![\"enter](\"https://i.imgur.com/QkATelm.png\")
How is deal probability calculated? On Avito, you must click on a button to reveal seller information. I assume Avito computes deal probability based on a combination of how many people view the ad and how many people click the button. Taking this assumption into account, I think people click the button for two reasons:
\n\nDuring my analysis of missing values, I produced the above graphic. About 30% of the top 10% deal_prob percentile are missing prices! I believe this, and among other things I've found, show that ambiguity is causing buyers to reach out to the seller and that action is being interpreted as a potential deal.
\n\nIt will be interesting to see whos solution will deal with this conflation of motives.
", "messages": [ { "id": "327042", "postDate": "05/10/2018 18:19:58", "content": "
How is deal probability calculated? On Avito, you must click on a button to reveal seller information. I assume Avito computes deal probability based on a combination of how many people view the ad and how many people click the button. Taking this assumption into account, I think people click the button for two reasons:
\n\nDuring my analysis of missing values, I produced the above graphic. About 30% of the top 10% deal_prob percentile are missing prices! I believe this, and among other things I've found, show that ambiguity is causing buyers to reach out to the seller and that action is being interpreted as a potential deal.
\n\nIt will be interesting to see whos solution will deal with this conflation of motives.
", "rawMarkdown": "![enter image description here][1]\n\nHow is deal probability calculated? On Avito, you must click on a button to reveal seller information. I assume Avito computes deal probability based on a combination of how many people view the ad and how many people click the button. Taking this assumption into account, I think people click the button for two reasons: \n\n 1. They want to contact the seller to purchace the item. \n 2. They want to contact the seller for more information. \n\nDuring my analysis of missing values, I produced the above graphic. About 30% of the top 10% deal_prob percentile are missing prices! I believe this, and among other things I've found, show that ambiguity is causing buyers to reach out to the seller and that action is being interpreted as a potential deal. \n\nIt will be interesting to see whos solution will deal with this conflation of motives. \n\n [1]: https://i.imgur.com/QkATelm.png", "votes": null }, { "id": "327112", "postDate": "05/10/2018 20:44:13", "content": "Excellent insight! Thank you!
", "rawMarkdown": "Excellent insight! Thank you!", "votes": null }, { "id": "327250", "postDate": "05/11/2018 05:52:51", "content": "Model will automatically take care of that, I believe as you will surely impute those values with a unique number most probably a -ve one. So the co- relation of that imputed rows should be high with deal probability
", "rawMarkdown": "Model will automatically take care of that, I believe as you will surely impute those values with a unique number most probably a -ve one. So the co- relation of that imputed rows should be high with deal probability", "votes": null }, { "id": "327301", "postDate": "05/11/2018 08:21:12", "content": "See my answer in https://www.kaggle.com/enfeizhan/does-missing-price-mean-price-negotiable/code this could help.
", "rawMarkdown": "See my answer in https://www.kaggle.com/enfeizhan/does-missing-price-mean-price-negotiable/code this could help.", "votes": null }, { "id": "327329", "postDate": "05/11/2018 09:48:54", "content": "Most of such ads are either from services category which don't have fixed price or they are items which are given away for free. Also sometimes people write price in description, mostly when they sell several items in one ad.
", "rawMarkdown": "Most of such ads are either from services category which don't have fixed price or they are items which are given away for free. Also sometimes people write price in description, mostly when they sell several items in one ad.", "votes": null }, { "id": "327453", "postDate": "05/11/2018 15:28:00", "content": "Interesting point and a possible business explanation.
\n\nAs long as you use a model that takes care of missing (or encoded missing) it should take care of that correlation. That's why most people prefer tree methods for that kind of heterogeneous with missings data.
", "rawMarkdown": "Interesting point and a possible business explanation.\n\nAs long as you use a model that takes care of missing (or encoded missing) it should take care of that correlation. That's why most people prefer tree methods for that kind of heterogeneous with missings data.", "votes": null }, { "id": "327841", "postDate": "05/12/2018 16:03:13", "content": "Very interesting thought. Following this thought that the \"deal probability\" is the ratio of how many people view the ad and how many people click the button, I am assuming the ones with the attractive picture but disappointing descriptions may result in an even lower deal probability... is that true?
