I'm working in weka with a data set which has count as the last attribute. The count attribute which is a numeric data type refers to the number of bikes that are rented on that particular day. Eg. 12, 15, 30, 11,45. I wanted to predict these values in my test data set.
0
votes
Weka typically assumes that the last attribute is the target attribute. Classification algorithms require the target to be nominal, but prediction algorithms (such as linear regression) allow for a numeric class. You can also change the target attribute in the preprocess tab of the Weka Explorer.
– Walter
2 Answers
1
votes
You can use linear regression with a numeric target attribute (Weka calls this the "class attribute"). In fact all of Weka's algorithms under the "classify" tab that are unavailable (because of the numeric class attribute) will be greyed out.
If you have lots of discrete independent variables, e.g. "weather", "is_weekend", "payment mode", "customer_gender",... you can also use Weka's "Discretize" filter to convert the numeric class attribute into a discrete one, e.g. interval classes such 0-10 bikes/day, 10-20 b./day etc.
In order to do so you must temporarily assign the class attibute to any another attribute, Do this with the selectbox at the top of the bottom right of the filter tab. then apply the discretize filter, but ONLY on the class bikes_per_day attribute (by default the filter dicretizes all attributes). Then re-assign the class attribute to the last, now discretized, attribute.
Then Weka's classification algorithms such as J48 will be available (no longer greyed out). They might give more concise, interpretable results, e.g. little decision trees such as: If season = summer and day= weekend and weather_forecasted=sunny then predicted_rented_class=40-50 bikes/day
