I'm working on a classification model aimed at identifying if behavioral activity within an account (b2b - one account, many contacts) can predict or not an opportunity generation ( a salesperson triggering a negotiation).
I have the following data aggregated per opportunity/period of time:
One record for each opportunity (positive case) generated in the previous 12M period (rolling), with the behavioral activity in the preceding 6 months happened before the opportunity generation date (if any)
One record for each week/account combination (negative case) in the previous 12M period (rolling), with the behavioral activity in the preceding 6 months happened before Sunday of that week. IF an opportunity was generated on that Sunday, then no record is generated for the account.
So typically a record would consist of a specific opportunity id and the behavioral data for that account in the previous 6 months OR the behavioral activity in the 6 months period for an account in each week of the previous 12 months. something like:
Accout Opp week day N_interactions_Content_X_previous6M 123 1 52 Monday 34 123 0 52 Sunday 36 234 0 52 Sunday 56
My question is with regard to the records where an opportunity exists BUT no Interactions are present in the previous 6 months. So positive cases where all the behavioral variable are set to 0 (all the variables) but a positive case generated (maybe because it was a stand-alone effort of the salesperson - which is not tracked in the systems)
The amount of these positive cases is almost 40% the number of positive cases WITH interactions.
I'm questioning if providing a set of records with all variables set to 0 but representing a positive case would contaminate the learning process of the model (the model would learn that when all variable are set to 0 there is a random chance to have positive cases generated)
Should I exclude these data? There is no Sales data in our systems so can't be added to the model. Does anyone know literature about how to manage these "positive empty" records?