\n\nAlso, I do a bit explore on Avito. They actually offer a paid service to place VIP customers ads on top, which also leads to higher viewing rate but low deal_probability
", "rawMarkdown": "Very interesting thought. Following this thought that the \"deal probability\" is the ratio of how many people view the ad and how many people click the button, I am assuming the ones with the attractive picture but disappointing descriptions may result in an even lower deal probability... is that true?\n\nAlso, I do a bit explore on Avito. They actually offer a paid service to place VIP customers ads on top, which also leads to higher viewing rate but low deal_probability", "votes": null }, { "id": "328131", "postDate": "05/13/2018 12:40:28", "content": "That's a nice observation. And warning to those using price.fillna(0) which will hide this.
", "rawMarkdown": "That's a nice observation. And warning to those using price.fillna(0) which will hide this.", "votes": null }, { "id": "328701", "postDate": "05/14/2018 23:34:46", "content": "What you are saying is that deal probability is the CTR (click-through rate). But based on what they say (\"This is the likelihood that an ad actually sold something\") I think it is the conversion rate, which is a step further from CTR. The CVR is calculated as number of actual conversions (i.e., people buying the product or hiring the service) divided by number of clicks on the ad. Also, if you take into account that they say \"It's not possible to verify every transaction with certainty, so this column's value can be any float from zero to one.\", that makes it more likely that it's the CVR, because convertions cannot always be tracked (maybe the advertiser left a number and the customer contacted him directly), unlike the CTR which you can always track, because a print and a click are events that obviously involve Avito's backend.
", "rawMarkdown": "What you are saying is that deal probability is the CTR (click-through rate). But based on what they say (\"This is the likelihood that an ad actually sold something\") I think it is the conversion rate, which is a step further from CTR. The CVR is calculated as number of actual conversions (i.e., people buying the product or hiring the service) divided by number of clicks on the ad. Also, if you take into account that they say \"It's not possible to verify every transaction with certainty, so this column's value can be any float from zero to one.\", that makes it more likely that it's the CVR, because convertions cannot always be tracked (maybe the advertiser left a number and the customer contacted him directly), unlike the CTR which you can always track, because a print and a click are events that obviously involve Avito's backend.", "votes": null }, { "id": "329838", "postDate": "05/17/2018 08:59:55", "content": "I think there is one more component of deal probability calculation - time when the ad is closed. The problem is that it is hard to get this info from data we have. Also, if ad sells multiple items, in case of shop, the ad will not be closed when single item is sold.
", "rawMarkdown": "I think there is one more component of deal probability calculation - time when the ad is closed. The problem is that it is hard to get this info from data we have. Also, if ad sells multiple items, in case of shop, the ad will not be closed when single item is sold.", "votes": null }, { "id": "329955", "postDate": "05/17/2018 16:44:28", "content": "On Avito, they do not have a direct record of who purchased what because sales are done directly between the seller and buyer. It's like the Russian craigs list.
", "rawMarkdown": "On Avito, they do not have a direct record of who purchased what because sales are done directly between the seller and buyer. It's like the Russian craigs list.", "votes": null }, { "id": "342256", "postDate": "06/13/2018 07:20:01", "content": "A number of other users and I have come to the conclusion that ads with more missing info in general (not just the 'price' column) do seem to get more interaction. This means that the reason missing info ads get more interaction is because users are contacting the seller in need of more information.
\n\nThis conclusion does seem to back up your hypothesis that deal_probability is actually CTR (click-through rate) and not CVR (conversion rate).
\n\nCVR has the formula: number of actual conversions (i.e., people buying the product or hiring the service)/number of clicks on the ad. If deal_probability was indeed CVR, then deal_probability would actually go down as the number of clicks, or interactions on the ad, goes up.
\n\nI might be wrong though. Please correct me if I am.
