神经网络的票房预测CNN-RNN-LSTM

1、摘要

本文主要讲解：神经网络的票房预测，三种实现方式CNN、RNN、LSTM
主要思路：

1、用不少于3种神经网络构建电影票房预测模型，并在同一个数据集上进行训练和预测，对比各个模型的准确率，并可视化。（使用CNN、RNN、LSTM）

2、对数据预处理并对数据进行可视化：分析预算对票房的影响、观众打分对票房影响、流行系数对票房影响、预算和观众打分对票房的影响，电影语言对票房的影响、票房的正态分布。

3、对神经网络结构、多个模型比对、训练过程进行可视化

4、特征分析

2、数据介绍

训练的数据集：Kaggle的movie数据集

其中主要特征有：

Budget：预算 
Revenue：票房 
Rating：观众打分 
totalVotes：观众打分数量 
popularity：流行系数

3、相关技术

CNN：卷积神经网络

RNN：循环神经网络

LSTM：长短期记忆网络，对于长期记忆的时间序列效果比RNN好

4、部分代码和步骤

此代码的依赖环境如下：

numpy==1.19.5
matplotlib==3.5.2
torch==1.9.0
pandas==1.1.5

代码输出如下：
在这里插入图片描述

主运行程序入口

import os
import warnings
from datetime import datetime

import matplotlib.pyplot as plt
import numpy as np
import optuna
import pandas as pd
import seaborn as sns
import torch
from sklearn.model_selection import KFold
from torch import nn

# 这两行代码解决 plt 中文显示的问题
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False

# fix random seed  设定随机种子，防止每次训练的结果不一样
def same_seeds(seed):
    torch.manual_seed(seed)  # 固定随机种子（CPU）
    if torch.cuda.is_available():  # 固定随机种子（GPU)
        torch.cuda.manual_seed(seed)  # 为当前GPU设置
        torch.cuda.manual_seed_all(seed)  # 为所有GPU设置
    np.random.seed(seed)  # 保证后续使用random函数时，产生固定的随机数
    torch.backends.cudnn.benchmark = False  # GPU、网络结构固定，可设置为True
    torch.backends.cudnn.deterministic = True  # 固定网络结构


same_seeds(22)
np.random.seed(22)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
warnings.filterwarnings("ignore")
os.chdir(r'C:\Projects\old\电影数据集多模型对比\data')
'''
CNN  RNN  LSTM
对神经网络结构、多模型对比、训练过程进行可视化
预测值和真实值误差小于0.1，每个模型预测的准确率0.8以上，至少一个0.9
特征分析
'''
# 迭代次数
# epoch_time = 5


# 文中提到的双曲线性单元激活函数
class HLU(nn.Module):
    def __init__(self):
        super(HLU, self).__init__()

    def forward(self, x, alpha=0.15):
        inputs_m = torch.where(x < 0, x, torch.zeros_like(x))
        return torch.where(x >= 0.0, x, inputs_m * alpha / (1 - inputs_m))


def cnn_model(X_train, y_train, X_val, y_val, X_test):
    class Model(nn.Module):
        def __init__(self):
            super().__init__()
            self.layer1 = nn.Sequential(
                nn.Conv1d(X_train.shape[1], 56, 1, 1),
                nn.BatchNorm1d(num_features=56),
                HLU(),
                nn.Dropout(0.29),
                nn.Conv1d(56, 28, 1, 2),
                nn.BatchNorm1d(num_features=28),
                HLU(),
                nn.Dropout(0.29),
                nn.Conv1d(28, 14, 1, 2),
                nn.BatchNorm1d(num_features=14),
                HLU(),
                nn.Dropout(0.29),
                nn.Flatten(),
                nn.Linear(14, 1)
            )

        def forward(self, x):
            x = self.layer1(x)
            return x

    model = Model()
    # img = tw.draw_model(model, [56, 198, 1])
    # img.save(r'CNN.jpg')
    # 学习率
    learn_rate = 0.2346
    loss_function = torch.nn.L1Loss()
    # 优化器，变量依次为待优化参数。学习率、动量
    optim = torch.optim.SGD(model.parameters(), learn_rate, momentum=0.5)
    # 训练集前固定写法（不写也无妨）
    model.train()
    loss_list = []
    for epoch in range(41):
        output = model(X_train)
        # 计算损失
        res_loss = loss_function(y_train, output)
        # 清零梯度
        optim.zero_grad()
        # 反向传播
        res_loss.backward()
        # 更新参数
        optim.step()
        loss_list.append(res_loss)
        print(f"训练batch:{i + 1},损失值:{res_loss}")
    plt.plot(loss_list)
    plt.title('CNN训练loss变化图')
    plt.ylabel('mae')
    plt.xlabel('Epoch')
    plt.legend(['mae'], loc='best')
    plt.savefig("CNN训练loss变化图.png", dpi=500, bbox_inches='tight')
    plt.show()
    # 测试集固定写法
    model.eval()
    # 测试集不需要梯度下降，加快计算效率
    with torch.no_grad():
        output = model(X_val)
        mae = loss_function(y_val, output)
    val_loss = mae.item()
    print(f"预测值和真实值误差:{val_loss}")
    return val_loss


