Time Series Analysis OR Classifications?

d_avolad_avola Member, University Professor Posts: 2 University Professor
edited October 2019 in Help
I have a time series dataset of the # of cell phone connections sorted by 2G, 3G & 4G technology platforms. In addition, I have data on cell phone storage bandwidths, processor speeds and camera megapixel bands over the same time period.   I want to determine which cell phone feature (ie. storage b/w, processor speed and/or megapixel b/w) can predict the growth or decline in cell phone connections for all technology platforms.   Not certain which statistical models works best?  Any help or advice is appreciated.


  • Options
    varunm1varunm1 Moderator, Member Posts: 1,207 Unicorn
    Hello @d_avola

    If your data have time-related components, you can create a time series processing (using appropriate windowing operators) and then use classification models to predict the growth or decline in connections. To get the feature importance, there are multiple ways, algorithm-specific importances can be given by Gradient boosted Tree and Random Forest. You can also look at the Explain predictions operator that provides feature importances for individual observations in your dataset, we calculate global weights from these local importances and use them as well to explain the overall importance of each feature based on the local importances.

    Here is one great explanation from @tftemme related to time series processing 

    Hope this helps. Please inform us if you are looking for any specific information.

    Be Safe. Follow precautions and Maintain Social Distancing

  • Options
    sgenzersgenzer Administrator, Moderator, Employee, RapidMiner Certified Analyst, Community Manager, Member, University Professor, PM Moderator Posts: 2,959 Community Manager
    that's an interesting problem @d_avola. I guess my first thought is - are you sure you want this to be a time series problem? It looks to me like you're trying to just predict "growth" or "decline". My engineering brain tells me to start simple. What horizon are you looking for? Say it's growth/decline after 6 months of a feature being introduced (I'm just making up 6 months - do whatever you want). I would grab the values + lag the data set 6 months back. The current values -> predicted values; the lagged values are your training/testing set. Does that make sense? Then you have a LOT more tools in your toolbox for modeling and a lot simpler problem to look at - at least at the outset. Personally I try everything possible to turn time series problems into normal regression / classification problems before resorting to TS tools. It's just a much more restraining environment.

Sign In or Register to comment.