Pengembangan Model CTGAN-LSTM Berbasis Sequential Pattern Mining (SPADE) dalam Prediksi Pukulan Bulu Tangkis

Abstract

Bulu tangkis merupakan olahraga dengan rangkaian pukulan yang berurutan dalam setiap rally. Urutan pukulan tersebut membentuk pola sekuensial yang dapat digunakan untuk memahami strategi permainan dan memprediksi pukulan berikutnya. Penelitian sebelumnya telah memanfaatkan metode sequential pattern mining, seperti Sequential Pattern Discovery using Equivalence Classes (SPADE), untuk menemukan pola urutan pukulan yang sering muncul. Pendekatan tersebut masih bersifat deskriptif karena hanya melihat pola yang sering terjadi tanpa membangun model prediksi. Data pola pukulan bulu tangkis memiliki distribusi Tipe pukulan yang tidak seimbang, sehingga model prediktif dapat lebih dominan mengenali Tipe pukulan mayoritas dan kurang optimal dalam mengenali Tipe pukulan minoritas. Penelitian ini dimulai dengan mengekstraksi urutan pukulan bulu tangkis menjadi rules pattern menggunakan SPADE. Membangun Conditional Tabular Generative Adversarial Network (CTGAN) untuk menghasilkan data sintetis yang menyerupai pola asli sekaligus menyeimbangkan Tipe pukulan data. Mengembangkan model Long-Short Term Memory (LSTM) many-to-one untuk memprediksi pukulan berikutnya, serta bertujuan untuk membandingkan pengaruh pemodelan LSTM sebelum dan sesudah augmentasi CTGAN. Data penelitian diperoleh dari dua pertandingan bulu tangkis tunggal putra tingkat dunia, yaitu BWF Championships dan Toyota Thailand Open pada video Youtube resmi. Data pukulan dicatat berdasarkan urutan dalam rally, kemudian diproses menjadi data sekuensial. Evaluasi dilakukan pada dua skema klasifikasi, yaitu skema 5 Tipe pukulan dan 10 Tipe pukulan. Skema 5 Tipe pukulan membedakan target berdasarkan jenis pukulan, yaitu Drive, Dropshot, Lob, Netting, dan Smash, sedangkan skema 10 Tipe pukulan membedakan target berdasarkan kombinasi jenis pukulan dan kode pemain yaitu Anders Antonsen dan Kunlavut Vitidsarn. Kinerja model dievaluasi menggunakan accuracy, balanced accuracy, dan F1-score berdasarkan 10 kali pengulangan. Uji t dilakukan untuk mengetahui apakah terdapat perbedaan kinerja antara model LSTM sebelum dilakukan CTGAN dan setelah dilakukan augmentasi dengan CTGAN signifikan secara statistik. Hasil penelitian menunjukkan bahwa CTGAN mampu menyeimbangkan distribusi data latih dan menghasilkan data sintetis pada skema 5 Tipe pukulan maupun 10 Tipe pukulan dengan nilai kelayakan TabSynDex pada kisaran 0.70-0.76 berdasarkan nilai tertinggi dengan pelatihan dari beberapa epoch. Nilai tersebut menunjukkan bahwa data sintetis cukup layak digunakan sebagai data tambahan dalam proses pelatihan model. Namun salah satu komponen nilai TabSynDex yaitu ML Eficacy masih memiliki nilai yang rendah dimana yang berarti bahwa data sintetis belum sepenuhnya baik digunakan pada machine learning. Hasil pemodelan dengan menggunakan 5-fold cross validation pada data train dan dilakukan 10 kali perulangan dengan seed yang berbeda-beda menunjukkan bahwa CTGAN-LSTM secara umum memberikan kinerja lebih baik dibandingkan LSTM saja, terutama pada skema 5 Tipe pukulan meski belum sepenuhnya baik pada data gabungan 2 match. Performa terbaik diperoleh pada data Toyota Thailand Open skema 5 Tipe pukulan dengan nilai rata-rata accuracy sebesar 0.8962, rata-rata balanced accuracy sebesar 0.8180, dan rata-rata F1-score sebesar 0.8259. Pada skema 10 Tipe pukulan, peningkatan performa lebih terbatas karena target klasifikasi lebih kompleks, yaitu menggabungkan jenis pukulan dan kode pemain. Hasil uji-t menunjukkan bahwa peningkatan CTGAN-LSTM paling kuat terjadi pada Toyota Thailand Open skema 5 Tipe pukulan sehingga ketiga metrik evaluasi diartikan terdapat perbedaan rata-rata yang signifikan dari kedua model. Sedangkan pada BWF Championships hanya metrik evaluasi accuracy saja yang tidak signifikan atau tidak terdapat perbedaan rata-rata dari kedua model sedangkan untuk balanced accuracy dan f1-score signifikan. Pada data gabungan dua pertandingan, peningkatan performa belum signifikan secara statistik atau tidak terdapat cukup bukti adanya perbedaan rata-rata dari kedua model. Dikarenakan karakteristik yang semakin beragam dan dua distribusi yang berbeda dari kedua model digabungkan sehingga cukup kompleks dalam pemodelan. Pengembangan model CTGAN-LSTM berbasis sequential pattern mining (SPADE) dapat digunakan sebagai pendekatan prediksi pukulan berikutnya pada data bulu tangkis, terutama untuk membantu menangani ketidakseimbangan Tipe pukulan, meskipun efektivitasnya tetap dipengaruhi oleh kompleksitas Tipe pukulan, karakteristik data, dan kualitas data sintetis yang dihasilkan.

