{ "id": 266596, "title": "Revisiting the nature of the targets [2nd place team]", "url": "/competitions/mlb-player-digital-engagement-forecasting/discussion/266596", "author_name": "JohnM", "post_date": "2021-08-19T17:00:59.991000", "votes": 14, "comment_count": 11, "views": 0, "content": "

Note: This is not our official team writeup, just some observations on the targets.

\n

There was some great discussion early on about the nature of the targets. I think most of us would like to know what they really are. Here are some thoughts I've had along the way.

\n

My hypothesis was that the targets represent 4 of the typical measures of digital engagement:

\n\n

These measures could be across any number of social media channels. I found Twitter to be the easiest for pulling data and ended up focusing there.

\n

For volume and sentiment, I looked at images of each target. Target2 has more high values than the others (Target4 also somewhat). You probably know this because MAE is largest there, and most models have trouble predicting the high values. Rows are the players and columns are the dates in these images. Values are the target scores.
\n\"tgt\"

\n

So maybe Targets 2 or 4 represent sentiment? Regardless of number of mentions, posts can be positive, neutral, or negative. An example is in April where 4 players scored 90+ on Target3. All were mentioned a lot on Twitter that day. Two of the players (DeGrom and Ohtani) had great pitching and got high scores on 2 and 4. The other two (Alvarado and Contreras) had very negative twitter traffic and scored very low.

\n

For likes and shares, my initial idea was to get counts of second-degree followers for each player as a proxy. I knew there were some 60M followers total and was surprised to find that there were over 20M unique followers. It took 3-4 days just to get the names of the followers (after a few botched attempts) and would have taken over a month to get follower counts of the 20M!

\n

As a fallback I looked at the 20M and how many players they followed, thinking maybe the targets were sensitive somehow to the distribution of uniqueness. Mike Trout and Yu Darvish are an example. Both have 2M+ followers, the most of any player by far. But Trout's followers tend to follow a lot of players, vs. Darvish whose followers tend to follow just him. Trout's mean target levels are 3-4 times Darvish's for targets 1,2 and 4 and 50% higher on 3.

\n

The binned features seemed promising. They showed high importance in the model and improved the score for basic models. But for more advanced models with lag features and such, the gains were not significant.

", "messages": [ { "id": 1481784, "postDate": "2021-08-19T17:00:59.990Z", "content": "

Note: This is not our official team writeup, just some observations on the targets.

\n

There was some great discussion early on about the nature of the targets. I think most of us would like to know what they really are. Here are some thoughts I've had along the way.

\n

My hypothesis was that the targets represent 4 of the typical measures of digital engagement:

\n\n

These measures could be across any number of social media channels. I found Twitter to be the easiest for pulling data and ended up focusing there.

\n

For volume and sentiment, I looked at images of each target. Target2 has more high values than the others (Target4 also somewhat). You probably know this because MAE is largest there, and most models have trouble predicting the high values. Rows are the players and columns are the dates in these images. Values are the target scores.
\n\"tgt\"

\n

So maybe Targets 2 or 4 represent sentiment? Regardless of number of mentions, posts can be positive, neutral, or negative. An example is in April where 4 players scored 90+ on Target3. All were mentioned a lot on Twitter that day. Two of the players (DeGrom and Ohtani) had great pitching and got high scores on 2 and 4. The other two (Alvarado and Contreras) had very negative twitter traffic and scored very low.

\n

For likes and shares, my initial idea was to get counts of second-degree followers for each player as a proxy. I knew there were some 60M followers total and was surprised to find that there were over 20M unique followers. It took 3-4 days just to get the names of the followers (after a few botched attempts) and would have taken over a month to get follower counts of the 20M!

\n

As a fallback I looked at the 20M and how many players they followed, thinking maybe the targets were sensitive somehow to the distribution of uniqueness. Mike Trout and Yu Darvish are an example. Both have 2M+ followers, the most of any player by far. But Trout's followers tend to follow a lot of players, vs. Darvish whose followers tend to follow just him. Trout's mean target levels are 3-4 times Darvish's for targets 1,2 and 4 and 50% higher on 3.

\n

The binned features seemed promising. They showed high importance in the model and improved the score for basic models. But for more advanced models with lag features and such, the gains were not significant.