def rnn_model(X_train, y_train, X_val, y_val, X_test):
    plt.figure(1, figsize=(30, 10))
    plt.ion()  # continuously plot

    class RNN(nn.Module):
        def __init__(self):
            super(RNN, self).__init__()

            self.rnn = nn.RNN(
                input_size=1,
                hidden_size=82,  # rnn hidden unit
                num_layers=1,  # number of rnn layer
                batch_first=True,
                dropout=0.29
                # input & output will has batch size as 1s dimension. e.g. (batch, time_step, input_size)
            )
            self.out = nn.Linear(82, 1)

        def forward(self, x):
            r_out, _ = self.rnn(x)
            r_out = r_out.reshape(-1, 82)
            r_out = self.out(r_out)
            return r_out

    batch_size = 16
    rnn = RNN()
    # img = tw.draw_model(rnn, [1, 83, 1])
    # img.save(r'RNN.jpg')
    rnn.to(device)
    rnn.train()
    optimizer = torch.optim.Adam(rnn.parameters(), lr=0.2346)  # optimize all cnn parameters
    loss_function = nn.L1Loss()
    length = len(X_train)
    loss_list = []
    for step in range(41):
        for j in range(0, length, batch_size):
            X_train_batch = X_train[j:j + batch_size]
            y_train_batch = y_train[j:j + batch_size]
            X_train_batch = X_train_batch.to(device)
            y_train_batch = y_train_batch.to(device)
            y_pred = rnn(X_train_batch)
            loss = loss_function(y_pred, y_train_batch)
            optimizer.zero_grad()  # clear gradients for this training step
            loss.backward()  # backpropagation, compute gradients
            optimizer.step()
        loss_list.append(loss)
        print(f"训练batch:{step + 1},损失值:{loss}")
    plt.plot(loss_list)
    plt.title('RNN训练loss变化图')
    plt.ylabel('mae')
    plt.xlabel('Epoch')
    plt.legend(['mae'], loc='best')
    plt.savefig("RNN训练loss变化图.png", dpi=500, bbox_inches='tight')
    plt.show()
    # test
    rnn.eval()
    vl = 0.0
    length = len(X_val)
    with torch.no_grad():
        for j in range(0, length, batch_size):
            X_val_batch = X_val[j:j + batch_size]
            y_val_batch = y_val[j:j + batch_size]
            X_val_batch = X_val_batch.to(device)
            y_val_batch = y_val_batch.to(device)
            output = rnn(X_val_batch)
            # 计算损失
            mae = loss_function(y_val_batch, output)
            vl += mae.item()
        batch_count = len(X_val) // batch_size
    val_loss = vl / batch_count
    print(f"预测值和真实值误差:{val_loss}")
    return val_loss


