{ "id": 6312, "title": "No luck with glmnet", "url": "/competitions/ams-2014-solar-energy-prediction-contest/discussion/6312", "author_name": "", "post_date": "2013-11-14T13:58:26.203Z", "votes": null, "comment_count": 4, "views": 2233, "content": "

Hi all

\n

I tried a very simple model glmnet under the assumption that it would work better than the simple Spline Interpolation Benchmark. But somehow it did not! I'm still very new to kaggle and ML in general. I made the code available as a gist https://gist.github.com/leobuettiker/7467042 

\n

I'm highly interested in any feedback. Why does it not work better than the benchmark? Did I do a coding mistake? (I don't hope so!) Is glmnet the wrong model? Are the parameters wrong? Any input that would help me to understand my mistakes is highly appreciated. I hope I might do better on other competitions.

", "messages": [ { "id": "33701", "postDate": "11/14/2013 13:58:26", "content": "

Hi all

\n

I tried a very simple model glmnet under the assumption that it would work better than the simple Spline Interpolation Benchmark. But somehow it did not! I'm still very new to kaggle and ML in general. I made the code available as a gist https://gist.github.com/leobuettiker/7467042 

\n

I'm highly interested in any feedback. Why does it not work better than the benchmark? Did I do a coding mistake? (I don't hope so!) Is glmnet the wrong model? Are the parameters wrong? Any input that would help me to understand my mistakes is highly appreciated. I hope I might do better on other competitions.

", "rawMarkdown": "", "votes": null }, { "id": "33708", "postDate": "11/14/2013 15:43:40", "content": "

i didnt fully grasp your model, but i think you should normalize your inputs first.

", "rawMarkdown": "", "votes": null }, { "id": "33719", "postDate": "11/14/2013 19:32:24", "content": "

Thank you for your answer. I basically take all available date (from all stations) and try to build a model over this date for each station. I hope glmnet does feature selection for me. As far as I understand the glmnet documentation it should do normalization automatically. I tried to do it by myself before the model creation and got a even worse result.  

", "rawMarkdown": "", "votes": null }, { "id": "33720", "postDate": "11/14/2013 19:47:45", "content": "

I think you should be able to get around 2,240K with glmnet.  There's a benchmark at http://fastml.com/predicting-solar-energy-from-weather-forecasts-plus-a-netcdf4-tutorial/ that does so in Python.

\n

I didn't look super carefully at your code, but it looks like you're only using "dswrf_sfc".  I tried ignoring the other files for a short time too, but they seem to be useful.

\n

Also, I don't think that the spline interpolation benchmark is horrible.  Splines have had a lot of smart people think about them for a long time, and a lot of that work was specifically for interpolation.  Sometimes the benchmarks are really weak (like the random one), and sometimes they are not bad (like this one, or the Expedia one).  Beating the better ones will generally require more than putting the same data into a slightly different algorithm :)

", "rawMarkdown": "", "votes": null }, { "id": "35033", "postDate": "11/22/2013 11:32:22", "content": "

Thanks for your answer Mike!

\n

I tried some models with more date set. But unfortunately I did not get any good results, probably I have a bug somewhere. But I think now it's time to move on.

\n

I do think the spline benchmark is a good one. But I'm still surprised that It was not easy to beat it with using the training data.

", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 33708, "author_name": "leustagos", "author_url": "", "post_date": "11/14/2013 15:43:40", "content": "

i didnt fully grasp your model, but i think you should normalize your inputs first.

", "votes": null, "replies": [] }, { "id": 33719, "author_name": "leobuettiker", "author_url": "", "post_date": "11/14/2013 19:32:24", "content": "

Thank you for your answer. I basically take all available date (from all stations) and try to build a model over this date for each station. I hope glmnet does feature selection for me. As far as I understand the glmnet documentation it should do normalization automatically. I tried to do it by myself before the model creation and got a even worse result.  

", "votes": null, "replies": [] }, { "id": 33720, "author_name": "mikescarpati", "author_url": "", "post_date": "11/14/2013 19:47:45", "content": "

I think you should be able to get around 2,240K with glmnet.  There's a benchmark at http://fastml.com/predicting-solar-energy-from-weather-forecasts-plus-a-netcdf4-tutorial/ that does so in Python.

\n

I didn't look super carefully at your code, but it looks like you're only using "dswrf_sfc".  I tried ignoring the other files for a short time too, but they seem to be useful.

\n

Also, I don't think that the spline interpolation benchmark is horrible.  Splines have had a lot of smart people think about them for a long time, and a lot of that work was specifically for interpolation.  Sometimes the benchmarks are really weak (like the random one), and sometimes they are not bad (like this one, or the Expedia one).  Beating the better ones will generally require more than putting the same data into a slightly different algorithm :)

", "votes": null, "replies": [] }, { "id": 35033, "author_name": "leobuettiker", "author_url": "", "post_date": "11/22/2013 11:32:22", "content": "

Thanks for your answer Mike!

\n

I tried some models with more date set. But unfortunately I did not get any good results, probably I have a bug somewhere. But I think now it's time to move on.

\n

I do think the spline benchmark is a good one. But I'm still surprised that It was not easy to beat it with using the training data.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "33701": "", "33708": "", "33719": "", "33720": "", "35033": "" }, "source": "meta" }