Hello everyone,
\n\nI've been scavenging the forums looking for a potential missing piece of the puzzle. I'm quite convinced that my next step is to differentiate between different types of supernovas (SNs). As Kyle Boon says, the somewhat obvious way to distinguish SN between each other is to compute their peak luminosity along with the width of the light curve's. In other words looking at the decline rate after the peak luminosity would seem to be a good idea.
\n\nThe issue I'm having is that a lot of the light curves are incomplete, and thus we don't really know their peak luminosity. It follows that without knowing the peak luminosity we can't compute the subsequent decline rates. I suppose that we have to \"guess\" where the peak luminosity is along with it's value. Maybe that gaussian processes are the way to go but I haven't had any success.
\n\nHas anyone else been thinking along these lines?
", "messages": [ { "id": "434442", "postDate": "12/06/2018 12:08:46", "content": "Hello everyone,
\n\nI've been scavenging the forums looking for a potential missing piece of the puzzle. I'm quite convinced that my next step is to differentiate between different types of supernovas (SNs). As Kyle Boon says, the somewhat obvious way to distinguish SN between each other is to compute their peak luminosity along with the width of the light curve's. In other words looking at the decline rate after the peak luminosity would seem to be a good idea.
\n\nThe issue I'm having is that a lot of the light curves are incomplete, and thus we don't really know their peak luminosity. It follows that without knowing the peak luminosity we can't compute the subsequent decline rates. I suppose that we have to \"guess\" where the peak luminosity is along with it's value. Maybe that gaussian processes are the way to go but I haven't had any success.
\n\nHas anyone else been thinking along these lines?
", "rawMarkdown": "Hello everyone,\n\nI've been scavenging the forums looking for a potential missing piece of the puzzle. I'm quite convinced that my next step is to differentiate between different types of supernovas (SNs). As Kyle Boon says, the somewhat obvious way to distinguish SN between each other is to compute their peak luminosity along with the width of the light curve's. In other words looking at the decline rate after the peak luminosity would seem to be a good idea.\n\nThe issue I'm having is that a lot of the light curves are incomplete, and thus we don't really know their peak luminosity. It follows that without knowing the peak luminosity we can't compute the subsequent decline rates. I suppose that we have to \"guess\" where the peak luminosity is along with it's value. Maybe that gaussian processes are the way to go but I haven't had any success.\n\nHas anyone else been thinking along these lines?", "votes": null }, { "id": "434478", "postDate": "12/06/2018 13:16:13", "content": "Although the peak is not there, it will often be not far away, perhaps only a few days. So my thought is to use the local rate of decline rather than the relation to the peak value. I tried to use a GAM approximation over all the passbands, but there seems to be too few measurements to do that. (For object_id > 1000000 there are only 130 measurements per object, for obejct_id < 1000000 there are apr 330 measurement per object (this is true for both train and test sets)). So now I will try to use a linear approximation in each passband and find the rates from that.
", "rawMarkdown": "Although the peak is not there, it will often be not far away, perhaps only a few days. So my thought is to use the local rate of decline rather than the relation to the peak value. I tried to use a GAM approximation over all the passbands, but there seems to be too few measurements to do that. (For object_id > 1000000 there are only 130 measurements per object, for obejct_id < 1000000 there are apr 330 measurement per object (this is true for both train and test sets)). So now I will try to use a linear approximation in each passband and find the rates from that.", "votes": null }, { "id": "434491", "postDate": "12/06/2018 13:42:21", "content": "I agree that the peak is probably close in most cases, but the question is how to quantify it.
", "rawMarkdown": "I agree that the peak is probably close in most cases, but the question is how to quantify it.", "votes": null }, { "id": "434569", "postDate": "12/06/2018 15:52:47", "content": "My point is: The peak is nice to have, but perhaps not necessary.
\n\nTry to take another look at the The PLAsTiCC Astronomy \"Starter Kit\" Figure 8: The diversity of Supernovae Light Curves.
\n\nIt seems that the gradient of the light curve might be a good substitute for the flux/flux-peak value.
\n\nF.ex. Type 1a SNe has a steep gradient between day 10 and day 30 where as Type 11 SNe does not, but Type 11 SNe has a steep gradient between day 100 and day 130. So if one can estimate the gradient in different time intervals around the peak it might be a good help to classification. If you try it out, then remember that Figure 8 has a logarithmic scale while the given flux is non-logarithmic.
