Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Predicting Soccer Highlights from Spatio-Temporal Match Event Streams
Authors: Tom Decroos, Vladimir Dzyuba, Jan Van Haaren, Jesse Davis
AAAI 2017 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | An empirical evaluation on a real-world dataset shows that POGBA outperforms the baseline algorithms in terms of both precision and recall. |
| Researcher Affiliation | Academia | Tom Decroos, Vladimir Dzyuba, Jan Van Haaren, Jesse Davis KU Leuven, Department of Computer Science, 3001 Leuven, Belgium EMAIL |
| Pseudocode | Yes | Algorithm 1 provides pseudocode for the full prediction pipeline. |
| Open Source Code | Yes | More extensive and detailed results are available in the online supplement.3 https://dtai.cs.kuleuven.be/sports/pogba |
| Open Datasets | No | The dataset consists of play-by-play data for 69 soccer matches from a Belgian professional soccer club. The paper does not provide explicit access information (link, citation, repository) for this dataset, nor does it state that it is a publicly available or open dataset. |
| Dataset Splits | No | For each match, we train the model on the 68 other matches and compute the precision and recall for the remaining match aggregated over both teams. This describes a leave-one-match-out cross-validation setup, but does not specify a distinct validation set or explicit train/validation/test split percentages/counts typically used for hyperparameter tuning. |
| Hardware Specification | No | The paper mentions 'Our unoptimized Python implementation classifies a window in approximately three seconds' but does not provide specific details about the hardware used for running experiments (e.g., CPU/GPU models, memory). |
| Software Dependencies | No | The paper mentions 'Our unoptimized Python implementation' but does not specify the version of Python or any other software dependencies with version numbers. |
| Experiment Setup | Yes | Based on our preliminary experiments and domain knowledge, we set the number of nearest neighbors k to 100, the exponential smoothing parameter α to 0.2, and the peak detection window w to 18. The choice of w is based on the analysis of the training data. It showed that, on average, a noteworthy event (e.g., a goal attempt) happens approximately every three minutes. Hence, we set the peak detection window to 18, which corresponds to 90 seconds, as this yields roughly one detected peak per three minutes. In the main experiment, we used ERT. |