", "rawMarkdown": "A number of other users and I have come to the conclusion that ads with more missing info in general (not just the 'price' column) do seem to get more interaction. This means that the reason missing info ads get more interaction is because users are contacting the seller in need of more information. \n\nThis conclusion does seem to back up your hypothesis that deal_probability is actually CTR (click-through rate) and not CVR (conversion rate). \n\nCVR has the formula: number of actual conversions (i.e., people buying the product or hiring the service)/number of clicks on the ad. If deal_probability was indeed CVR, then deal_probability would actually go down as the number of clicks, or interactions on the ad, goes up.\n\nI might be wrong though. Please correct me if I am.", "votes": null } ], "comments": [ { "id": 327112, "author_name": "marketneutral", "author_url": "", "post_date": "05/10/2018 20:44:13", "content": "Excellent insight! Thank you!
", "votes": null, "replies": [] }, { "id": 327250, "author_name": "vishaljindal0", "author_url": "", "post_date": "05/11/2018 05:52:51", "content": "Model will automatically take care of that, I believe as you will surely impute those values with a unique number most probably a -ve one. So the co- relation of that imputed rows should be high with deal probability
", "votes": null, "replies": [] }, { "id": 327301, "author_name": "rinatmks", "author_url": "", "post_date": "05/11/2018 08:21:12", "content": "See my answer in https://www.kaggle.com/enfeizhan/does-missing-price-mean-price-negotiable/code this could help.
", "votes": null, "replies": [] }, { "id": 327329, "author_name": "dariakharlan", "author_url": "", "post_date": "05/11/2018 09:48:54", "content": "Most of such ads are either from services category which don't have fixed price or they are items which are given away for free. Also sometimes people write price in description, mostly when they sell several items in one ad.
", "votes": null, "replies": [] }, { "id": 327453, "author_name": "arroqc", "author_url": "", "post_date": "05/11/2018 15:28:00", "content": "Interesting point and a possible business explanation.
\n\nAs long as you use a model that takes care of missing (or encoded missing) it should take care of that correlation. That's why most people prefer tree methods for that kind of heterogeneous with missings data.
", "votes": null, "replies": [] }, { "id": 327841, "author_name": "zehaiwang", "author_url": "", "post_date": "05/12/2018 16:03:13", "content": "Very interesting thought. Following this thought that the \"deal probability\" is the ratio of how many people view the ad and how many people click the button, I am assuming the ones with the attractive picture but disappointing descriptions may result in an even lower deal probability... is that true?
\n\nAlso, I do a bit explore on Avito. They actually offer a paid service to place VIP customers ads on top, which also leads to higher viewing rate but low deal_probability
", "votes": null, "replies": [] }, { "id": 328131, "author_name": "seroppi", "author_url": "", "post_date": "05/13/2018 12:40:28", "content": "That's a nice observation. And warning to those using price.fillna(0) which will hide this.
", "votes": null, "replies": [] }, { "id": 328701, "author_name": "axelstram", "author_url": "", "post_date": "05/14/2018 23:34:46", "content": "What you are saying is that deal probability is the CTR (click-through rate). But based on what they say (\"This is the likelihood that an ad actually sold something\") I think it is the conversion rate, which is a step further from CTR. The CVR is calculated as number of actual conversions (i.e., people buying the product or hiring the service) divided by number of clicks on the ad. Also, if you take into account that they say \"It's not possible to verify every transaction with certainty, so this column's value can be any float from zero to one.\", that makes it more likely that it's the CVR, because convertions cannot always be tracked (maybe the advertiser left a number and the customer contacted him directly), unlike the CTR which you can always track, because a print and a click are events that obviously involve Avito's backend.
", "votes": null, "replies": [ { "id": 329955, "author_name": "andrewrib", "author_url": "", "post_date": "05/17/2018 16:44:28", "content": "On Avito, they do not have a direct record of who purchased what because sales are done directly between the seller and buyer. It's like the Russian craigs list.
", "votes": null, "replies": [] } ] }, { "id": 329838, "author_name": "alexfir", "author_url": "", "post_date": "05/17/2018 08:59:55", "content": "I think there is one more component of deal probability calculation - time when the ad is closed. The problem is that it is hard to get this info from data we have. Also, if ad sells multiple items, in case of shop, the ad will not be closed when single item is sold.