", "rawMarkdown": "Note: This is not our official team writeup, just some observations on the targets.\n\nThere was some great discussion early on about the nature of the targets. I think most of us would like to know what they really are. Here are some thoughts I've had along the way.\n\nMy hypothesis was that the targets represent 4 of the typical measures of digital engagement:\n- volume/mentions\n- sentiment\n- likes and shares\n- responsiveness of player\n- search traffic\n- other?\n\nThese measures could be across any number of social media channels. I found Twitter to be the easiest for pulling data and ended up focusing there.\n\nFor volume and sentiment, I looked at images of each target. Target2 has more high values than the others (Target4 also somewhat). You probably know this because MAE is largest there, and most models have trouble predicting the high values. Rows are the players and columns are the dates in these images. Values are the target scores.\n![tgt](https://i.imgur.com/yRv9g9B.png)\n\nSo maybe Targets 2 or 4 represent sentiment? Regardless of number of mentions, posts can be positive, neutral, or negative. An example is in April where 4 players scored 90+ on Target3. All were mentioned a lot on Twitter that day. Two of the players (DeGrom and Ohtani) had great pitching and got high scores on 2 and 4. The other two (Alvarado and Contreras) had very negative twitter traffic and scored very low.\n\nFor likes and shares, my initial idea was to get counts of second-degree followers for each player as a proxy. I knew there were some 60M followers total and was surprised to find that there were over 20M unique followers. It took 3-4 days just to get the names of the followers (after a few botched attempts) and would have taken over a month to get follower counts of the 20M! \n\nAs a fallback I looked at the 20M and how many players they followed, thinking maybe the targets were sensitive somehow to the distribution of uniqueness. Mike Trout and Yu Darvish are an example. Both have 2M+ followers, the most of any player by far. But Trout's followers tend to follow a lot of players, vs. Darvish whose followers tend to follow just him. Trout's mean target levels are 3-4 times Darvish's for targets 1,2 and 4 and 50% higher on 3. \n\nThe binned features seemed promising. They showed high importance in the model and improved the score for basic models. But for more advanced models with lag features and such, the gains were not significant.\n", "votes": 14 }, { "id": 1487362, "postDate": "2021-08-23T15:19:30.633Z", "content": "

Great Insights!

\n

I guess second degree followers don't contribute much because most of us might be already capturing that information by including player historical targets, same reason why twitter followers don't contribute to model.

", "rawMarkdown": "Great Insights!\n\nI guess second degree followers don't contribute much because most of us might be already capturing that information by including player historical targets, same reason why twitter followers don't contribute to model. ", "votes": 2, "replies": [ { "id": 1513173, "postDate": "2021-09-14T23:02:14.980Z", "content": "

Very likely. BTW, congrats on the great finish!

", "rawMarkdown": "Very likely. BTW, congrats on the great finish!", "votes": 1 }, { "id": 1515062, "postDate": "2021-09-16T17:30:39.507Z", "content": "

Thanks! Congratulations to your team as well.

", "rawMarkdown": "Thanks! Congratulations to your team as well." } ] }, { "id": 1483491, "postDate": "2021-08-20T17:10:37.560Z", "content": "

As I sorta mentioned in my solution writeup, I think it would really only make sense for these features to relate to social media engagement. I've noticed that the MLB went from one of the most locked-down content presences of all of the US's major sports to being one of the more prolific within the past few years, so to me it seems like they are trying to figure out how to best capitalize on their social media posts given what happens in game.

\n

Another possibility is that they're considering rule changes, and want to have good predictive models so that they can toy around with factors like pace of play changes and see how that would theoretically affect the levels of engagement that they get. They may have been inspired by the way that the NFL is using competitions on Kaggle to inspire rule changes.

", "rawMarkdown": "As I sorta mentioned in my solution writeup, I think it would really only make sense for these features to relate to social media engagement. I've noticed that the MLB went from one of the most locked-down content presences of all of the US's major sports to being one of the more prolific within the past few years, so to me it seems like they are trying to figure out how to best capitalize on their social media posts given what happens in game. \n\nAnother possibility is that they're considering rule changes, and want to have good predictive models so that they can toy around with factors like pace of play changes and see how that would theoretically affect the levels of engagement that they get. They may have been inspired by the way that the NFL is using competitions on Kaggle to inspire rule changes.", "votes": 2, "replies": [ { "id": 1485529, "postDate": "2021-08-22T07:28:35.963Z", "content": "

I also think that these targets can be linked to their website as the data is easier to collect than social media
\nSo I think some targets might be :

\n", "rawMarkdown": "I also think that these targets can be linked to their website as the data is easier to collect than social media \nSo I think some targets might be :\n- Number of visitors to a player's page\n- Number of tickets sold after visiting a player's page\netc...", "votes": 2 }, { "id": 1489352, "postDate": "2021-08-24T22:56:05.630Z", "content": "

@amedprof I think you're on track after doing some web research. Google and MLB have set up several ways to engage fans on their cloud platform. This site mentions several things. The Statcast insight explorer, Film Room, integration with Google Ads, data distribution to team sites, etc.

