基於對比學習與注意力機制之足球先發陣容推薦模型

Abstract

對足球隊而言，決定比賽的先發陣容是一項關鍵但複雜的任務。在本研究中，我們提出了一個基於對比學習與注意力機制的足球先發陣容推薦模型(CASTRec)，以協助球隊建構合適的先發陣容。首先，我們使用四個模組提取四個類別的特徵，分別是球員近期的先發狀態、球員與敵隊球員的互動、球員和同隊隊友的互動、團隊的整體表現。具體來說，我們利用圖表示球員間的互動及球隊間的對戰關係，再藉由圖注意力機制更新圖中每個節點的特徵；同時，我們也使用長短期記憶架構處理球員和球隊在近期比賽中的表現。然後，我們利用多頭注意力機制學習球員特徵向量間的互相影響力。接著，透過對比學習將擁有相同先發狀態的球員特徵拉近，並將擁有不同先發狀態的球員特徵推遠，以強化球員的特徵。最後，我們使用多層感知器推薦先發陣容。實驗結果顯示，我們的推薦模型優於比較方法，而且易於延伸至其他的應用。我們的模型可以幫助教練們建構最佳的先發陣容，亦可預測對手的先發陣容，協助教練們擬定與敵隊交手時的戰術與策略。

Determining the optimal starting lineup for an upcoming game is a crucial and complex task for a soccer team. In this study, we propose a novel Contrastive-and-Attention-based STarting lineup Recommendation model, called CASTRec, for selecting soccer starting lineups. The proposed model contains seven modules. We first extract the features regarding players’ historical status, inter-team and intra-team interactions among players, and overall team performance in the first four modules. Specifically, we represent inter-team player interactions, intra-team player interactions and team matchups by graphs and utilize the graph attention mechanism to update the node representations. Also, we incorporate the long short-term memory to summarize the performance of players and teams across recent games. Next, we apply the multi-head attention mechanism to learn the inter-relationships among the players. Subsequently, we employ the contrastive learning to enhance the representations learned by pulling closer the representations of the players with the same starting status and pushing farther apart those of the players with different starting status. Last, we use the multilayer perceptron (MLP) to recommend the starting lineup. The experimental results show that our model outperforms the compared models and can be easily extended to other applications such as outcome, goal, shot and pass predictions. Our model can help coaches build an optimal starting lineup for an upcoming game and predict the starting lineups of opposing teams, which in turn helps coaches implement effective tactical strategies.