ニューラルネットの精度を上げたい

Question

前提・実現したいこと

python、ニューラルネット初心者です。

python3を用いてニューラルネットを構築しています。
CSVデータに保存したデータを使って、学習を行い未知のデータに対して10段階の分類を行うことが最終目標です。
最終的な精度が30～40％と低いままで向上しないため、精度が上がるようなアドバイスをいただきたいです。

回答にあたって、必要な情報が欠けていましたら補足いたしますので、教えていただけますと幸いです。

発生している問題・エラーメッセージ

Epoch 1/200
237/237 [==============================] - 1s 2ms/step - loss: 0.1512 - accuracy: 0.3215 - val_loss: 0.1505 - val_accuracy: 0.3241

...

Epoch 200/200
237/237 [==============================] - 0s 1ms/step - loss: 0.1440 - accuracy: 0.3781 - val_loss: 0.1507 - val_accuracy: 0.3141
13/13 [==============================] - 0s 795us/step - loss: 0.1507 - accuracy: 0.3141

Test loss: 0.15066926181316376
Test accuracy: 0.31407034397125244

該当のソースコード

from __future__ import print_function
import csv

import pandas as pd
from pandas import Series,DataFrame

from sklearn import svm
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score

import numpy as np
import matplotlib.pyplot as plt

import tensorflow as tf

from tensorflow import keras
from tensorflow.keras import models
from tensorflow.keras import optimizers
from tensorflow.keras.datasets import fashion_mnist
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense, Dropout
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input


#CSVファイルの読み込み
data_set = pd.read_csv("dataset_new2.csv",sep=",",header=0)
#説明変数（データセットの入力）
x = DataFrame(data_set.drop("result",axis=1))

#目的変数（得点の10段階評価）
y = DataFrame(data_set["result"])


#説明変数・目的変数をそれぞれ訓練データ・テストデータに分割
x_train,x_test,y_train,y_test = train_test_split(x,y,test_size=0.05)

#データの整形
x_train = x_train.astype(np.float)
x_test = x_test.astype(np.float)

y_train = tf.keras.utils.to_categorical(y_train,5)
y_test = tf.keras.utils.to_categorical(y_test,5)

#ニューラルネットワークの実装①
model = tf.keras.Sequential()

model.add(Dense(50, activation='relu', input_shape=(14,)))
model.add(Dropout(0.2))

model.add(Dense(50, activation='relu', input_shape=(14,)))
model.add(Dropout(0.2))

model.add(Dense(50, activation='relu', input_shape=(14,)))
model.add(Dropout(0.2))

model.add(Dense(5, activation='softmax'))

model.summary()
print("\n")

#ニューラルネットワークの実装②
model.compile(loss='mean_squared_error',optimizer=tf.keras.optimizers.Adam(
    learning_rate=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-07, amsgrad=False,
    name='Adam'),metrics=['accuracy'])

# optimizer=tf.keras.optimizers.RMSprop(
#     learning_rate=0.001, rho=0.9, momentum=0.0, epsilon=1e-07, centered=False,
#     name='RMSprop'),metrics=['accuracy'])



#ニューラルネットワークの学習
history = model.fit(x_train, y_train,batch_size=32,epochs=200,verbose=1,validation_data=(x_test, y_test))

#ニューラルネットワークの推論
score = model.evaluate(x_test,y_test,verbose=1)
print("\n")
print("Test loss:",score[0])
print("Test accuracy:",score[1])