\n

These elements alone seem a bit narrow for defining digital engagement. But they are easier for Google to track and may have a better connection to $$. So maybe this and some twitter stuff.

", "rawMarkdown": "@amedprof I think you're on track after doing some web research. Google and MLB have set up several ways to engage fans on their cloud platform. [This site](https://cloud.google.com/customers/mlb) mentions several things. The Statcast insight explorer, Film Room, integration with Google Ads, data distribution to team sites, etc.\n\nThese elements alone seem a bit narrow for defining digital engagement. But they are easier for Google to track and may have a better connection to $$. So maybe this and some twitter stuff.", "votes": 1 }, { "id": 1489358, "postDate": "2021-08-24T23:09:39.090Z", "content": "
\n

it seems like they are trying to figure out how to best capitalize on their social media posts

\n
\n

@marktenenholtz this makes sense. They have a Players list with around 500 players on it and they tweet pretty often.

", "rawMarkdown": "> it seems like they are trying to figure out how to best capitalize on their social media posts\n\n@marktenenholtz this makes sense. They have a [Players](https://twitter.com/i/lists/6199354) list with around 500 players on it and they tweet pretty often.", "votes": 1 }, { "id": 1490794, "postDate": "2021-08-25T21:31:39.530Z", "content": "

Oh, and there's this: https://www.mlb.com/careers/opportunities?gh_jid=3258647 :)

", "rawMarkdown": "Oh, and there's this: https://www.mlb.com/careers/opportunities?gh_jid=3258647 :)", "votes": 3 } ] }, { "id": 1518240, "postDate": "2021-09-20T14:09:13.530Z", "content": "

Cool! I like your work! I'd be happy if you can check my new Deep Learning tutorial, that covers LSTM model and many useful techniques😊

", "rawMarkdown": "Cool! I like your work! I'd be happy if you can check my new Deep Learning tutorial, that covers LSTM model and many useful techniques😊\n", "votes": -2 }, { "id": 1513396, "postDate": "2021-09-15T06:29:41.477Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 1513903, "postDate": "2021-09-15T14:15:02.833Z", "content": "

Thanks! I couldn't find the extra time for Explainability with all the time spent on the machine learning part. Also, I have this strange rule about only explaining things when the customer tells me what it is I'm explaining to them 😏

", "rawMarkdown": "Thanks! I couldn't find the extra time for Explainability with all the time spent on the machine learning part. Also, I have this strange rule about only explaining things when the customer tells me what it is I'm explaining to them 😏" } ] } ], "comments": [ { "id": 1487362, "author_name": "Mohsin hasan", "author_url": "", "post_date": "2021-08-23T15:19:30.633000", "content": "

Great Insights!

\n

I guess second degree followers don't contribute much because most of us might be already capturing that information by including player historical targets, same reason why twitter followers don't contribute to model.

", "votes": 2, "replies": [ { "id": 1513173, "author_name": "JohnM", "author_url": "", "post_date": "2021-09-14T23:02:14.980000", "content": "

Very likely. BTW, congrats on the great finish!

", "votes": 1, "replies": [] }, { "id": 1515062, "author_name": "Mohsin hasan", "author_url": "", "post_date": "2021-09-16T17:30:39.507000", "content": "

Thanks! Congratulations to your team as well.

", "votes": 0, "replies": [] } ] }, { "id": 1483491, "author_name": "Mark Tenenholtz", "author_url": "", "post_date": "2021-08-20T17:10:37.560000", "content": "

As I sorta mentioned in my solution writeup, I think it would really only make sense for these features to relate to social media engagement. I've noticed that the MLB went from one of the most locked-down content presences of all of the US's major sports to being one of the more prolific within the past few years, so to me it seems like they are trying to figure out how to best capitalize on their social media posts given what happens in game.

\n

Another possibility is that they're considering rule changes, and want to have good predictive models so that they can toy around with factors like pace of play changes and see how that would theoretically affect the levels of engagement that they get. They may have been inspired by the way that the NFL is using competitions on Kaggle to inspire rule changes.

", "votes": 2, "replies": [ { "id": 1485529, "author_name": "Amed", "author_url": "", "post_date": "2021-08-22T07:28:35.963000", "content": "

I also think that these targets can be linked to their website as the data is easier to collect than social media
\nSo I think some targets might be :

\n", "votes": 2, "replies": [] }, { "id": 1489352, "author_name": "JohnM", "author_url": "", "post_date": "2021-08-24T22:56:05.630000", "content": "

@amedprof I think you're on track after doing some web research. Google and MLB have set up several ways to engage fans on their cloud platform. This site mentions several things. The Statcast insight explorer, Film Room, integration with Google Ads, data distribution to team sites, etc.

