籃球比賽的勝負預測及選手的攻防能力是籃球運動分析中的重要課題，Lee (2025) 將 Gu et al. (2021) 提出用於線上遊戲的Neural Attentional Cooperation-competition (NeuralAC) 模型應用於NBA比賽場景。但在預測 NBA 比賽結果時，除了要考量球員本身較穩定的能力，也需要納入球員近期的表現狀態。Lee (2025) 在原始 NAC 架構中加入貝氏建模、指數移動平均特徵、以及深度學習的非線性映射機制，提出 Bayes-EMA NeuralAC 模型。雖然這個模型提升了預測準確率，但 Lee (2025) 並未明確討論表現的提升主要來自以上哪些因子; 此外，對於各特徵之全域數值差異也無納入考量。因此，本研究先對動態特徵進行標準化，以降低不同特徵之間數值尺度差異所造成的影響，接著再檢驗EMA 特徵、非線性映射機制、標準化與貝氏建模對模型表現的影響。本研究亦藉由將已移除的動態特徵重新加回，進行特徵敏感度分析，並採用隨機梯度下降法進行參數估計。根據單一賽季 NBA 資料集的實驗結果，EMA 特徵與非線性映射機制是提升預測表現的主要來源。相較之下，標準化僅帶來小幅度的提升而貝氏設定在整合模型中反而使預測表現微幅下降；透過將已移除變數重新加回的對比驗證，進一步證實被移除者均為冗餘特徵，使模型能維持穩定的預測性能。最後，本研究透過球星案例分析展示模型的可解釋性。

Predicting NBA game outcomes and evaluating players' offensive and defensive abilities are important issues in basketball analytics. Lee (2025) applied the Neural Attentional Cooperation-competition (NeuralAC) model, originally proposed by Gu et al. (2021) for online gaming, to NBA game prediction. NBA game outcomes depend not only on players' mean abilities but also on their recent performance. Thus, Lee (2025) extended the original NeuralAC framework by incorporating Bayesian modeling, Exponential Moving Average (EMA) features, and nonlinear mapping mechanism, called the Bayes-EMA NeuralAC model. Although this model improved prediction accuracy, Lee (2025) did not discuss which factor made more contribution to the improvement, moreover, the differences in the range among features have not been taken care of. Therefore, the present study first standardizes the features to reduce the influence of scale differences and then investigates the effects of EMA features, the nonlinear mapping mechanism, standardization, and Bayesian modeling on model performance. In addition, feature sensitivity analysis is conducted by adding back the removed features, and model parameters are estimated using Stochastic Gradient Descent (SGD). Based on experiments using a single-season NBA dataset, the results show that EMA features and the nonlinear mapping mechanism are the main factors improving prediction performance, while standardization yields only a marginal improvement, and incorporating the Bayesian setting into the integrated model slightly degrades predictive performance. The feature sensitivity analysis indicates that after standardization, features can be correctly removed while maintaining stable model performance. Finally, elite player case analysis is used to demonstrate the interpretability of the model.

Chen, S. and Joachims, T. (2016). Modeling intransitivity in matchup and comparison data. In Proceedings of the Ninth ACM International Conference on Web Search and Data Mining, pages 227–236.

Gu, Y., Liu, Q., Zhang, K., Huang, Z., Wu, R., and Tao, J. (2021). Neuralac: Learning cooperation and competition effects for match outcome prediction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 4072–4080.

Herbrich, R., Minka, T., and Graepel, T. (2006). Trueskill™: a bayesian skill rating system. Advances in neural information processing systems, 19.

Huang, T.-K., Lin, C.-J., and Weng, R. (2004). A generalized bradley-terry model: From group competition to individual skill. Advances in neural information processing systems, 17.

Huang, T.-K., Lin, C.-J., and Weng, R. C. (2006). Ranking individuals by group comparisons. In Proceedings of the 23rd international conference on Machine learning, pages 425–432.

Koren, Y., Bell, R., and Volinsky, C. (2009). Matrix factorization techniques for recommender systems. Computer, 42(8):30–37.

Lee, Y.-C. (2025). Sports analytics with bayesian skill updates and deep neural interaction models (in chinese). Master’s thesis, National Chengchi University, Taipei, Taiwan.