The 211 helpline refers individuals to various social services, such as shelters and food pantries all across the Greater Houston Area. Callers are referred to partner agencies based on their need and location after a brief interview by the 211 agent handling the call.
Currently, agents at 211 have to rely on their domain knowledge and manual lookup processes to find the appropriate site to refer the caller. This creates inefficiencies for the agency team and leads to many 211 calls taking longer than necessary. Since the volume of calls that can be supported is limited by the number of agents and how long each call takes, lengthy calls can have a large impact on the effectiveness of the 211 service.
Using historical data of over 350,000 calls from January 2018 to September 2019, we trained a deep neural network to predict where an agent would refer the caller to based on information like call date & time, caller demographics, and the needs requested by the caller.
Our model was able to accurately identify referral sites in the top 10 predicted for 87% of callers. In other words, 87% of the time agents at 211 would be able to find the appropriate referral site without any manual processes.
By using machine learning to generate predictions, we would be able to dramatically increase the efficiency that 211 agents are able to make referrals and thus the overall effectiveness of the service. If we can speed up the time of each 211 call, we can make sure all callers are directed to the resources they need.
Future applications for this model could involve developing software to augment an agent's user-experience to incorporate the predicted referrals or adapting it into a standalone web service to serve a public audience.