Every day employees were sorting and deleting non-relevant emails which took quite a lot of their labor time. The customer needed to reduce the time to filter out spam and automate some other routine operations.
Key Challenges
- Reducing spam in the email client
- Reducing employee email processing time using an email classification bot
- Automating routine tasks
Implemented our solution
Since the standard filters of the email service failed to perform email classification and employees had to spend too much time sorting emails themselves, ICL Services specialists offered the use of machine learning-based solutions in order to reduce the amount of routine and non-relevant email filtering operations.
Prior to the start of the work, a preliminary study was carried out. The ICL Services team documented then current problems, collected input data and analyzed the customer's infrastructure. The audit helped sel ect the optimal architectural solution.
The customer's entire email infrastructure was hosted on MS Azure and, using a native environment, it was possible to deploy the solution completely seamlessly for the customer in a few days.
We defined the parameters for email classification and built a model
Machine learning models must be trained using historical data. For that reason, the initial task was to collect and mark up the available correspondence history. And the task was simple in a way because the HR department sorted useful emails (CVs and HR-relevant information) fr om spam anyway. The problem was something else–the amount of emails available for training was limited literally to a one week period.
The project started off with data collection. The data was converted to the required format, marked up and fed for training to the ML model training engine built into MS Azure. Unfortunately, the accuracy of the first ML model was about 70% due to the lack of data.
We improved accuracy
It was obvious that more data was required to improve accuracy. The ideal scenario would have been to ask the customer to accumulate more spam in the marked up data and then use it to finish the ML model but it required more time. Thus, an alternative option was selected. In just a few days, the developers collected spam data sets fr om public sources, translated them into German and uploaded them to the ML model. This made it possible to increase the filtering accuracy from 70% to 92% and complete the initial tasks: filtering out what the model considered to be certain spam, automating auto-replies when the ML model was certain of what type of email had been received and leaving only a minor "grey area" wh ere the error probability was high for manual processing. What is more, thanks to data accumulation and recurrent retraining of the ML model, the filtering accuracy has been growing continuously and is already 94% at the moment.
Results
- ICL Services developed and implemented a machine learning model for spam filtering;
- The customer obtained an out-of-the-box automated email processing solution integrated with Microsoft Azure in only a few days;
- With 92% spam detection accuracy and auto-replies, HR employees reduced the time they spent manually handling emails;
- The company mitigated reputational risks–job seekers no longer needed to wait several days for a response but received it instantly.