def objective(trial):
    hidden_size = trial.suggest_int('hidden_size', 60, 100)
    epoch_time = trial.suggest_int('epoch_time', 33, 55)
    batch_size = trial.suggest_int('batch_size', 16, 16)
    dropout = trial.suggest_uniform('dropout', 0.28, 0.3)
    lr = trial.suggest_uniform('lr', 0.23, 0.25)
    num_layers = trial.suggest_categorical('num_layers', [1])
    print([hidden_size, epoch_time, batch_size, dropout, lr, num_layers])
    input_size, hidden_size, num_layers, output_size = 1, hidden_size, num_layers, 1
    model = LSTM(input_size, hidden_size, num_layers, output_size, dropout)
    model = model.to(device)
    loss_function = nn.L1Loss()
    X_train_p = pd.read_csv('./X_train.csv')
    y_train_p = pd.read_csv('./y_train.csv')
    data = pd.concat([X_train_p, y_train_p], axis=1)
    data = data[~data.isin([np.nan, np.inf, -np.inf]).any(1)]
    data.reset_index(inplace=True)
    X_train_p = data.iloc[:, :-1]
    y_train_p = pd.DataFrame(data.iloc[:, -1])
    k = 2
    fold = list(KFold(k, shuffle=True, random_state=random_seed).split(X_train_p))
    maes = []
    for i, (train, val) in enumerate(fold):
        X_train = X_train_p.loc[train, :].values
        y_train = y_train_p.loc[train, :].values
        X_val = X_train_p.loc[val, :].values
        y_val = y_train_p.loc[val, :].values.ravel()
        X_train = torch.FloatTensor(X_train.reshape((X_train.shape[0], X_train.shape[1], 1)))
        y_train = torch.FloatTensor(y_train.ravel())
        X_val = torch.FloatTensor(X_val.reshape((X_val.shape[0], X_val.shape[1], 1)))
        y_val = torch.FloatTensor(y_val)
        length = len(X_train)
        model.train()
        optimizer = torch.optim.SGD(model.parameters(), lr=lr)
        for i in range(epoch_time):
            for j in range(0, length, batch_size):
                X_train_batch = X_train[j:j + batch_size]
                y_train_batch = y_train[j:j + batch_size]
                X_train_batch = X_train_batch.to(device)
                y_train_batch = y_train_batch.to(device)
                y_pred = model(X_train_batch)
                loss = loss_function(y_pred, y_train_batch)
                optimizer.zero_grad()
                loss.backward()
                optimizer.step()
        model.eval()
        length = len(X_val)
        X_val = X_val.to(device)
        y_val = y_val.to(device)
        # 测试集不需要梯度下降，加快计算效率
        vl = 0.0
        with torch.no_grad():
            for j in range(0, length, batch_size):
                X_val_batch = X_val[j:j + batch_size]
                y_val_batch = y_val[j:j + batch_size]
                X_val_batch = X_val_batch.to(device)
                y_val_batch = y_val_batch.to(device)
                output = model(X_val_batch)
                # 计算损失
                mae = loss_function(y_val_batch, output)
                vl += mae.item()
            batch_count = len(X_val) // batch_size
        val_loss = vl / batch_count
        maes.append(val_loss)
    mae = np.array(maes).mean()
    return mae


X_train_p = pd.read_csv('./X_train.csv')
X_test = pd.read_csv('./X_test.csv')
y_train_p = pd.read_csv('./y_train.csv')
data = pd.concat([X_train_p, y_train_p], axis=1)
data = data[~data.isin([np.nan, np.inf, -np.inf]).any(1)]
data.reset_index(inplace=True)

X_train_p = data.iloc[:, :-1]
y_train_p = pd.DataFrame(data.iloc[:, -1])
random_seed = 2019
k = 2
fold = list(KFold(k, shuffle=True, random_state=random_seed).split(X_train_p))
np.random.seed(random_seed)

for i, (train, val) in enumerate(fold):
    print(i + 1, "fold")

    X_train = X_train_p.loc[train, :].values
    y_train = y_train_p.loc[train, :].values
    X_val = X_train_p.loc[val, :].values
    y_val = y_train_p.loc[val, :].values.ravel()
    mean_y_val = y_val.mean()
    X_train = torch.FloatTensor(X_train.reshape((X_train.shape[0], X_train.shape[1], 1)))
    y_train = torch.FloatTensor(y_train.ravel())
    X_val = torch.FloatTensor(X_val.reshape((X_val.shape[0], X_val.shape[1], 1)))
    y_val = torch.FloatTensor(y_val)

    # """ cnn_model
    start = datetime.now()
    cnn_mae = cnn_model(X_train, y_train, X_val, y_val, X_test)
    print("cnn_model mae ", "{0:.5f}".format(cnn_mae),
          '(' + str(int((datetime.now() - start).seconds)) + 's)')
    cnn_acc = round(1 - cnn_mae / mean_y_val, 3)
    print('cnn_model 准确率:', cnn_acc)

    # #
    # # # """ rnn_model
    start = datetime.now()
    rnn_mae = rnn_model(X_train, y_train, X_val, y_val, X_test)
    print("rnn_model mae.", "{0:.5f}".format(rnn_mae),
          '(' + str(int((datetime.now() - start).seconds)) + 's)')
    rnn_acc = round(1 - rnn_mae / mean_y_val, 3)
    print('rnn_model 准确率:', rnn_acc)