", "votes": null, "replies": [] }, { "id": 342256, "author_name": "iandzindo", "author_url": "", "post_date": "06/13/2018 07:20:01", "content": "A number of other users and I have come to the conclusion that ads with more missing info in general (not just the 'price' column) do seem to get more interaction. This means that the reason missing info ads get more interaction is because users are contacting the seller in need of more information.
\n\nThis conclusion does seem to back up your hypothesis that deal_probability is actually CTR (click-through rate) and not CVR (conversion rate).
\n\nCVR has the formula: number of actual conversions (i.e., people buying the product or hiring the service)/number of clicks on the ad. If deal_probability was indeed CVR, then deal_probability would actually go down as the number of clicks, or interactions on the ad, goes up.
\n\nI might be wrong though. Please correct me if I am.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "327042": "![enter image description here][1]\n\nHow is deal probability calculated? On Avito, you must click on a button to reveal seller information. I assume Avito computes deal probability based on a combination of how many people view the ad and how many people click the button. Taking this assumption into account, I think people click the button for two reasons: \n\n 1. They want to contact the seller to purchace the item. \n 2. They want to contact the seller for more information. \n\nDuring my analysis of missing values, I produced the above graphic. About 30% of the top 10% deal_prob percentile are missing prices! I believe this, and among other things I've found, show that ambiguity is causing buyers to reach out to the seller and that action is being interpreted as a potential deal. \n\nIt will be interesting to see whos solution will deal with this conflation of motives. \n\n [1]: https://i.imgur.com/QkATelm.png", "327112": "Excellent insight! Thank you!", "327250": "Model will automatically take care of that, I believe as you will surely impute those values with a unique number most probably a -ve one. So the co- relation of that imputed rows should be high with deal probability", "327301": "See my answer in https://www.kaggle.com/enfeizhan/does-missing-price-mean-price-negotiable/code this could help.", "327329": "Most of such ads are either from services category which don't have fixed price or they are items which are given away for free. Also sometimes people write price in description, mostly when they sell several items in one ad.", "327453": "Interesting point and a possible business explanation.\n\nAs long as you use a model that takes care of missing (or encoded missing) it should take care of that correlation. That's why most people prefer tree methods for that kind of heterogeneous with missings data.", "327841": "Very interesting thought. Following this thought that the \"deal probability\" is the ratio of how many people view the ad and how many people click the button, I am assuming the ones with the attractive picture but disappointing descriptions may result in an even lower deal probability... is that true?\n\nAlso, I do a bit explore on Avito. They actually offer a paid service to place VIP customers ads on top, which also leads to higher viewing rate but low deal_probability", "328131": "That's a nice observation. And warning to those using price.fillna(0) which will hide this.", "328701": "What you are saying is that deal probability is the CTR (click-through rate). But based on what they say (\"This is the likelihood that an ad actually sold something\") I think it is the conversion rate, which is a step further from CTR. The CVR is calculated as number of actual conversions (i.e., people buying the product or hiring the service) divided by number of clicks on the ad. Also, if you take into account that they say \"It's not possible to verify every transaction with certainty, so this column's value can be any float from zero to one.\", that makes it more likely that it's the CVR, because convertions cannot always be tracked (maybe the advertiser left a number and the customer contacted him directly), unlike the CTR which you can always track, because a print and a click are events that obviously involve Avito's backend.", "329838": "I think there is one more component of deal probability calculation - time when the ad is closed. The problem is that it is hard to get this info from data we have. Also, if ad sells multiple items, in case of shop, the ad will not be closed when single item is sold.", "329955": "On Avito, they do not have a direct record of who purchased what because sales are done directly between the seller and buyer. It's like the Russian craigs list.", "342256": "A number of other users and I have come to the conclusion that ads with more missing info in general (not just the 'price' column) do seem to get more interaction. This means that the reason missing info ads get more interaction is because users are contacting the seller in need of more information. \n\nThis conclusion does seem to back up your hypothesis that deal_probability is actually CTR (click-through rate) and not CVR (conversion rate). \n\nCVR has the formula: number of actual conversions (i.e., people buying the product or hiring the service)/number of clicks on the ad. If deal_probability was indeed CVR, then deal_probability would actually go down as the number of clicks, or interactions on the ad, goes up.\n\nI might be wrong though. Please correct me if I am." }, "source": "meta" }