\n

These elements alone seem a bit narrow for defining digital engagement. But they are easier for Google to track and may have a better connection to $$. So maybe this and some twitter stuff.

", "votes": 1, "replies": [] }, { "id": 1489358, "author_name": "JohnM", "author_url": "", "post_date": "2021-08-24T23:09:39.090000", "content": "
\n

it seems like they are trying to figure out how to best capitalize on their social media posts

\n
\n

@marktenenholtz this makes sense. They have a Players list with around 500 players on it and they tweet pretty often.

", "votes": 1, "replies": [] }, { "id": 1490794, "author_name": "Mark Tenenholtz", "author_url": "", "post_date": "2021-08-25T21:31:39.530000", "content": "

Oh, and there's this: https://www.mlb.com/careers/opportunities?gh_jid=3258647 :)

", "votes": 3, "replies": [] } ] }, { "id": 1518240, "author_name": "Victoria Maslova", "author_url": "", "post_date": "2021-09-20T14:09:13.530000", "content": "

Cool! I like your work! I'd be happy if you can check my new Deep Learning tutorial, that covers LSTM model and many useful techniques😊

", "votes": -2, "replies": [] }, { "id": 1513396, "author_name": "", "author_url": "", "post_date": "2021-09-15T06:29:41.477000", "content": "", "votes": 0, "replies": [ { "id": 1513903, "author_name": "JohnM", "author_url": "", "post_date": "2021-09-15T14:15:02.833000", "content": "

Thanks! I couldn't find the extra time for Explainability with all the time spent on the machine learning part. Also, I have this strange rule about only explaining things when the customer tells me what it is I'm explaining to them 😏

", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1481784": "Note: This is not our official team writeup, just some observations on the targets.\n\nThere was some great discussion early on about the nature of the targets. I think most of us would like to know what they really are. Here are some thoughts I've had along the way.\n\nMy hypothesis was that the targets represent 4 of the typical measures of digital engagement:\n- volume/mentions\n- sentiment\n- likes and shares\n- responsiveness of player\n- search traffic\n- other?\n\nThese measures could be across any number of social media channels. I found Twitter to be the easiest for pulling data and ended up focusing there.\n\nFor volume and sentiment, I looked at images of each target. Target2 has more high values than the others (Target4 also somewhat). You probably know this because MAE is largest there, and most models have trouble predicting the high values. Rows are the players and columns are the dates in these images. Values are the target scores.\n![tgt](https://i.imgur.com/yRv9g9B.png)\n\nSo maybe Targets 2 or 4 represent sentiment? Regardless of number of mentions, posts can be positive, neutral, or negative. An example is in April where 4 players scored 90+ on Target3. All were mentioned a lot on Twitter that day. Two of the players (DeGrom and Ohtani) had great pitching and got high scores on 2 and 4. The other two (Alvarado and Contreras) had very negative twitter traffic and scored very low.\n\nFor likes and shares, my initial idea was to get counts of second-degree followers for each player as a proxy. I knew there were some 60M followers total and was surprised to find that there were over 20M unique followers. It took 3-4 days just to get the names of the followers (after a few botched attempts) and would have taken over a month to get follower counts of the 20M! \n\nAs a fallback I looked at the 20M and how many players they followed, thinking maybe the targets were sensitive somehow to the distribution of uniqueness. Mike Trout and Yu Darvish are an example. Both have 2M+ followers, the most of any player by far. But Trout's followers tend to follow a lot of players, vs. Darvish whose followers tend to follow just him. Trout's mean target levels are 3-4 times Darvish's for targets 1,2 and 4 and 50% higher on 3. \n\nThe binned features seemed promising. They showed high importance in the model and improved the score for basic models. But for more advanced models with lag features and such, the gains were not significant.\n", "1487362": "Great Insights!\n\nI guess second degree followers don't contribute much because most of us might be already capturing that information by including player historical targets, same reason why twitter followers don't contribute to model. ", "1483491": "As I sorta mentioned in my solution writeup, I think it would really only make sense for these features to relate to social media engagement. I've noticed that the MLB went from one of the most locked-down content presences of all of the US's major sports to being one of the more prolific within the past few years, so to me it seems like they are trying to figure out how to best capitalize on their social media posts given what happens in game. \n\nAnother possibility is that they're considering rule changes, and want to have good predictive models so that they can toy around with factors like pace of play changes and see how that would theoretically affect the levels of engagement that they get. They may have been inspired by the way that the NFL is using competitions on Kaggle to inspire rule changes.", "1518240": "Cool! I like your work! I'd be happy if you can check my new Deep Learning tutorial, that covers LSTM model and many useful techniques😊\n", "1513396": "" } }