.156514 0.341093 0.7420 0.861406 0.040933 0.585341 0.000006 0.1960 0.373 0.620955 0.001148
27624 0.052101 0.699393 0.7900 0.845236 0.058135 0.024398 0.000000 0.0453 0.734 0.388031 0.004795
27625 0.071836 0.582996 0.8470 0.820906 0.034611 0.011245 0.037200 0.1180 0.622 0.416482 0.002089

27626 rows × 11 columns

Train Model with Hyperparameter Tuning

Hyperparameter Tuning Using Keras-Tuner

# remember always make sure that the X_train/val/test and y_train/val/test are in float32 datatype
# and y_train/val/test is in form of 1 column
X_train = X_train_scaled_df.values.astype(dtype=np.float32)
y_train = y_train.reshape(-1, 1).astype(dtype=np.float32)

X_val = X_val_scaled_df.values.astype(dtype=np.float32)
y_val = y_val.reshape(-1, 1).astype(dtype=np.float32)

print(f'X_train shape: {X_train.shape}')
print(f'y_train shape: {y_train.shape}')

print(f'X_val shape: {X_val.shape}')
print(f'y_val shape: {y_val.shape}')
X_train shape: (221008, 11)
y_train shape: (221008, 1)
X_val shape: (27626, 11)
y_val shape: (27626, 1)
input_size = X_train.shape[1]

def build_model(hp):
    model = tf.keras.Sequential()
    model.add(tf.keras.layers.Input(shape=(input_size,)))

    n_hid_layers = hp.Int('n_hid_layers', min_value=1, max_value=3, step=1)
    for layer in range(n_hid_layers):
        n_neurons = hp.Int(f'n_neurons_{layer}', min_value=8, max_value=128, step=8)
        act = hp.Choice(f'activation_{layer}', 
                        ['linear', 'relu', 'sigmoid', 'tanh'])
        model.add(tf.keras.layers.Dense(n_neurons, activation=act))
    
    model.add(tf.keras.layers.Dense(4, activation='softmax')) # multiclass

    lr = hp.Float('learning_rate', 1e-5, 1e-2)
    opt = tf.keras.optimizers.AdamW(learning_rate=lr)
    model.compile(loss='sparse_categorical_crossentropy', optimizer=opt, metrics=['accuracy']) # multiclass

    return model
tf.keras.backend.clear_session()

tuner = kt.BayesianOptimization(hypermodel=build_model,
                                objective='val_loss',
                                max_trials=10,
                                directory='tuner_dir_0',
                                project_name='tune_spotify_model',
                                overwrite=False)

tuner.search_space_summary()
Search space summary
Default search space size: 4
n_hid_layers (Int)
{'default': None, 'conditions': [], 'min_value': 1, 'max_value': 3, 'step': 1, 'sampling': 'linear'}
n_neurons_0 (Int)
{'default': None, 'conditions': [], 'min_value': 8, 'max_value': 128, 'step': 8, 'sampling': 'linear'}
activation_0 (Choice)
{'default': 'linear', 'conditions': [], 'values': ['linear', 'relu', 'sigmoid', 'tanh'], 'ordered': False}
learning_rate (Float)
{'default': 1e-05, 'conditions': [], 'min_value': 1e-05, 'max_value': 0.01, 'step': None, 'sampling': 'linear'}

Find the best hyperparameters.

tuner.search(X_train, y_train, validation_data=(X_val, y_val),
             epochs=100, batch_size=256)
Trial 10 Complete [00h 02m 24s]
val_loss: 0.2080817073583603

Best val_loss So Far: 0.18831150233745575
Total elapsed time: 00h 23m 55s
tuner.results_summary(3)
Results summary
Results in tuner_dir_0\tune_spotify_model
Showing 3 best trials
Objective(name="val_loss", direction="min")

Trial 08 summary
Hyperparameters:
n_hid_layers: 3
n_neurons_0: 48
activation_0: tanh
learning_rate: 0.000635494810129079
n_neurons_1: 120
activation_1: tanh
n_neurons_2: 40
activation_2: tanh
Score: 0.18831150233745575

Trial 09 summary
Hyperparameters:
n_hid_layers: 2
n_neurons_0: 96
activation_0: relu
learning_rate: 0.008251039712329468
n_neurons_1: 40
activation_1: linear
n_neurons_2: 48
activation_2: tanh
Score: 0.2080817073583603

