{ "id": 358590, "title": "📔 Things learned on the first week - Second week plan for beginners 💡", "url": "/competitions/tabular-playground-series-oct-2022/discussion/358590", "author_name": "Pastor Soto", "post_date": "2022-10-08T15:56:53.673000", "votes": 11, "comment_count": 0, "views": 0, "content": "

The first week was an exploratory week to understand the competition, the data, the initial suggestions and the game itself.

\n

My previous week plan was:

\n
    \n
  1. Understand online learning
    2. \n
  2. Check the discussion forums for ideas
    3. \n
  3. Understand the game
    4. \n
  4. Check NFL notebooks
    5. \n
\n

Let's go over one by one with the findings:

\n

Understand online learning

\n

Online learning is a method that allow to make predictions in live environment, it's a useful tool with a lot of applications such as stock market, sports events, weather and many others. It consumes less memory than batch prediction and tools like river mimics the sklearn syntax which make it really easy to implement.

\n

The cons, is that the resources are limited, and I couldn't find an example in a kaggle competition, also the performance is lower than batch prediction according to the river documentation. How to handle missing values can also be a problem if you perform some type of calculation on historical data.

\n

I am also not sure how does it work internally, maybe is a loop over every row in the model, if so it sounds computational expensive, and maybe a better approach will be using a sample of the training data with a batch prediction. I saw some results here that might support that path.

\n

Check discussion forum for ideas

\n

Key ideas I found were:

\n\n
\n

Missing values needs to be explored a little further, which could start by checking similar approaches for other competitions, what happened when a player leaves the field? My guess is that we care to place in a position when the other team increase their chances to score and the team of the player decrease their chances to score. We might want to find this position as a naive imputation.

\n

I need to explore more visualization, using things such as the one use in other competitions and in football.

\n
\n

Understand the game

\n

There are several discussions that allow to understand the game, and watching some minutes can also give you a better intuition on what's going on.

\n

The key things that I found were:

\n\n

Things I still don't know:

\n\n

Check NFL notebook

\n\n

Things I still don't know:
\nHow do they handle missing values?

\n

Second week plan

\n

Find a way to handle missing values

\n\n

Create a visualization notebook using football approaches

\n\n

Prepare the third week plan

\n", "messages": [ { "id": 1978221, "postDate": "2022-10-08T15:56:53.673Z", "content": "

The first week was an exploratory week to understand the competition, the data, the initial suggestions and the game itself.

\n

My previous week plan was:

\n
    \n
  1. Understand online learning
    2. \n
  2. Check the discussion forums for ideas
    3. \n
  3. Understand the game
    4. \n
  4. Check NFL notebooks
    5. \n
\n

Let's go over one by one with the findings:

\n

Understand online learning

\n

Online learning is a method that allow to make predictions in live environment, it's a useful tool with a lot of applications such as stock market, sports events, weather and many others. It consumes less memory than batch prediction and tools like river mimics the sklearn syntax which make it really easy to implement.

\n

The cons, is that the resources are limited, and I couldn't find an example in a kaggle competition, also the performance is lower than batch prediction according to the river documentation. How to handle missing values can also be a problem if you perform some type of calculation on historical data.

\n

I am also not sure how does it work internally, maybe is a loop over every row in the model, if so it sounds computational expensive, and maybe a better approach will be using a sample of the training data with a batch prediction. I saw some results here that might support that path.

\n

Check discussion forum for ideas

\n

Key ideas I found were:

\n\n
\n

Missing values needs to be explored a little further, which could start by checking similar approaches for other competitions, what happened when a player leaves the field? My guess is that we care to place in a position when the other team increase their chances to score and the team of the player decrease their chances to score. We might want to find this position as a naive imputation.

\n

I need to explore more visualization, using things such as the one use in other competitions and in football.

\n
\n

Understand the game

\n

There are several discussions that allow to understand the game, and watching some minutes can also give you a better intuition on what's going on.

\n

The key things that I found were:

\n\n

Things I still don't know:

\n\n

Check NFL notebook

\n\n

Things I still don't know:
\nHow do they handle missing values?

