\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6537187%2Fedb3f7ce80c300fc29e19a1420a717bb%2Ffe.PNG?generation=1664635982874140&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6537187%2Ff35a97c0d04a45c97735f3e57323c2cd%2Fdist.PNG?generation=1664636007250965&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6537187%2Fc8249466bc04b2ca6292d53e219dce52%2Fpoly.PNG?generation=1664636020275844&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6537187%2Faa203a9b5bad7b86d07c2827942f1f69%2Fphy.PNG?generation=1664636040782988&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6537187%2Fbd3ee504f38b08aed9de01030fb34ab9%2Fdir.PNG?generation=1664636058489247&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6537187%2F633a27ddf8d90c333a4c7c50e94d8b7e%2Fgraph.PNG?generation=1664636070977481&alt=media\")
Thanks for these cool features @ravishah1
", "rawMarkdown": "Thanks for these cool features @ravishah1", "votes": null }, { "id": "1971625", "postDate": "10/04/2022 17:42:49", "content": "Thank you @ravishah1, amazing analysis!
", "rawMarkdown": "Thank you @ravishah1, amazing analysis!", "votes": null }, { "id": "1971973", "postDate": "10/04/2022 21:15:05", "content": "Some good advice, thank you.
", "rawMarkdown": "Some good advice, thank you.", "votes": null }, { "id": "1976936", "postDate": "10/07/2022 16:44:50", "content": "I have a question!!!
\nIt says \"test.csv: Test set. Unlike the train set, the rows are scrambled.\"
\nwhat does it mean, does is mean that all the rows in test set are shuffled so that each row can belongs to any game, any event randomly without the time sequence preserved?
\nthe 'event_time' is not given in the test set, and we are expected to predict team_[A|B]_scoring_within_10sec, how do we know at what point does the last 10sec begins without having 'event_time' in test set, because I assume the predictions for both teams should be 0 (or close to 0) before 10sec begin as it is in the train set
You're right.
\nThis is the challenging part, where we cannot use direct time-series prediction. We would treat each row as independent event, and guess the goals based on current state.
\nExample: if the ball location is very close to Goal A, then the probability of scoring will be higher
Thanks for the feedback…, but how about the last part, If time is not present then do we still need to predict whether the given row is belongs to (<10sec) or (>10sec) in addition to which team wins.
\nI still doubt because by looking at single row having just static information about positions and velocities of an arbitrary 'moment' of a game without any reference to what happened couple of moments ago, how to predict whether that moment belongs to (<10sec) and which team will win. am I missing something?🤔
How can we use acceleration and trajectory calculation for the test dataset if there is no timeline? Probably a silly question, but I can't figure out how to use this information correctly
", "rawMarkdown": "How can we use acceleration and trajectory calculation for the test dataset if there is no timeline? Probably a silly question, but I can't figure out how to use this information correctly", "votes": null }, { "id": "1977965", "postDate": "10/08/2022 11:26:08", "content": "We cant predict whether this row is <10 or >10 sec, as some games end in just 5 seconds 😄
\nSo, we need a ML model to learn from dataset patterns, and calculate the probabilities.
\nThere are several suggestion to predict next few frame based on current position and velocity, but this is hard as there are too many unknowns.
Thanks for these cool features @ravishah1
", "votes": null, "replies": [] }, { "id": 1971625, "author_name": "marianocaccavale", "author_url": "", "post_date": "10/04/2022 17:42:49", "content": "Thank you @ravishah1, amazing analysis!
", "votes": null, "replies": [] }, { "id": 1971973, "author_name": "stefancomanita", "author_url": "", "post_date": "10/04/2022 21:15:05", "content": "Some good advice, thank you.
", "votes": null, "replies": [] }, { "id": 1976936, "author_name": "charithamaduranga", "author_url": "", "post_date": "10/07/2022 16:44:50", "content": "I have a question!!!
\nIt says \"test.csv: Test set. Unlike the train set, the rows are scrambled.\"
\nwhat does it mean, does is mean that all the rows in test set are shuffled so that each row can belongs to any game, any event randomly without the time sequence preserved?
\nthe 'event_time' is not given in the test set, and we are expected to predict team_[A|B]_scoring_within_10sec, how do we know at what point does the last 10sec begins without having 'event_time' in test set, because I assume the predictions for both teams should be 0 (or close to 0) before 10sec begin as it is in the train set
You're right.
\nThis is the challenging part, where we cannot use direct time-series prediction. We would treat each row as independent event, and guess the goals based on current state.