該当のデータセット

year,day,time,num,sup_x,sup_y,sup_s,sup_t,em_x,em_y,em_s,em_t,hm_aw,capacty,result
0.994,0.603,0.924,0.506,0.196,0.248,0.194,0.36,0.192,0.126,0.164,0.316,0,0.691,0
0.994,0.603,0.924,0.271,0.092,0.128,0.171,0.124,0.368,0.841,0.486,0.558,0,0.235,1
0.994,0.603,0.949,0.233,0.279,0.168,0.343,0.363,0.124,0.129,0.166,0.165,0,0.29,2
0.994,0.603,0.949,0.333,0.155,0.161,0.16,0.168,0.263,0.13,0.26,0.343,0,0.553,1
0.994,0.603,0.95,0.228,0.203,0.198,0.244,0.349,0.187,0.109,0.171,0.257,0,0.563,2
0.994,0.603,0.95,0.398,0.195,0.193,0.225,0.254,0.313,0.187,0.239,0.417,0,0.704,2
0.994,0.604,0.949,0.271,0.212,0.188,0.258,0.321,0.363,0.201,0.182,0.517,0,0.273,2
0.994,0.604,0.949,0.384,0.406,0.275,0.215,0.494,0.155,0.33,0.17,0.279,0,1,2
0.994,0.672,0.896,0.169,0.192,0.126,0.164,0.316,0.195,0.193,0.225,0.254,0,0.51,1
0.994,0.672,0.946,0.662,0.155,0.33,0.17,0.279,0.212,0.188,0.258,0.321,0,0.585,1
0.994,0.672,0.946,0.201,0.187,0.109,0.171,0.257,0.237,0.084,0.163,0.278,0,0.661,0
0.994,0.672,0.947,0.199,0.363,0.201,0.182,0.517,0.092,0.128,0.171,0.124,0,0.373,4
0.994,0.672,0.949,0.801,0.368,0.841,0.486,0.558,0.196,0.248,0.194,0.36,0,0.881,4
0.994,0.672,0.949,0.322,0.263,0.13,0.26,0.343,0.406,0.275,0.215,0.494,0,0.363,1
0.994,0.672,0.949,0.485,0.313,0.187,0.239,0.417,0.203,0.198,0.244,0.349,0,0.704,3
0.994,0.672,0.949,0.257,0.233,0.111,0.177,0.239,0.279,0.168,0.343,0.363,0,0.285,1
0.994,0.672,0.949,0.291,0.124,0.129,0.166,0.165,0.155,0.161,0.16,0.168,0,0.298,0
0.994,0.678,0.946,0.23,0.363,0.201,0.182,0.517,0.313,0.187,0.239,0.417,0,0.373,3
0.994,0.678,0.946,0.166,0.124,0.129,0.166,0.165,0.233,0.111,0.177,0.239,0,0.298,4
0.994,0.678,0.948,0.297,0.187,0.109,0.171,0.257,0.406,0.275,0.215,0.494,0,0.661,2
0.994,0.678,0.949,0.547,0.203,0.198,0.244,0.349,0.368,0.841,0.486,0.558,0,0.563,2
0.994,0.678,0.949,0.181,0.237,0.084,0.163,0.278,0.279,0.168,0.343,0.363,0,0.297,0
0.994,0.678,0.949,0.616,0.155,0.33,0.17,0.279,0.263,0.13,0.26,0.343,0,0.585,2
0.994,0.678,0.949,0.246,0.195,0.193,0.225,0.254,0.092,0.128,0.171,0.124,0,0.28,4
0.994,0.679,0.9,0.162,0.192,0.126,0.164,0.316,0.155,0.161,0.16,0.168,0,0.51,0
0.994,0.679,0.949,0.276,0.212,0.188,0.258,0.321,0.196,0.248,0.194,0.36,0,0.273,3
0.994,0.681,0.947,0.181,0.263,0.13,0.26,0.343,0.363,0.201,0.182,0.517,0,0.363,4
0.994,0.681,0.948,0.191,0.313,0.187,0.239,0.417,0.212,0.188,0.258,0.321,0,0.21,3
0.994,0.681,0.949,0.55,0.368,0.841,0.486,0.558,0.155,0.33,0.17,0.279,0,0.881,3
0.994,0.681,0.949,0.296,0.196,0.248,0.194,0.36,0.124,0.129,0.166,0.165,0,0.664,4
0.994,0.681,0.949,0.193,0.092,0.128,0.171,0.124,0.203,0.198,0.244,0.349,0,0.235,2
0.994,0.681,0.949,0.18,0.233,0.111,0.177,0.239,0.187,0.109,0.171,0.257,0,0.285,4
0.994,0.681,0.949,0.285,0.279,0.168,0.343,0.363,0.192,0.126,0.164,0.316,0,0.29,3
0.994,0.681,0.949,0.301,0.155,0.161,0.16,0.168,0.195,0.193,0.225,0.254,0,0.553,3
0.994,0.681,0.95,0.186,0.406,0.275,0.215,0.494,0.237,0.084,0.163,0.278,0,0.214,1
0.994,0.684,0.899,0.145,0.237,0.084,0.163,0.278,0.263,0.13,0.26,0.343,0,0.213,1
0.994,0.684,0.921,0.224,0.203,0.198,0.244,0.349,0.192,0.126,0.164,0.316,0,0.563,0
0.994,0.684,0.924,0.227,0.092,0.128,0.171,0.124,0.313,0.187,0.239,0.417,0,0.235,1
0.994,0.684,0.946,0.283,0.124,0.129,0.166,0.165,0.212,0.188,0.258,0.321,0,0.298,1
0.994,0.684,0.949,0.646,0.155,0.33,0.17,0.279,0.155,0.161,0.16,0.168,0,0.585,3
0.994,0.684,0.949,0.258,0.195,0.193,0.225,0.254,0.196,0.248,0.194,0.36,0,0.28,2
0.994,0.684,0.95,0.373,0.187,0.109,0.171,0.257,0.368,0.841,0.486,0.558,0,0.661,1
0.994,0.685,0.949,0.142,0.233,0.111,0.177,0.239,0.406,0.275,0.215,0.494,0,0.285,0
0.994,0.685,0.95,0.28,0.363,0.201,0.182,0.517,0.279,0.168,0.343,0.363,0,0.373,3
0.994,0.687,0.921,0.183,0.212,0.188,0.258,0.321,0.263,0.13,0.26,0.343,0,0.273,1
0.994,0.687,0.945,0.151,0.203,0.198,0.244,0.349,0.363,0.201,0.182,0.517,0,0.563,2
0.994,0.687,0.946,0.148,0.195,0.193,0.225,0.254,0.233,0.111,0.177,0.239,0,0.28,1
0.994,0.687,0.949,0.223,0.279,0.168,0.343,0.363,0.155,0.33,0.17,0.279,0,0.29,3
0.994,0.687,0.949,0.078,0.237,0.084,0.163,0.278,0.124,0.129,0.166,0.165,0,0.297,2
0.994,0.687,0.949,0.336,0.196,0.248,0.194,0.36,0.187,0.109,0.171,0.257,0,0.664,2
0.994,0.687,0.949,0.152,0.192,0.126,0.164,0.316,0.313,0.187,0.239,0.417,0,0.51,2
0.994,0.687,0.949,0.445,0.155,0.161,0.16,0.168,0.368,0.841,0.486,0.558,0,0.553,2
0.994,0.687,0.95,0.198,0.406,0.275,0.215,0.494,0.092,0.128,0.171,0.124,0,0.214,3
0.994,0.752,0.8,0.364,0.406,0.275,0.215,0.494,0.263,0.13,0.26,0.343,0,1,1