Trial 03 summary
Hyperparameters:
n_hid_layers: 1
n_neurons_0: 32
activation_0: relu
learning_rate: 0.004561689776561494
n_neurons_1: 16
activation_1: sigmoid
n_neurons_2: 104
activation_2: relu
Score: 0.2286299765110016

Retrain Best Model

model = build_model(tuner.get_best_hyperparameters()[0])

model.summary()
Model: "sequential_1"

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ dense_3 (Dense)                 │ (None, 48)             │           576 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ dense_4 (Dense)                 │ (None, 120)            │         5,880 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ dense_5 (Dense)                 │ (None, 40)             │         4,840 │
├─────────────────────────────────┼────────────────────────┼───────────────┤
│ dense_6 (Dense)                 │ (None, 4)              │           164 │
└─────────────────────────────────┴────────────────────────┴───────────────┘

 Total params: 11,460 (44.77 KB)

 Trainable params: 11,460 (44.77 KB)

 Non-trainable params: 0 (0.00 B)

early_stop = tf.keras.callbacks.EarlyStopping(patience=10, monitor='val_loss',
                                              restore_best_weights=True,
                                              verbose=1)

history = model.fit(X_train, y_train, validation_data=(X_val, y_val),
                    epochs=500, batch_size=256, callbacks=[early_stop])
Epoch 1/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 3s 2ms/step - accuracy: 0.8158 - loss: 0.4543 - val_accuracy: 0.8425 - val_loss: 0.3840

Epoch 2/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8489 - loss: 0.3678 - val_accuracy: 0.8511 - val_loss: 0.3580

Epoch 3/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8553 - loss: 0.3492 - val_accuracy: 0.8584 - val_loss: 0.3417

Epoch 4/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8614 - loss: 0.3328 - val_accuracy: 0.8645 - val_loss: 0.3248

Epoch 5/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8661 - loss: 0.3180 - val_accuracy: 0.8644 - val_loss: 0.3175

Epoch 6/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8706 - loss: 0.3054 - val_accuracy: 0.8732 - val_loss: 0.3022

Epoch 7/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8741 - loss: 0.2967 - val_accuracy: 0.8718 - val_loss: 0.2994

Epoch 8/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8766 - loss: 0.2899 - val_accuracy: 0.8785 - val_loss: 0.2857

Epoch 9/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8794 - loss: 0.2837 - val_accuracy: 0.8797 - val_loss: 0.2852

Epoch 10/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.8817 - loss: 0.2787 - val_accuracy: 0.8827 - val_loss: 0.2767

Epoch 11/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.8827 - loss: 0.2742 - val_accuracy: 0.8847 - val_loss: 0.2717

Epoch 12/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.8848 - loss: 0.2709 - val_accuracy: 0.8848 - val_loss: 0.2716

Epoch 13/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.8860 - loss: 0.2670 - val_accuracy: 0.8873 - val_loss: 0.2648

Epoch 14/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8875 - loss: 0.2640 - val_accuracy: 0.8879 - val_loss: 0.2634

Epoch 15/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8888 - loss: 0.2612 - val_accuracy: 0.8895 - val_loss: 0.2611

Epoch 16/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8897 - loss: 0.2588 - val_accuracy: 0.8887 - val_loss: 0.2610

Epoch 17/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8909 - loss: 0.2568 - val_accuracy: 0.8920 - val_loss: 0.2577

Epoch 18/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8911 - loss: 0.2552 - val_accuracy: 0.8856 - val_loss: 0.2639

Epoch 19/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8926 - loss: 0.2532 - val_accuracy: 0.8931 - val_loss: 0.2507

Epoch 20/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8927 - loss: 0.2516 - val_accuracy: 0.8923 - val_loss: 0.2522

Epoch 21/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.8937 - loss: 0.2502 - val_accuracy: 0.8906 - val_loss: 0.2551

Epoch 22/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.8940 - loss: 0.2482 - val_accuracy: 0.8883 - val_loss: 0.2593

Epoch 23/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8948 - loss: 0.2474 - val_accuracy: 0.8942 - val_loss: 0.2470

Epoch 24/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.8956 - loss: 0.2455 - val_accuracy: 0.8929 - val_loss: 0.2491

