使用长短期记忆网络预测NBA比赛胜负

Predicting outcome of NBA games using long short-term memory network

深度学习与机器学习的方法已广泛应用于NBA(美国篮球职篮联赛)的比赛胜负的预测中,然而过去的方法未对过去几场比赛的数据进行建模,忽略了比赛双方近期状态的有效表示。为了解决这个问题,提出了基于长短期记忆(LSTM)网络的方法对NBA常规赛的比赛胜负进行预测。该方法分别以比赛中的两支球队过去几场比赛的数据作为LSTM的输入,以该场比赛结果作为输出,训练能够预测比赛胜负的模型。本质上是使用球队在该赛季的历史数据的平均值作为该球队的实力,以近几场比赛的数据序列作为该球队状态的体现。在实验中比较了其他几种预测NBA比赛胜负的方法(支持向量机、卷积神经网络、逻辑回归模型等方法),数据来自2014-2019年间的5个赛季的NBA常规赛数据。结果表明,模型的预测准确率达到(69.09%),高于其他几种模型。

Deep learning and machine learning models have been widely used in the prediction of NBA(National Basketball League)game results. However,the past methods lack modeling of past games data,thus ignoring the effective representation of the recent state of the two teams of the game. In order to solve this problem,a model based on LSTM(Long Short-Term Memory)network was proposed to predict the outcome of the NBA regular season game. The data of past few games of the two teams in the game was used as the input of LSTM,and the result of the game was used as the output to train a model that can predict the outcome of the game. Essentially,the average value of the team ’s historical data during the season was used as the strength of the team,and the data sequence of the recent gameswas used as the embodiment of the team ’ s state. The LSTM can better represent the recent state of the team. In the experiment,several other methods of predicting the outcome of NBA games(support vector machine,convolutional neural network,logistic regression model,etc.)were compared with the proposed model. The data came from the NBA regular season data for the five seasons from2014 to 2019. The results show that the prediction accuracy of the proposed model reaches 69. 09%,which is higher than those of other models.