Figure 6
The figure is divided into three panels labeled (a), (b), and (c) arranged side-by-side. The left panel (a), titled “Cross League Generalization”, is a line chart with “Relative Performance (percent)” on the vertical axis ranging from 40 to 100 with an interval of 10 percent. The horizontal axis is labeled “Target League” and lists four leagues: La Liga, Serie A, Bundesliga, and Ligue 1. A gray label near the top left reads “Trained on Premier League”. Four colored lines represent different methods: a purple circle line for Our H D R L, a yellow square marker line for Transformer, a blue triangular marker line for V A E P, and a green diamond marker line for G R F, as indicated in a legend near the lower left. Across all four target leagues, the Our H D R L line stays high, around 80 to 83 percent relative performance, with a highlighted point at La Liga labeled “81.2 percent”, while the transformer line stays around 51 to 55 percent of relative performance. The other two methods stay lower, ranging between 42 and 50 percent, and both peak at Bundesliga. The middle panel (b), titled “Cross Sport Transfer Learning”, is a line chart with “Number of Target Sport Matches” on the horizontal axis ranging from 0 to 1000 with an interval of 200. The vertical axis is labeled “Accuracy (percent)” and ranges from 45 to 95 with an interval of 5 percent. Two curves compare training regimes when transferring from soccer to basketball. A red line with a circular marker labeled “From Scratch” starts from 45 percent at 0 matches and increases gradually, ending around 86 percent at 1000 matches. A purple line with a circular marker line labeled “With Transfer” begins higher, around 73.5 percent at 0 matches, and rises more quickly, ending around 90 percent by 1000 matches. The area between the two curves is shaded light purple. Text near the 100 match point notes “73.4 percent” for the transfer curve and “5.8 percent faster convergence”, and a short vertical line connects both curves at 100 matches. The right panel (c), titled “Temporal Generalization”, is a grouped bar chart showing performance over three test seasons: 2021 to 22, 2022 to 23, and 2023 to 24 on the horizontal axis. The vertical axis is labeled “Performance (percent)” and ranges from 55 to 90 with an interval of 5 percent. For each season, a purple bar represents a “Static Model”, and a green bar represents an “Adaptive Model”, as indicated in a legend at the bottom left. In 2021 to 22, both bars are high and equal, with labels of 85.3 percent for both. For 2022 to 23, the Static Model bar drops to about 79.2, while the Adaptive Model reaches 83.1 percent. For 2023 to 24, the Static Model decreases further to about 69.4 percent, whereas the Adaptive Model maintains higher performance near 81.7 percent. A dashed grey line connecting the tops of the Static Model bars slopes downward from left to right, annotated by a callout labeled “Trained on 2020 to 21”. Note: All numerical data values are approximated.Generalization performance across different contexts