Epoch 25/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8958 - loss: 0.2443 - val_accuracy: 0.8960 - val_loss: 0.2443

Epoch 26/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.8963 - loss: 0.2431 - val_accuracy: 0.8971 - val_loss: 0.2429

Epoch 27/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8971 - loss: 0.2410 - val_accuracy: 0.8993 - val_loss: 0.2396

Epoch 28/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8973 - loss: 0.2412 - val_accuracy: 0.8928 - val_loss: 0.2502

Epoch 29/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8981 - loss: 0.2394 - val_accuracy: 0.8972 - val_loss: 0.2410

Epoch 30/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8985 - loss: 0.2386 - val_accuracy: 0.8989 - val_loss: 0.2414

Epoch 31/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8989 - loss: 0.2374 - val_accuracy: 0.9003 - val_loss: 0.2356

Epoch 32/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.8997 - loss: 0.2359 - val_accuracy: 0.8991 - val_loss: 0.2381

Epoch 33/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9001 - loss: 0.2344 - val_accuracy: 0.8935 - val_loss: 0.2492

Epoch 34/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9008 - loss: 0.2338 - val_accuracy: 0.8956 - val_loss: 0.2444

Epoch 35/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9007 - loss: 0.2321 - val_accuracy: 0.8989 - val_loss: 0.2383

Epoch 36/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9015 - loss: 0.2311 - val_accuracy: 0.9013 - val_loss: 0.2329

Epoch 37/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9016 - loss: 0.2307 - val_accuracy: 0.9009 - val_loss: 0.2338

Epoch 38/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9026 - loss: 0.2288 - val_accuracy: 0.9031 - val_loss: 0.2300

Epoch 39/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9031 - loss: 0.2277 - val_accuracy: 0.9030 - val_loss: 0.2318

Epoch 40/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9035 - loss: 0.2271 - val_accuracy: 0.9035 - val_loss: 0.2290

Epoch 41/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9032 - loss: 0.2264 - val_accuracy: 0.9037 - val_loss: 0.2261

Epoch 42/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 3s 3ms/step - accuracy: 0.9039 - loss: 0.2254 - val_accuracy: 0.9039 - val_loss: 0.2264

Epoch 43/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 3s 3ms/step - accuracy: 0.9044 - loss: 0.2247 - val_accuracy: 0.9055 - val_loss: 0.2229

Epoch 44/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9047 - loss: 0.2239 - val_accuracy: 0.9041 - val_loss: 0.2281

Epoch 45/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9052 - loss: 0.2221 - val_accuracy: 0.9024 - val_loss: 0.2275

Epoch 46/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9055 - loss: 0.2226 - val_accuracy: 0.9061 - val_loss: 0.2212

Epoch 47/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9058 - loss: 0.2217 - val_accuracy: 0.9068 - val_loss: 0.2198

Epoch 48/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9070 - loss: 0.2190 - val_accuracy: 0.9063 - val_loss: 0.2217

Epoch 49/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9070 - loss: 0.2188 - val_accuracy: 0.9052 - val_loss: 0.2221

Epoch 50/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9076 - loss: 0.2175 - val_accuracy: 0.9083 - val_loss: 0.2196

Epoch 51/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9074 - loss: 0.2172 - val_accuracy: 0.9064 - val_loss: 0.2207

Epoch 52/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9078 - loss: 0.2163 - val_accuracy: 0.9097 - val_loss: 0.2149

Epoch 53/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9080 - loss: 0.2156 - val_accuracy: 0.9090 - val_loss: 0.2166

Epoch 54/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9088 - loss: 0.2148 - val_accuracy: 0.9106 - val_loss: 0.2128

Epoch 55/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9090 - loss: 0.2136 - val_accuracy: 0.9083 - val_loss: 0.2148

Epoch 56/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9097 - loss: 0.2119 - val_accuracy: 0.9109 - val_loss: 0.2118

Epoch 57/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9098 - loss: 0.2118 - val_accuracy: 0.9089 - val_loss: 0.2148

Epoch 58/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9096 - loss: 0.2119 - val_accuracy: 0.9020 - val_loss: 0.2269

Epoch 59/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9104 - loss: 0.2104 - val_accuracy: 0.9093 - val_loss: 0.2143

