Just thought I'd start this discussion for fun, given this is obviously not going to change for this competition. But does anyone else think the much more appropriate DV here would be a binary whether or not the customer spends?
\n\nIt seems to me that optimizing and learning how to predict if a person will make a purchase (1) or not (0) would be much more beneficial to google than predicting how much people spend, when over 95% of people who come in the store spend nothing.
", "messages": [ { "id": 389997, "postDate": "2018-09-19T15:16:29.293Z", "content": "I would say both might be useful. \nIf you spend a given cost per click to acquire a visitor, this cost might vary depending on your ad campaign, channel, keywords, etc. And you'd want to know how much earnings per click you are generating (ideally you would look at margin, not revenue). That way you can tell whether your campaign is profitable or not.
\n\nFor example, visitors coming from referrals buy 5% of the time whereas visitors coming from display ads spend 2% of the time. However, display ads visitors have average transactions that are twice as big. So just looking at conversion rate (i.e. whether we have a purchase or not) doesn't tell the whole story.
", "rawMarkdown": "I would say both might be useful. \nIf you spend a given cost per click to acquire a visitor, this cost might vary depending on your ad campaign, channel, keywords, etc. And you'd want to know how much earnings per click you are generating (ideally you would look at margin, not revenue). That way you can tell whether your campaign is profitable or not.\n\nFor example, visitors coming from referrals buy 5% of the time whereas visitors coming from display ads spend 2% of the time. However, display ads visitors have average transactions that are twice as big. So just looking at conversion rate (i.e. whether we have a purchase or not) doesn't tell the whole story.", "votes": 3, "replies": [ { "id": 390005, "postDate": "2018-09-19T15:22:07.037Z", "content": "Very true!
", "rawMarkdown": "Very true!" } ] }, { "id": 390143, "postDate": "2018-09-19T19:17:37.760Z", "content": "Defining a this competition as something like a CTR (Click Through Rate Prediction) problem was an option but I think they wanted to predict the value of each user interaction, by doing this they going one level deeper to find out how much a user is worth by predicting his expenditure, then we can categorize the users based on their predicted values and create something like CRF loyalty plan for users!
", "rawMarkdown": "Defining a this competition as something like a CTR (Click Through Rate Prediction) problem was an option but I think they wanted to predict the value of each user interaction, by doing this they going one level deeper to find out how much a user is worth by predicting his expenditure, then we can categorize the users based on their predicted values and create something like CRF loyalty plan for users!", "votes": 1 }, { "id": 390009, "postDate": "2018-09-19T15:23:55.697Z", "content": "Yes, I agree with you. Given the number of rows with transaction value, and the number of relevant variables with high cardinality, I would have made this competition to just predict whether the person will purchase or not.
", "rawMarkdown": "Yes, I agree with you. Given the number of rows with transaction value, and the number of relevant variables with high cardinality, I would have made this competition to just predict whether the person will purchase or not.", "votes": 1 }, { "id": 389999, "postDate": "2018-09-19T15:18:54.613Z", "content": "I'd say that business usually better understands what it needs :)
\n\nI think that both ideas (revenue prediction or binary classification) are valid, but the company decided that they need information about the revenue. Maybe it will help to target potentially profitable clients.
", "rawMarkdown": "I'd say that business usually better understands what it needs :)\n\nI think that both ideas (revenue prediction or binary classification) are valid, but the company decided that they need information about the revenue. Maybe it will help to target potentially profitable clients.", "votes": 1, "replies": [ { "id": 390004, "postDate": "2018-09-19T15:21:06.340Z", "content": "Oh, I totally agree! I was just having a little thought experiment. My intuition is that getting more people to spend money would have a greater outcome than just determining how we can get those who spend to spend more - but two equally valid sides of the same coin!