Badminton is a sport characterized by a series of consecutive shots in each rally. These shot sequences form sequential patterns that can be used to understand game strategies and predict the next shot. Previous research has utilized sequential pattern mining methods, such as Sequential Pattern Discovery using Equivalence Classes (SPADE), to identify frequently occurring stroke sequence patterns. These approaches remain descriptive because they only examine frequently occurring patterns without building predictive models. Badminton stroke pattern data has an imbalanced class distribution, so predictive models may be more dominant in recognizing the majority class and less optimal in recognizing the minority class. Therefore, this study develops an integrated approach combining SPADE, Conditional Tabular Generative Adversarial Network (CTGAN), and Long Short-Term Memory (LSTM) to predict the next shot based on sequential patterns. This study began by extracting badminton shot sequences into rule patterns using SPADE, building a CTGAN to generate synthetic data that resembles the original patterns while balancing the data classes, developing a many-to-one LSTM model to predict the next shot, and aiming to compare the impact of modeling LSTM before and after CTGAN augmentation. The research data was obtained from two world-class men’s singles badminton matches the BWF Championships and the Toyota Thailand Open with youtube video BWF. Shot data was recorded based on the sequence within the rally and then processed into sequential data. The evaluation was conducted on two classification schemes: a 5-class scheme and a 10-class scheme. The 5-class scheme distinguishes targets based on shot type namely, Drive, Dropshot, Lob, Netting, and Smash while the 10-class scheme distinguishes targets based on a combination of shot type and player code, specifically Anders Antonsen and Kunlavut Vitidsarn. Model performance was evaluated using accuracy, balanced accuracy, and F1-score based on ten repetitions. Additionally, t-test were performed to determine whether there was a statistically significant difference in performance between the LSTM model before CTGAN and after augmentation with CTGAN. The results of the study show that CTGAN is capable of balancing the distribution of training data and generating synthetic data in both 5-class and 10-class schemes, with TabSynDex validity scores ranging from 0.70 to 0.76 based on the highest values obtained after several epochs of training. These values indicate that the synthetic data is sufficiently valid for use as supplementary data in the model training process. However, one component of the TabSynDex score ML Eficacy still has a low value, which means that the synthetic data is not yet fully suitable for use in machine learning. Modeling results using 5-fold cross-validation on the training data, performed 10 times with different seeds, show that CTGAN-LSTM generally performs better than LSTM alone, especially in the 5-class scheme, although it is not yet fully effective on the combined 2-match data. The best performance was achieved on the Toyota Thailand Open dataset using the 5-class scheme, with an average accuracy of 0.8962, an average balanced accuracy of 0.8180, and an average F1-score of 0.8259. In the 10-class scheme, the improvement in performance was more limited because the classification target was more complex, combining shot type and player code. The results of the t-test show that the strongest improvement from CTGAN-LSTM occurred in the 5-class Toyota Thailand Open dataset, indicating that there were significant differences in the mean values of all three evaluation metrics between the two models. In contrast, for the BWF Championships dataset, only the accuracy metric showed no significant difference meaning there was no significant difference in the mean values between the two models while balanced accuracy and F1-score showed significant differences. In the combined data from the two tournaments, the performance improvement was not yet statistically significant or there was insufficient evidence of a mean difference between the two models due to increasingly diverse characteristics and the combination of two different distributions from the models, making the modeling process quite complex. The development of the CTGAN-LSTM model based on sequential pattern mining (SPADE) can be used as an approach to predict the next shot in badminton data, particularly to help address class imbalance, although its effectiveness remains influenced by class complexity, data characteristics, and the quality of the generated synthetic data.