Epoch 60/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9107 - loss: 0.2090 - val_accuracy: 0.9087 - val_loss: 0.2157

Epoch 61/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9109 - loss: 0.2091 - val_accuracy: 0.9089 - val_loss: 0.2127

Epoch 62/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9111 - loss: 0.2079 - val_accuracy: 0.9111 - val_loss: 0.2079

Epoch 63/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9115 - loss: 0.2071 - val_accuracy: 0.9104 - val_loss: 0.2104

Epoch 64/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9119 - loss: 0.2062 - val_accuracy: 0.9129 - val_loss: 0.2063

Epoch 65/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9129 - loss: 0.2050 - val_accuracy: 0.9113 - val_loss: 0.2073

Epoch 66/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9125 - loss: 0.2054 - val_accuracy: 0.9125 - val_loss: 0.2038

Epoch 67/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9125 - loss: 0.2039 - val_accuracy: 0.9125 - val_loss: 0.2054

Epoch 68/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9139 - loss: 0.2026 - val_accuracy: 0.9106 - val_loss: 0.2075

Epoch 69/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9139 - loss: 0.2024 - val_accuracy: 0.9133 - val_loss: 0.2041

Epoch 70/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9139 - loss: 0.2018 - val_accuracy: 0.9137 - val_loss: 0.2027

Epoch 71/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9143 - loss: 0.2012 - val_accuracy: 0.9128 - val_loss: 0.2034

Epoch 72/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9145 - loss: 0.2015 - val_accuracy: 0.9147 - val_loss: 0.2013

Epoch 73/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9152 - loss: 0.1995 - val_accuracy: 0.9129 - val_loss: 0.2048

Epoch 74/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9149 - loss: 0.1995 - val_accuracy: 0.9110 - val_loss: 0.2068

Epoch 75/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9157 - loss: 0.1975 - val_accuracy: 0.9141 - val_loss: 0.2016

Epoch 76/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9153 - loss: 0.1978 - val_accuracy: 0.9153 - val_loss: 0.2000

Epoch 77/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9161 - loss: 0.1969 - val_accuracy: 0.9165 - val_loss: 0.1979

Epoch 78/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9153 - loss: 0.1977 - val_accuracy: 0.9104 - val_loss: 0.2087

Epoch 79/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9166 - loss: 0.1953 - val_accuracy: 0.9138 - val_loss: 0.2035

Epoch 80/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9165 - loss: 0.1966 - val_accuracy: 0.9124 - val_loss: 0.2052

Epoch 81/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9158 - loss: 0.1957 - val_accuracy: 0.9141 - val_loss: 0.1990

Epoch 82/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9170 - loss: 0.1941 - val_accuracy: 0.9124 - val_loss: 0.1990

Epoch 83/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9172 - loss: 0.1940 - val_accuracy: 0.9157 - val_loss: 0.1956

Epoch 84/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9174 - loss: 0.1938 - val_accuracy: 0.9151 - val_loss: 0.2020

Epoch 85/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9174 - loss: 0.1932 - val_accuracy: 0.9160 - val_loss: 0.1974

Epoch 86/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9182 - loss: 0.1915 - val_accuracy: 0.9174 - val_loss: 0.1946

Epoch 87/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9180 - loss: 0.1918 - val_accuracy: 0.9167 - val_loss: 0.1969

Epoch 88/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9185 - loss: 0.1916 - val_accuracy: 0.9160 - val_loss: 0.1945

Epoch 89/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9186 - loss: 0.1901 - val_accuracy: 0.9165 - val_loss: 0.1956

Epoch 90/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9181 - loss: 0.1914 - val_accuracy: 0.9169 - val_loss: 0.1941

Epoch 91/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9194 - loss: 0.1892 - val_accuracy: 0.9111 - val_loss: 0.2105

Epoch 92/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9187 - loss: 0.1893 - val_accuracy: 0.9190 - val_loss: 0.1924

Epoch 93/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9194 - loss: 0.1886 - val_accuracy: 0.9147 - val_loss: 0.1967

Epoch 94/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9194 - loss: 0.1884 - val_accuracy: 0.9151 - val_loss: 0.1954

Epoch 95/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9192 - loss: 0.1881 - val_accuracy: 0.9190 - val_loss: 0.1901