", "rawMarkdown": "Oh, I totally agree! I was just having a little thought experiment. My intuition is that getting more people to spend money would have a greater outcome than just determining how we can get those who spend to spend more - but two equally valid sides of the same coin!", "votes": 1 } ] }, { "id": 391322, "postDate": "2018-09-21T15:40:57.580Z", "content": "What can be useful for host is an interesting question. In this competition we know what already happened, like how many times a visitor clicked during any specific visit, number of visits for an year, etc. And I believe info about visit is more predictive, than geo location for example. I can imagine how geo info can be used for marketing, but not visitor behavior on GStore. So if we can predict whether visitor will buy something or even revenue amount, how to use it? If user just bought it, GStore knows it without ML, but it is not possible to know if visitor will have 22 hits next time when he/she will spend 23000$.
\n\nBtw, what does DV mean? Dependent variable?
", "rawMarkdown": "What can be useful for host is an interesting question. In this competition we know what already happened, like how many times a visitor clicked during any specific visit, number of visits for an year, etc. And I believe info about visit is more predictive, than geo location for example. I can imagine how geo info can be used for marketing, but not visitor behavior on GStore. So if we can predict whether visitor will buy something or even revenue amount, how to use it? If user just bought it, GStore knows it without ML, but it is not possible to know if visitor will have 22 hits next time when he/she will spend 23000$. \n\nBtw, what does DV mean? Dependent variable?", "votes": 2, "replies": [ { "id": 391328, "postDate": "2018-09-21T15:58:58.853Z", "content": "I just created a topic mentioning that particular issue :)
", "rawMarkdown": "I just created a topic mentioning that particular issue :)" }, { "id": 391377, "postDate": "2018-09-21T17:59:42.687Z", "content": "True! and Yes, dependent variable; that's my psychology training coming through haha
", "rawMarkdown": "True! and Yes, dependent variable; that's my psychology training coming through haha", "votes": 1 } ] }, { "id": 389932, "postDate": "2018-09-19T13:38:28.407Z", "content": "Just thought I'd start this discussion for fun, given this is obviously not going to change for this competition. But does anyone else think the much more appropriate DV here would be a binary whether or not the customer spends?
\n\nIt seems to me that optimizing and learning how to predict if a person will make a purchase (1) or not (0) would be much more beneficial to google than predicting how much people spend, when over 95% of people who come in the store spend nothing.
", "rawMarkdown": "Just thought I'd start this discussion for fun, given this is obviously not going to change for this competition. But does anyone else think the much more appropriate DV here would be a binary whether or not the customer spends?\n\nIt seems to me that optimizing and learning how to predict if a person will make a purchase (1) or not (0) would be much more beneficial to google than predicting how much people spend, when over 95% of people who come in the store spend nothing.", "votes": 2 }, { "id": 390680, "postDate": "2018-09-20T16:27:50.287Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 389997, "author_name": "S D", "author_url": "", "post_date": "2018-09-19T15:16:29.293000", "content": "I would say both might be useful. \nIf you spend a given cost per click to acquire a visitor, this cost might vary depending on your ad campaign, channel, keywords, etc. And you'd want to know how much earnings per click you are generating (ideally you would look at margin, not revenue). That way you can tell whether your campaign is profitable or not.
\n\nFor example, visitors coming from referrals buy 5% of the time whereas visitors coming from display ads spend 2% of the time. However, display ads visitors have average transactions that are twice as big. So just looking at conversion rate (i.e. whether we have a purchase or not) doesn't tell the whole story.
", "votes": 3, "replies": [ { "id": 390005, "author_name": "DylanW", "author_url": "", "post_date": "2018-09-19T15:22:07.037000", "content": "Very true!
", "votes": 0, "replies": [] } ] }, { "id": 390143, "author_name": "Kain", "author_url": "", "post_date": "2018-09-19T19:17:37.760000", "content": "Defining a this competition as something like a CTR (Click Through Rate Prediction) problem was an option but I think they wanted to predict the value of each user interaction, by doing this they going one level deeper to find out how much a user is worth by predicting his expenditure, then we can categorize the users based on their predicted values and create something like CRF loyalty plan for users!
", "votes": 1, "replies": [] }, { "id": 390009, "author_name": "Fernando Constantino", "author_url": "", "post_date": "2018-09-19T15:23:55.697000", "content": "Yes, I agree with you. Given the number of rows with transaction value, and the number of relevant variables with high cardinality, I would have made this competition to just predict whether the person will purchase or not.
