3 Replies Latest reply on Feb 7, 2017 9:12 AM by aminec RSS
    ilopez Explorer

    TWA_predictive_scoring_tips&guide

    Hi everyone,

     

    I am currently using the TW Analytics module, the trial of 30 days, and I am facing some doubts relating to the interpretation of predictive scoring analysis.

     

    1) I have loaded my dataset and I have created my model. Moreover, I am able to download the predictive scoring .csv file from the platform but, actually, I would need some related guide to understand if the analysis is consistent with those results.

     

    2) Since the accuracy of my model is really close to 100%, is there an explanation guide or some reference tips that could get my model more realistic?

     

     

    Really thankful if you can help me.

     

    Kindly,

    Ivan

      • Re: TWA_predictive_scoring_tips&guide
        jgreiner Apprentice

        Hi Ivan,

         

        1. Here is a link to the API guide which goes over all the jobs that are possible to run on ThingWorx Analytics.  For descriptions on the Prediction Models I would Check out the Prediction Mode Generation Section and Scoring sections of the Document.
        2. ThingWorx Analytics self-validates its predictions by putting aside a relevant set of data (holdout set) to verify against. Prediction accuracy is qualified on several performance metrics, primarily RMSE and Pearson Correlation. So the model accuracy was so high because that is how the hold out set performed when it was scored against the generated model.

         

        Let me know if you have any further questions.

         

        Thanks,

         

        John

          • Re: TWA_predictive_scoring_tips&guide
            ilopez Explorer

            Hi John,


            thanks for your kind reply, but I am not able to open those links due to some unauthorized access.

             

            Can you just post here the related content please?

             

            Thank you again.

            Ivan

              • Re: TWA_predictive_scoring_tips&guide
                aminec Apprentice

                Hi Ivan,

                 

                Now as far understanding the outcome is concerned in analytics is on a case by case basis so no general interpretations could be provided for specific results. However, a broad definition of the concepts you mentioned would be :

                 

                -The Goal: A “Goal” in ThingWorx Analytics is the same as an outcome variable or “dependent variable”.

                -The features or Independent Variables : is the information available to help us predict or explain an outcome whereas the dependent variable (Goal) is the outcome we are trying to predict or understand.

                -The SignalsA “Signal” represents the strength or weakness a particular input (independent variable) has in relation to the Topic (dependent variable)

                -Model: This is a mathematical Equation representing the relationship between the Goal and the independent Variables created through ThingWorx Analytics Machine learning algorithms. Its accuracy is measured through different methods like RMSE or Pearson correlations. The Model is trained based on the historical Data available in the Dataset you provide

                 

                For the interpretation of the results, you would look at the signals first. These would provide a first view on the descriptive analytics side towards the goal you are trying to predict. ThingWorx Analytics would not create those signals but rather reveal them through trends in your Dataset.

                 

                And for your model I would suggest to remove the feature with the strongest signal from the Dataset (through the use of an exclusive filter). After that try training your model again. If you are obtaining a "too good to be true" model this feature might have biased correlation to the goal.(although it is not necessarily the case ).

                 

                I hope this helps.

                 

                Best Regards,

                Amine