To help you think about AutoML, these are usually paid and sometimes unpaid services that help you with automating the pipeline of solving your problem with ML without needing much understanding or effort like utilising enterprise services like IBM Watson/Azure ML/AWS, AutoKeras, Google AutoML . This might require making API calls or using the enterprise tools.
In professional world, faster is better, so you could use AutoML, but for competitions to create a fair playing ground, only open source libraries or algoritems on a specific license and also for organisers it might cost them some money to scale your solution.
So, yes you can use any RNN or NN variant algorithm from open source frameworks Tensorflow/Pytorch, same way you can use any ML algorithm from Sckit Learn
Come on dr ice, be cool, use your creativity and impressive skillset to come up with an amazing solution. Look at flamethrower, he is so hot on the LB, we can do with a bit of winter. (r)nn != automl (you should know this well).
Thanks doc, I appreciate the feedback and I'm from sunny south africa.
This is a tricky competition - more data cleaning, less machine learning, perhaps a bit of inspiration!
I played a lot earlier, before somebody discovered just by using median you can actually score quite good. I've not stopped, but this one for me is boring at the moment.
Not sure what flame and tomy are doing, must be special though. I see they are both from Nigeria, as are you!? Perhaps it is something in the water, so you should definitely give this one a shot.
Seriously though, I see you have an excellent zindi track record and have stats in your skillset, so you should have some good ideas for this one!
Actually doc, I've tried lots of the usual time series models for this one (e.g. arima and exp smoothing) and actually got good results (I was inside of top 20 before mediam model took over).
Nothing amazing, but not too bad either. NN and more complex models simply did not work for me here.
Now since then I tried hard to get a more complex model, some NN based wonder that can reclaim some LB space. Still a bit early, and I have my doubts on the success of NN here, but will see soon.
For me the alternative to explore, given the median model, is to model around the median. That is really a boring option, perhaps fitting for this data, but so I hope to find something elsewhere to pursue.
Hi,
To help you think about AutoML, these are usually paid and sometimes unpaid services that help you with automating the pipeline of solving your problem with ML without needing much understanding or effort like utilising enterprise services like IBM Watson/Azure ML/AWS, AutoKeras, Google AutoML . This might require making API calls or using the enterprise tools.
In professional world, faster is better, so you could use AutoML, but for competitions to create a fair playing ground, only open source libraries or algoritems on a specific license and also for organisers it might cost them some money to scale your solution.
So, yes you can use any RNN or NN variant algorithm from open source frameworks Tensorflow/Pytorch, same way you can use any ML algorithm from Sckit Learn
Succinct! I get it now. Thanks Chief! Keep throwing them flames!!!
Come on dr ice, be cool, use your creativity and impressive skillset to come up with an amazing solution. Look at flamethrower, he is so hot on the LB, we can do with a bit of winter. (r)nn != automl (you should know this well).
Man like Skaak!, I like your manner of speaking though, it's kinda wierd.
Saw your discuss earlier, couldn't quite place the accent; where you from?
In any case, appreciate the feedback!
Thanks doc, I appreciate the feedback and I'm from sunny south africa.
This is a tricky competition - more data cleaning, less machine learning, perhaps a bit of inspiration!
I played a lot earlier, before somebody discovered just by using median you can actually score quite good. I've not stopped, but this one for me is boring at the moment.
Not sure what flame and tomy are doing, must be special though. I see they are both from Nigeria, as are you!? Perhaps it is something in the water, so you should definitely give this one a shot.
Seriously though, I see you have an excellent zindi track record and have stats in your skillset, so you should have some good ideas for this one!
Actually doc, I've tried lots of the usual time series models for this one (e.g. arima and exp smoothing) and actually got good results (I was inside of top 20 before mediam model took over).
Nothing amazing, but not too bad either. NN and more complex models simply did not work for me here.
Now since then I tried hard to get a more complex model, some NN based wonder that can reclaim some LB space. Still a bit early, and I have my doubts on the success of NN here, but will see soon.
For me the alternative to explore, given the median model, is to model around the median. That is really a boring option, perhaps fitting for this data, but so I hope to find something elsewhere to pursue.