", "votes": 1, "replies": [] }, { "id": 389999, "author_name": "Andrey Lukyanenko", "author_url": "", "post_date": "2018-09-19T15:18:54.613000", "content": "I'd say that business usually better understands what it needs :)
\n\nI think that both ideas (revenue prediction or binary classification) are valid, but the company decided that they need information about the revenue. Maybe it will help to target potentially profitable clients.
", "votes": 1, "replies": [ { "id": 390004, "author_name": "DylanW", "author_url": "", "post_date": "2018-09-19T15:21:06.340000", "content": "Oh, I totally agree! I was just having a little thought experiment. My intuition is that getting more people to spend money would have a greater outcome than just determining how we can get those who spend to spend more - but two equally valid sides of the same coin!
", "votes": 1, "replies": [] } ] }, { "id": 391322, "author_name": "Alexander Firsov", "author_url": "", "post_date": "2018-09-21T15:40:57.580000", "content": "What can be useful for host is an interesting question. In this competition we know what already happened, like how many times a visitor clicked during any specific visit, number of visits for an year, etc. And I believe info about visit is more predictive, than geo location for example. I can imagine how geo info can be used for marketing, but not visitor behavior on GStore. So if we can predict whether visitor will buy something or even revenue amount, how to use it? If user just bought it, GStore knows it without ML, but it is not possible to know if visitor will have 22 hits next time when he/she will spend 23000$.
\n\nBtw, what does DV mean? Dependent variable?
", "votes": 2, "replies": [ { "id": 391328, "author_name": "S D", "author_url": "", "post_date": "2018-09-21T15:58:58.853000", "content": "I just created a topic mentioning that particular issue :)
", "votes": 0, "replies": [] }, { "id": 391377, "author_name": "DylanW", "author_url": "", "post_date": "2018-09-21T17:59:42.687000", "content": "True! and Yes, dependent variable; that's my psychology training coming through haha
", "votes": 1, "replies": [] } ] }, { "id": 390680, "author_name": "", "author_url": "", "post_date": "2018-09-20T16:27:50.287000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "389997": "I would say both might be useful. \nIf you spend a given cost per click to acquire a visitor, this cost might vary depending on your ad campaign, channel, keywords, etc. And you'd want to know how much earnings per click you are generating (ideally you would look at margin, not revenue). That way you can tell whether your campaign is profitable or not.\n\nFor example, visitors coming from referrals buy 5% of the time whereas visitors coming from display ads spend 2% of the time. However, display ads visitors have average transactions that are twice as big. So just looking at conversion rate (i.e. whether we have a purchase or not) doesn't tell the whole story.", "390143": "Defining a this competition as something like a CTR (Click Through Rate Prediction) problem was an option but I think they wanted to predict the value of each user interaction, by doing this they going one level deeper to find out how much a user is worth by predicting his expenditure, then we can categorize the users based on their predicted values and create something like CRF loyalty plan for users!", "390009": "Yes, I agree with you. Given the number of rows with transaction value, and the number of relevant variables with high cardinality, I would have made this competition to just predict whether the person will purchase or not.", "389999": "I'd say that business usually better understands what it needs :)\n\nI think that both ideas (revenue prediction or binary classification) are valid, but the company decided that they need information about the revenue. Maybe it will help to target potentially profitable clients.", "391322": "What can be useful for host is an interesting question. In this competition we know what already happened, like how many times a visitor clicked during any specific visit, number of visits for an year, etc. And I believe info about visit is more predictive, than geo location for example. I can imagine how geo info can be used for marketing, but not visitor behavior on GStore. So if we can predict whether visitor will buy something or even revenue amount, how to use it? If user just bought it, GStore knows it without ML, but it is not possible to know if visitor will have 22 hits next time when he/she will spend 23000$. \n\nBtw, what does DV mean? Dependent variable?", "389932": "Just thought I'd start this discussion for fun, given this is obviously not going to change for this competition. But does anyone else think the much more appropriate DV here would be a binary whether or not the customer spends?\n\nIt seems to me that optimizing and learning how to predict if a person will make a purchase (1) or not (0) would be much more beneficial to google than predicting how much people spend, when over 95% of people who come in the store spend nothing.", "390680": "" } }