\n\nIn the discussion https://www.kaggle.com/c/PLAsTiCC-2018/discussion/72646 Mithrillion has tried to estimate the peak with help of an autoencoder: but well, it seems difficult at best.
\n\nIf I get the time I will try to curve fit a skewed gaussian distribution, I think it could yield useful parameters even if the peak is a little off.
", "rawMarkdown": "My point is: The peak is nice to have, but perhaps not necessary.\n\nTry to take another look at the The PLAsTiCC Astronomy \"Starter Kit\" Figure 8: The diversity of Supernovae Light Curves.\n\nIt seems that the *gradient* of the light curve might be a good substitute for the flux/flux-peak value.\n\nF.ex. Type 1a SNe has a steep gradient between day 10 and day 30 where as Type 11 SNe does not, but Type 11 SNe has a steep gradient between day 100 and day 130. So if one can estimate the gradient in different time intervals around the peak it might be a good help to classification. If you try it out, then remember that Figure 8 has a logarithmic scale while the given flux is non-logarithmic.\n\nIn the discussion https://www.kaggle.com/c/PLAsTiCC-2018/discussion/72646 Mithrillion has tried to estimate the peak with help of an autoencoder: but well, it seems difficult at best.\n\nIf I get the time I will try to curve fit a skewed gaussian distribution, I think it could yield useful parameters even if the peak is a little off.", "votes": null }, { "id": "434572", "postDate": "12/06/2018 16:01:05", "content": "I think I understand what you're getting at. My issue with the starter kit kernal is that the light curves used as examples are nothing like the ones we have to with. Ours seem to be much more sparse, alas.
", "rawMarkdown": "I think I understand what you're getting at. My issue with the starter kit kernal is that the light curves used as examples are nothing like the ones we have to with. Ours seem to be much more sparse, alas.", "votes": null }, { "id": "436143", "postDate": "12/09/2018 17:17:58", "content": "@Peter Yes, but if you don't have the peak, how do you count the time elapsed from the brightness peak ?
", "rawMarkdown": "Peter Yes, but if you don't have the peak, how do you count the time elapsed from the brightness peak ?", "votes": null }, { "id": "436532", "postDate": "12/10/2018 13:35:41", "content": "Yes, that is where I just assume that when the peak is very close it is the same \"is there\". I don't know what else to do.
", "rawMarkdown": "Yes, that is where I just assume that when the peak is very close it is the same \"is there\". I don't know what else to do.", "votes": null }, { "id": "438047", "postDate": "12/13/2018 02:44:19", "content": "i am having trouble with this one as well. i have read that the decay rate really helps but so far my attempts in incorporating the decay rate i calculated doesn't help my model
", "rawMarkdown": "i am having trouble with this one as well. i have read that the decay rate really helps but so far my attempts in incorporating the decay rate i calculated doesn't help my model", "votes": null } ], "comments": [ { "id": 434478, "author_name": "petersorensen360", "author_url": "", "post_date": "12/06/2018 13:16:13", "content": "Although the peak is not there, it will often be not far away, perhaps only a few days. So my thought is to use the local rate of decline rather than the relation to the peak value. I tried to use a GAM approximation over all the passbands, but there seems to be too few measurements to do that. (For object_id > 1000000 there are only 130 measurements per object, for obejct_id < 1000000 there are apr 330 measurement per object (this is true for both train and test sets)). So now I will try to use a linear approximation in each passband and find the rates from that.
", "votes": null, "replies": [ { "id": 434491, "author_name": "maxhalford", "author_url": "", "post_date": "12/06/2018 13:42:21", "content": "I agree that the peak is probably close in most cases, but the question is how to quantify it.
", "votes": null, "replies": [] }, { "id": 434569, "author_name": "petersorensen360", "author_url": "", "post_date": "12/06/2018 15:52:47", "content": "My point is: The peak is nice to have, but perhaps not necessary.
\n\nTry to take another look at the The PLAsTiCC Astronomy \"Starter Kit\" Figure 8: The diversity of Supernovae Light Curves.
\n\nIt seems that the gradient of the light curve might be a good substitute for the flux/flux-peak value.
\n\nF.ex. Type 1a SNe has a steep gradient between day 10 and day 30 where as Type 11 SNe does not, but Type 11 SNe has a steep gradient between day 100 and day 130. So if one can estimate the gradient in different time intervals around the peak it might be a good help to classification. If you try it out, then remember that Figure 8 has a logarithmic scale while the given flux is non-logarithmic.