\nExample: if the ball location is very close to Goal A, then the probability of scoring will be higher
Thanks for the feedback…, but how about the last part, If time is not present then do we still need to predict whether the given row is belongs to (<10sec) or (>10sec) in addition to which team wins.
\nI still doubt because by looking at single row having just static information about positions and velocities of an arbitrary 'moment' of a game without any reference to what happened couple of moments ago, how to predict whether that moment belongs to (<10sec) and which team will win. am I missing something?🤔
We cant predict whether this row is <10 or >10 sec, as some games end in just 5 seconds 😄
\nSo, we need a ML model to learn from dataset patterns, and calculate the probabilities.
\nThere are several suggestion to predict next few frame based on current position and velocity, but this is hard as there are too many unknowns.
How can we use acceleration and trajectory calculation for the test dataset if there is no timeline? Probably a silly question, but I can't figure out how to use this information correctly
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1965858": "# Feature Overview\n- we have a lot of really cool features to work with \n- there is nearly 10 GB of data\n- the data can be broken down into 3 main groups: position-based player data, position-based ball data, temporal data\n\n\n# Spatial Features\n\n### Distance Features\n- When working with spatial data, there are a lot of distance features that can be created\n- In face, since there are 7 objects (6 players + 1 ball) and you can find the distance between any 2 object, you can use the combination formula to determine there are (7!)/(2!(7-2)!) = 21 features that can be created using a single distance metric between the points\n- There are many ways to calculate distance, below is the Minkowski Formula which is a generalized formula that can calculate Manhattan distance, Euclidean distance, and additional Minkowski distances based on the parameter n\n\n\n### Polygon Features\n- Since we have objects on a coordinate plane, we can create a polygon object and extract features from that\n- The tricky part is making sure you actually create an accurate polygon since you only have points and not edges \n- If you can get a polygon to work, you can extract a lot of shape features such as area, centroid, etc\n\n\n# Temporal Features\n\n### Physics Features\n- With the introduction of time as a feature, you can get more complicated features by using physics\n- Some basic kinematic equations are below to help you get started\n- You can add velocity and acceleration as features or even try and use them to predict the future location of the cars\n\n\n### Direction Features\n- Temporal data also allows use to see where the car came from\n- Using this information, we can map a trajectory of the car\n\n\n# Graph Representation\n- A very complicated idea is to create a graph representation of all the objects\n- Each node could represent the (x, y, z) coords \n- Each edge could represent the distance between nodes\n- It may be challenging to decide how you want to set it up (how to create edges, create edges with all points or only based on threshold, use adjacency matrix or other data structure, etc)\n- Once you generate a graph, you could try feeding it into some sort of Graph Neural Network (GNN)\n- Note you probably want to start with the ideas above as this can be very complicated and tabular feature engineering is usually very powerful anyway\n", "1971550": "Thanks for these cool features @ravishah1", "1971625": "Thank you @ravishah1, amazing analysis!", "1971973": "Some good advice, thank you.", "1976936": "I have a question!!!\nIt says ***\"test.csv: Test set. Unlike the train set, the rows are scrambled.\"***\nwhat does it mean, does is mean that all the rows in test set are shuffled so that each row can belongs to any game, any event randomly without the time sequence preserved?\nthe 'event_time' is not given in the test set, and we are expected to predict team_[A|B]_scoring_within_10sec, how do we know at what point does the last 10sec begins without having 'event_time' in test set, because I assume the predictions for both teams should be 0 (or close to 0) before 10sec begin as it is in the train set", "1976959": "You're right.\nThis is the challenging part, where we cannot use direct time-series prediction. We would treat each row as independent event, and guess the goals based on current state.\nExample: if the ball location is very close to Goal A, then the probability of scoring will be higher", "1977012": "Thanks for the feedback..., but how about the last part, If time is not present then do we still need to predict whether the given row is belongs to (<10sec) or (>10sec) in addition to which team wins. \nI still doubt because by looking at single row having just static information about positions and velocities of an arbitrary 'moment' of a game without any reference to what happened couple of moments ago, how to predict whether that moment belongs to (<10sec) and which team will win. am I missing something?🤔", "1977951": "How can we use acceleration and trajectory calculation for the test dataset if there is no timeline? Probably a silly question, but I can't figure out how to use this information correctly", "1977965": "We cant predict whether this row is <10 or >10 sec, as some games end in just 5 seconds 😄\nSo, we need a ML model to learn from dataset patterns, and calculate the probabilities.\nThere are several suggestion to predict next few frame based on current position and velocity, but this is hard as there are too many unknowns." }, "source": "meta" }