\n

Second week plan

\n

Find a way to handle missing values

\n\n

Create a visualization notebook using football approaches

\n\n

Prepare the third week plan

\n", "rawMarkdown": "The first week was an exploratory week to understand the competition, the data, the initial suggestions and the game itself.\n\nMy previous week plan was:\n\n1. Understand online learning\n2. Check the discussion forums for ideas\n3. Understand the game\n4. Check NFL notebooks\n\nLet's go over one by one with the findings:\n\n### Understand online learning\n\nOnline learning is a method that allow to make predictions in live environment, it's a useful tool with a lot of applications such as stock market, sports events, weather and many others. It consumes less memory than batch prediction and tools like [river](https://riverml.xyz/0.13.0/examples/batch-to-online/) mimics the sklearn syntax which make it really easy to implement. \n\nThe cons, is that the resources are limited, and I couldn't find an example in a kaggle competition, also the performance is lower than batch prediction according to the river documentation. How to handle missing values can also be a problem if you perform some type of calculation on historical data.\n\nI am also not sure how does it work internally, maybe is a loop over every row in the model, if so it sounds computational expensive, and maybe a better approach will be using a sample of the training data with a batch prediction. I saw some results here that might support that path. \n\n### Check discussion forum for ideas\n\nKey ideas I found were:\n\n- [Reduce the size of the dataset](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/356540)\n- [Handle missing values](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/357158)\n- [Online learning resources](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/356584)\n- Some visualizations\n\n> Missing values needs to be explored a little further, which could start by checking similar approaches for other competitions, what happened when a player leaves the field? My guess is that we care to place in a position when the other team increase their chances to score and the team of the player decrease their chances to score. We might want to find this position as a naive imputation. \n\n> I need to explore more visualization, using things such as the one use in other competitions and in football. \n\n### Understand the game\n\nThere are several discussions that allow to understand the game, and watching some minutes can also give you a better intuition on what's going on. \n\nThe key things that I found were:\n\n- The game is 3 vs 3\n- The game is similar to football, where the objective is to score a ball in the opposite field\n- Players can be destroyed by other player for 3 seconds and reaper in their own field\n\nThings I still don't know:\n- Why each game has more than one event?\n- What's the size of the field?\n\n### Check NFL notebook\n\n- The winning solution was a neural network\n- They simplified the game by using a runner and defenders and remove teammates\n- The size of the dataset was significantly lower that this competition\n- They had to optimize their memory to came up with the winning solution\n\nThings I still don't know:\nHow do they handle missing values?\n\n### Second week plan\n\n**Find a way to handle missing values**\n\n- Check the NFL notebook to see how did they handle those values\n- Check discussion forum and notebooks to see how people are handling missing values\n- Understand the field and the position for, remove players that make more sense (lower probability of their team and increase the probability of the other team)\n\n**Create a visualization notebook using football approaches**\n\n- Check this [YouTube channel](https://www.youtube.com/channel/UCUBFJYcag8j2rm_9HkrrA7w) that has tracking approaches for football to draw a field and applied speed and position of the players\n- Check the discussion forum and notebooks to see EDA ideas\n- Create an EDA notebook that allow to understand the features and came up with a modeling plan\n- From the EDA perform some feature engineering on the dataset\n\n**Prepare the third week plan**\n\n- Create baseline model", "votes": 11 } ], "comments": [], "raw_markdown_by_id": { "1978221": "The first week was an exploratory week to understand the competition, the data, the initial suggestions and the game itself.\n\nMy previous week plan was:\n\n1. Understand online learning\n2. Check the discussion forums for ideas\n3. Understand the game\n4. Check NFL notebooks\n\nLet's go over one by one with the findings:\n\n### Understand online learning\n\nOnline learning is a method that allow to make predictions in live environment, it's a useful tool with a lot of applications such as stock market, sports events, weather and many others. It consumes less memory than batch prediction and tools like [river](https://riverml.xyz/0.13.0/examples/batch-to-online/) mimics the sklearn syntax which make it really easy to implement. \n\nThe cons, is that the resources are limited, and I couldn't find an example in a kaggle competition, also the performance is lower than batch prediction according to the river documentation. How to handle missing values can also be a problem if you perform some type of calculation on historical data.\n\nI am also not sure how does it work internally, maybe is a loop over every row in the model, if so it sounds computational expensive, and maybe a better approach will be using a sample of the training data with a batch prediction. I saw some results here that might support that path. \n\n### Check discussion forum for ideas\n\nKey ideas I found were:\n\n- [Reduce the size of the dataset](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/356540)\n- [Handle missing values](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/357158)\n- [Online learning resources](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/356584)\n- Some visualizations\n\n> Missing values needs to be explored a little further, which could start by checking similar approaches for other competitions, what happened when a player leaves the field? My guess is that we care to place in a position when the other team increase their chances to score and the team of the player decrease their chances to score. We might want to find this position as a naive imputation. \n\n> I need to explore more visualization, using things such as the one use in other competitions and in football. \n\n### Understand the game\n\nThere are several discussions that allow to understand the game, and watching some minutes can also give you a better intuition on what's going on. \n\nThe key things that I found were:\n\n- The game is 3 vs 3\n- The game is similar to football, where the objective is to score a ball in the opposite field\n- Players can be destroyed by other player for 3 seconds and reaper in their own field\n\nThings I still don't know:\n- Why each game has more than one event?\n- What's the size of the field?\n\n### Check NFL notebook\n\n- The winning solution was a neural network\n- They simplified the game by using a runner and defenders and remove teammates\n- The size of the dataset was significantly lower that this competition\n- They had to optimize their memory to came up with the winning solution\n\nThings I still don't know:\nHow do they handle missing values?\n\n### Second week plan\n\n**Find a way to handle missing values**\n\n- Check the NFL notebook to see how did they handle those values\n- Check discussion forum and notebooks to see how people are handling missing values\n- Understand the field and the position for, remove players that make more sense (lower probability of their team and increase the probability of the other team)\n\n**Create a visualization notebook using football approaches**\n\n- Check this [YouTube channel](https://www.youtube.com/channel/UCUBFJYcag8j2rm_9HkrrA7w) that has tracking approaches for football to draw a field and applied speed and position of the players\n- Check the discussion forum and notebooks to see EDA ideas\n- Create an EDA notebook that allow to understand the features and came up with a modeling plan\n- From the EDA perform some feature engineering on the dataset\n\n**Prepare the third week plan**\n\n- Create baseline model" } }