試したこと

バッチサイズの変更、学習率の変更、最適化アルゴリズムの変更、学習回数の変更、ユニット数の変更を行いましたが、精度が大きく向上することはありませんでした。

データセットが不均衡なデータになっていたため、サンプリングを行い、少数派のデータが占める割合を増やしましたが、精度が向上することはありませんでした。
サンプリングを行った場合、10段階の分類→5段階の分類に変更しました。

データセットの正規化を行いましたが、精度が向上することはありませんでした。

公開されているライブラリのデータセットを用いて学習を行うと90％以上の成果が得られました。

補足情報（FW/ツールのバージョンなど）

python3
tensorflow 2.6.2
keras 2.6.0

使用しているデータセットはこちらです。
戯画ファイル便を使用しての共有になります。
https://13.gigafile.nu/1216-c23c9e68e2185c05ada3cb58098943c01

Comments

[jbpb0]

> 公開されているライブラリのデータセットを用いて学習を行うと90％以上の成果が得られました。

それならデータ依存の現象だから、データをもっとたくさん開示してくれないと、他人には何が起きてるのか分かりません
データ全部とは言いませんが、他人が学習させて現象を再現できる程度のデータ量は要ります

[jbpb0]

「input_shape=(14,)」が3ヶ所にありますけど、一番最初の以外は要らないのでは？
(未確認ですが)

[hate_tometo]

ご回答ありがとうございます。

データの開示についてですが、字数制限以内（10000字）でできる限りのデータを載せました。
必要な情報の不足申し訳ありません。

ご指摘いただいた、input_shapeについてですが、ご指摘通りの修正を行ったところ動作が確認できましたので、不必要な入力でした。
ありがとうございます。

[hate_tometo]

ギガファイル便での共有でしたら、可能でしたのでURLを添付いたしました。