Client background
Our client is a North American data service provider. It collected large volumes of data from many sources but saw limited ways to monetize it.
Business challenge
The data was unstructured. Without extensive additional processing, it had no usable dimensions for analytics. The client wanted to start with the transaction entity, which could show the dynamics behind a benefactor’s payment. Even a simple rules-based system couldn’t analyze it effectively. Some entities had very similar names, which made processing impossible.
The client asked SoftServe to find a way to extract monetizable information from its data.
Activities and solution
Our experts worked with the client’s team and concluded that a deep learning model was required. The model would use active learning principles to build the training dataset interactively and faster.
To solve the classification problem, we used entity names as the target variable and transaction names as the model input. The architecture combines embedding and bidirectional LSTM layers with a DistilBERT transformer tokenizer. With around 2M trainable parameters, the model stays under 25MB.
Value delivered
The model gave the client a structure for analyzing its data. It performs multilabel classification for about 1,500 entities and is built on a final training dataset of around 30,000 rows. It reached nearly 95% entity recognition accuracy on entirely new data. Running on Amazon Web Services (AWS), the solution processes more than 3M transactions per day.
Structuring and monetizing large data sets takes hands-on experience building deep learning solutions. That experience let the client put its data to use and pursue new opportunities.
Client feedback
One of the client’s co-founders said they were “incredibly impressed with the SoftServe experience.” They “loved that SoftServe came with prior conviction on what would be needed to complete the project to help it get off the ground in the early days.” They also “appreciated the willingness to transition to high priority items as the project progressed, even if they deviated from the original gameplan.” The co-founder “found the collaboration and teamwork elements to be wonderful” and looks forward to working with SoftServe again.