    # """ lstm_model
    class LSTM(nn.Module):
        def __init__(self, input_size, hidden_size, num_layers, output_size, dropout):
            super().__init__()
            self.input_size = input_size
            self.hidden_size = hidden_size

            self.num_layers = num_layers
            self.output_size = output_size
            self.lstm = nn.LSTM(self.input_size, self.hidden_size, self.num_layers, dropout=dropout,
                                batch_first=True)
            self.linear = nn.Linear(self.hidden_size, self.output_size)

        def forward(self, input_seq):
            x, _ = self.lstm(input_seq)
            x = x.reshape(-1, self.hidden_size)
            pred = self.linear(x)
            return pred


    """ Run optimize. 
    Set n_trials and/or timeout (in sec) for optimization by Optuna
    """
    study = optuna.create_study(direction='minimize')
    study.optimize(objective, n_trials=11)
    print('Best trial number: ', study.best_trial.number)
    print('Best value:', study.best_trial.value)
    print('Best parameters: \n', study.best_trial.params)
    parameters = study.best_trial.params
    # parameters = {'hidden_size': 82, 'epoch_time': 41, 'batch_size': 16, 'dropout': 0.2904, 'lr': 0.2346,
    #               'num_layers': 1}
    hidden_size = parameters['hidden_size']
    batch_size = parameters['batch_size']
    dropout = parameters['dropout']
    lr = parameters['lr']
    num_layers = parameters['num_layers']
    epoch_time = parameters['epoch_time']
    input_size, hidden_size, num_layers, output_size = 1, hidden_size, num_layers, 1
    model = LSTM(input_size, hidden_size, num_layers, output_size, dropout)
    # img = tw.draw_model(model,  [1, 82, 1])
    # img.save(r'LSTM.jpg')
    model = model.to(device)
    loss_function = nn.L1Loss().to(device)
    optimizer = torch.optim.Adam(model.parameters(), lr=lr)
    model.train()
    loss_list = []
    for i in range(epoch_time):
        for j in range(0, len(X_train), batch_size):
            X_train_batch = X_train[j:j + batch_size]
            y_train_batch = y_train[j:j + batch_size]
            X_train_batch = X_train_batch.to(device)
            y_train_batch = y_train_batch.to(device)
            y_pred = model(X_train_batch)
            loss = loss_function(y_pred, y_train_batch)
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
        loss_list.append(loss)
        print(f"训练batch:{i + 1},损失值:{loss}")
    plt.plot(loss_list)
    plt.title('LSTM训练loss变化图')
    plt.ylabel('mae')
    plt.xlabel('Epoch')
    plt.legend(['mae'], loc='best')
    plt.savefig("LSTM训练loss变化图.png", dpi=500, bbox_inches='tight')
    plt.show()
    model.eval()
    length = len(X_val)
    X_val = X_val.to(device)
    y_val = y_val.to(device)
    # 测试集不需要梯度下降，加快计算效率
    vl = 0.0
    with torch.no_grad():
        for j in range(0, length, batch_size):
            X_val_batch = X_val[j:j + batch_size]
            y_val_batch = y_val[j:j + batch_size]
            X_val_batch = X_val_batch.to(device)
            y_val_batch = y_val_batch.to(device)
            output = model(X_val_batch)
            # 计算损失
            mae = loss_function(y_val_batch, output)
            vl += mae.item()
        batch_count = len(X_val) // batch_size
    lstm_mae = vl / batch_count
    print(f"预测值和真实值误差:{lstm_mae}")
    lstm_acc = round(1 - lstm_mae / mean_y_val, 3)
    print('lstm_model 准确率:', lstm_acc)


def labels(ax):
    for p in ax.patches:
        width = p.get_width()  # get bar length
        ax.text(width,  # set the text at 1 unit right of the bar
                p.get_y() + p.get_height() / 2,  # get Y coordinate + X coordinate / 2
                '{:1.3f}'.format(width),  # set variable to display, 2 decimals
                ha='left',  # horizontal alignment
                va='center')  # vertical alignment


# 模型比较
compare = pd.DataFrame({"Model": ["CNN", "RNN", "LSTM"],
                        "MAE": [cnn_mae, rnn_mae, lstm_mae],
                        "Accuracy": [cnn_acc, rnn_acc, lstm_acc]})
plt.figure(figsize=(14, 14))
plt.subplot(211)
compare = compare.sort_values(by="MAE", ascending=False)
ax = sns.barplot(x="MAE", y="Model", data=compare, palette="Blues_d")
labels(ax)
plt.subplot(212)
compare = compare.sort_values(by="Accuracy", ascending=False)
ax = sns.barplot(x="Accuracy", y="Model", data=compare, palette="Blues_d")
labels(ax)
plt.show()

完整数据和代码链接：
神经网络的票房预测CNN-RNN-LSTM