How to build a predictive model with multiple features?

Question

I built an R RandomForest Regression model. The source training data is a historical monthly report of all closed tickets, and the data for forecasting/prediction is a report of open tickets. These reports are generated by another team.

I test/train the model using two years of historical closed ticket data, and predict (forecast) a ŷ Completion Date for each open ticket.

The closed tickets training data looks like this:

ID	Dollars	Fruit	Etc	StartDate	CompletionDate
AA088	500	Apple	...	1/1/2020	2/15/2020
AB100	1000	Apple	...	1/1/2020	5/15/2020
AB101	2000	Banana	...	1/1/2020	5/15/2020
BB723	5000	Apple	...	1/5/2020	3/20/2020
BB724	3000	Lime	...	1/5/2020	3/20/2020
BB725	1000	Orange	...	1/5/2020	3/20/2020

The open ticket data looks similar, except it lacks CompletionDate, and sometimes various fields are "Unknown" at this time.

To build the model, I withhold "ID", make all categorical values factors, use CompletionDate as my y variable, and train the RandomForest on a majority of available features.

Recently, the team that generates this data threw a curve ball, rather than each row being a single record, rows are line-items of a higher level ticket! A majority of tickets have only one line-item, the remaining tickets can have between 2 and 6 line-items.

ID_Parent	ID_Row	Dollars	Fruit	Etc	StartDate	CompletionDate
AA	088	500	Apple	...	1/1/2020	2/15/2020
AB	100	1000	Apple	...	1/1/2020	5/15/2020
AB	101	2000	Banana	...	1/1/2020	5/15/2020
BB	723	5000	Apple	...	1/5/2020	3/20/2020
BB	724	3000	Lime	...	1/5/2020	3/20/2020
BB	725	1000	Orange	...	1/5/2020	3/20/2020

I have considered to summarize (rollup) records, which is easy for numeric value like Dollar (Sum(Dollars)). I could concatenate the multiple categorical values, however, each factor is independent and has strong predictive value to the model (i.e. line items with "Apple" has a weight / meaning that would be lost if I simply concatenated as a string with the other row's value)

ID_Parent	SumDollars	ConcatenatedFruit	Etc	StartDate	CompletionDate
AA	500	Apple	...	1/1/2020	2/15/2020
AB	3000	Apple, Banana	...	1/1/2020	5/15/2020
BB	9000	Apple, Lime, Orange	...	1/5/2020	3/20/2020

How should I handle a categorical feature like Fruit that contains multiple factors?

Can RandomForest accept a feature that contains multiple factors? Do I need to use a different type of model?

Answer 1

My recommendations is to OneHotEncode this variable, to finally obtain something like this:

ID_Parent	SumDollars	ConcatenatedFruit_Apple	ConcatenatedFruit_Banana	ConcatenatedFruit_Lime	ConcatenatedFruit_Orange	Etc	StartDate	CompletionDate
AA	500	1	0	0	0	...	1/1/2020	2/15/2020
AB	3000	1	1	0	0	...	1/1/2020	5/15/2020
AB	9000	1	0	1	1	...	5/1/2020	3/20/2020

Moreover, if you OneHotEncoded this way, random forest can deal perfectly with this categorical feature.

Here I provide you one code that will do what I commented:

import pandas as pd
df = pd.DataFrame({'id': [0, 1, 2], 'class': ['2 3', '1 3', '3 5']})
df['class'] = df['class'].apply(lambda x: x.split(' '))
df_long = df.explode('class')
df_one_hot_encoded = pd.concat([df, pd.get_dummies(df_long['class'],prefix='class', prefix_sep='_')], axis=1)
df_one_hot_encoded_compact = df_one_hot_encoded.groupby('id').max().reset_index()

I've extracted it from here (answered by OmaymaS)