\n\nIn the discussion https://www.kaggle.com/c/PLAsTiCC-2018/discussion/72646 Mithrillion has tried to estimate the peak with help of an autoencoder: but well, it seems difficult at best.
\n\nIf I get the time I will try to curve fit a skewed gaussian distribution, I think it could yield useful parameters even if the peak is a little off.
", "votes": null, "replies": [] }, { "id": 434572, "author_name": "maxhalford", "author_url": "", "post_date": "12/06/2018 16:01:05", "content": "I think I understand what you're getting at. My issue with the starter kit kernal is that the light curves used as examples are nothing like the ones we have to with. Ours seem to be much more sparse, alas.
", "votes": null, "replies": [] }, { "id": 436143, "author_name": "vlarmet", "author_url": "", "post_date": "12/09/2018 17:17:58", "content": "@Peter Yes, but if you don't have the peak, how do you count the time elapsed from the brightness peak ?
", "votes": null, "replies": [] }, { "id": 436532, "author_name": "petersorensen360", "author_url": "", "post_date": "12/10/2018 13:35:41", "content": "Yes, that is where I just assume that when the peak is very close it is the same \"is there\". I don't know what else to do.
", "votes": null, "replies": [] } ] }, { "id": 438047, "author_name": "niclasdoce", "author_url": "", "post_date": "12/13/2018 02:44:19", "content": "i am having trouble with this one as well. i have read that the decay rate really helps but so far my attempts in incorporating the decay rate i calculated doesn't help my model
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "434442": "Hello everyone,\n\nI've been scavenging the forums looking for a potential missing piece of the puzzle. I'm quite convinced that my next step is to differentiate between different types of supernovas (SNs). As Kyle Boon says, the somewhat obvious way to distinguish SN between each other is to compute their peak luminosity along with the width of the light curve's. In other words looking at the decline rate after the peak luminosity would seem to be a good idea.\n\nThe issue I'm having is that a lot of the light curves are incomplete, and thus we don't really know their peak luminosity. It follows that without knowing the peak luminosity we can't compute the subsequent decline rates. I suppose that we have to \"guess\" where the peak luminosity is along with it's value. Maybe that gaussian processes are the way to go but I haven't had any success.\n\nHas anyone else been thinking along these lines?", "434478": "Although the peak is not there, it will often be not far away, perhaps only a few days. So my thought is to use the local rate of decline rather than the relation to the peak value. I tried to use a GAM approximation over all the passbands, but there seems to be too few measurements to do that. (For object_id > 1000000 there are only 130 measurements per object, for obejct_id < 1000000 there are apr 330 measurement per object (this is true for both train and test sets)). So now I will try to use a linear approximation in each passband and find the rates from that.", "434491": "I agree that the peak is probably close in most cases, but the question is how to quantify it.", "434569": "My point is: The peak is nice to have, but perhaps not necessary.\n\nTry to take another look at the The PLAsTiCC Astronomy \"Starter Kit\" Figure 8: The diversity of Supernovae Light Curves.\n\nIt seems that the *gradient* of the light curve might be a good substitute for the flux/flux-peak value.\n\nF.ex. Type 1a SNe has a steep gradient between day 10 and day 30 where as Type 11 SNe does not, but Type 11 SNe has a steep gradient between day 100 and day 130. So if one can estimate the gradient in different time intervals around the peak it might be a good help to classification. If you try it out, then remember that Figure 8 has a logarithmic scale while the given flux is non-logarithmic.\n\nIn the discussion https://www.kaggle.com/c/PLAsTiCC-2018/discussion/72646 Mithrillion has tried to estimate the peak with help of an autoencoder: but well, it seems difficult at best.\n\nIf I get the time I will try to curve fit a skewed gaussian distribution, I think it could yield useful parameters even if the peak is a little off.", "434572": "I think I understand what you're getting at. My issue with the starter kit kernal is that the light curves used as examples are nothing like the ones we have to with. Ours seem to be much more sparse, alas.", "436143": "Peter Yes, but if you don't have the peak, how do you count the time elapsed from the brightness peak ?", "436532": "Yes, that is where I just assume that when the peak is very close it is the same \"is there\". I don't know what else to do.", "438047": "i am having trouble with this one as well. i have read that the decay rate really helps but so far my attempts in incorporating the decay rate i calculated doesn't help my model" }, "source": "meta" }