Creating a Custom Classification Model for Leading Media Intelligence Firm in AWS.

Auxo Digital designed and built an intelligent custom classification model capable of providing a market-leading media intelligence customer with efficient data processing, sorting and validation based on specific parameters.

The model, developed using a machine learning operations (MLOps) approach, leverages AWS tools and artificial intelligence (AI) capabilities to automatically detect sentiment for potential incitement across numerous online sources.

Combining the deep engineering skills of a multi-talented team with the evolving capabilities of the AWS architecture, Auxo Digital developed the model to meet the very specific requirements of the customer. The technology, capable of scaling on demand and handling increasing levels of complexity, was designed to flex and evolve according to the needs of the client.

Intelligent learning, capable engineering

The classification model was used by the client to undertake automated sentiment detection, particularly around political incitement, which meant it had to be trained to match globally accepted definitions of the legal concept.

The specificity of this requirement meant the content and data collected by the model had to be accurately analysed. However, due to how easily the data could be misinterpreted or misread, the model had to be capable of finding the best results despite the volume of data.

It was a challenging ask and one that meant following a clearly defined MLOps approach:

Technology Capabilities

MLOps

Digital Innovation

Artificial Intelligence

AWS

01

Data Collection

Data collection was an essential first step. Each data point had to be collected and manually classified, to form the foundation from which the model ‘learned’ its requirements during the training phase. The data had to include a good variety and volume of examples, while data augmentation was used to scale the training data.

02

Data Pre-processing

The data was cleaned and pre-processed in preparation for training.

03

Training

The model was trained using the data and tried to learn the correlation between the input (content) and the output (classification).

04

Performance Evaluation

Once trained, the model’s performance was evaluated. If it was optimal, it could be used to analyse bulk data or be deployed to an endpoint for real-time analysis.

05

Iterative Improvement

The model was improved by using insights from the performance evaluation to enhance the training data and tweak model training parameters. The MLOps cycle was then iterated a few times before the model reentered production.

The CCM’s use case for the customer was developed and built using Auxo Digital’s proprietary 5 step software development methodology, augmented with the process detailed above.

Once the model had been trained and was ready for deployment, it was then continuously monitored to ensure its performance remained relevant and capable. Any shifts in performance resulted in the model returning to the training phase.

As the model was improved, trained and retrained, it became increasingly adept at identifying nuances within the data and providing relevant feedback and insights.

Business Value:

  • Customised, trained model capable of collecting data according to customer specifications

  • Built-in validation parameters

  • Intelligent data classification

The Auxo Digital Difference

The classification model was immensely successful for the customer, providing exceptional data and insights and meeting the very clear brief.

The technology as a whole has varied applications – Auxo Digital’s approach to MLOps allows for the creation of classification models that can fit within any business, offering a unique and powerful way of translating data into value.

The models are capable of identifying conversational fragments across myriad data streams including social media, and deliver return on investment across highly relevant and critical insights that can be used to customise communications and service delivery.