Epoch 96/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9195 - loss: 0.1872 - val_accuracy: 0.9168 - val_loss: 0.1932

Epoch 97/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9195 - loss: 0.1877 - val_accuracy: 0.9152 - val_loss: 0.1951

Epoch 98/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9204 - loss: 0.1864 - val_accuracy: 0.9152 - val_loss: 0.1942

Epoch 99/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9203 - loss: 0.1857 - val_accuracy: 0.9165 - val_loss: 0.1931

Epoch 100/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9204 - loss: 0.1858 - val_accuracy: 0.9170 - val_loss: 0.1963

Epoch 101/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9205 - loss: 0.1855 - val_accuracy: 0.9160 - val_loss: 0.1929

Epoch 102/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9212 - loss: 0.1846 - val_accuracy: 0.9174 - val_loss: 0.1902

Epoch 103/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9210 - loss: 0.1845 - val_accuracy: 0.9172 - val_loss: 0.1909

Epoch 104/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9211 - loss: 0.1838 - val_accuracy: 0.9184 - val_loss: 0.1913

Epoch 105/500

864/864 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step - accuracy: 0.9212 - loss: 0.1844 - val_accuracy: 0.9181 - val_loss: 0.1905

Epoch 105: early stopping

Restoring model weights from the end of the best epoch: 95.

Model Evaluation

loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(len(loss))

plt.plot(epochs, loss)
plt.plot(epochs, val_loss)
plt.legend(['loss', 'val_loss'], loc='upper right')
plt.title('Train Loss vs Val Loss')
plt.show()

  • Learning Behavior: Both training loss and validation loss exhibit a smooth, steady decline from around 0.45 down to ~0.18 over 100+ epochs.

  • Generalization Capability: The validation loss curve closely tracks the training loss curve without diverging upward. This indicates that the model generalizes remarkably well to unseen data and shows no signs of overfitting.

  • Convergence: The minor fluctuations in val_loss are typical in mini-batch gradient descent. Overall, the model reaches a stable convergence around epoch 80–100.

# remember always make sure that the X_train/test and y_train/test are in float32 datatype
# and y_train/test is in form of 1 column
X_test = X_test_scaled_df.values.astype(dtype=np.float32)
y_test = y_test.reshape(-1, 1).astype(dtype=np.float32)
model.evaluate(X_test, y_test)
864/864 ━━━━━━━━━━━━━━━━━━━━ 1s 1ms/step - accuracy: 0.9218 - loss: 0.1833

[0.1833219975233078, 0.9217765927314758]
y_pred_softprob = model(X_test)
y_pred = np.argmax(y_pred_softprob, axis=-1)
print(y_pred)
[0 0 1 ... 0 1 0]
from sklearn.metrics import classification_report, confusion_matrix, ConfusionMatrixDisplay

print("--- Classification Report ---")
report = classification_report(y_test, y_pred)
print('Classification Report:\n')
print(report)

print("--- Confusion Matrix ---")
cm = confusion_matrix(y_test, y_pred)
disp = ConfusionMatrixDisplay(confusion_matrix=cm)
disp.plot()
--- Classification Report ---
Classification Report:

              precision    recall  f1-score   support

         0.0       0.93      0.94      0.93      8211
         1.0       0.91      0.92      0.91     10511
         2.0       0.89      0.86      0.87      4632
         3.0       0.97      0.96      0.97      4272

    accuracy                           0.92     27626
   macro avg       0.93      0.92      0.92     27626
weighted avg       0.92      0.92      0.92     27626

--- Confusion Matrix ---

The model achieves an overall accuracy of 92% across a test set of 27,626 samples, with both macro F1-score (0.92) and weighted F1-score (0.92) reflecting high performance across all classes.

The primary source of error occurs between Class 2 and Class 1, where 580 true Class 2 samples were misclassified as Class 1, and 445 true Class 1 samples were misclassified as Class 2. This suggests some feature overlap between these two categories.

Conclusion:

The Multi-Layer Perceptron (MLP) model demonstrates robust learning dynamics and high predictive capability across all metrics. Despite the initial class imbalance, the high macro-average F1-score (0.92) confirms that minority classes (such as Class 2 and Class 3) are classified with strong accuracy and reliability.

Save Model

model.save('model/spotify_mlp.keras')