Cognitive computing: Building blocks
Expediting IT services
In a study published by IBM in January 2021, 60% of CEOs and CTOs in mid-size and large companies said their company’s IT modernization program was not prepared for the future. Digital transformation is difficult at best, but can be impeded further if existing IT resources are not used effectively. Moreover, in an era of aggressive hacking, organizations will want to be aware as soon as possible if anomalies in IT performance are detected. One tool that helps improve IT performance is cognitive service management (CSM). Supported by digital automation, AI, and machine learning, CSM helps reduce the downtime of these more complex environments and streamline the digital transformation process.
BMC Software has been assisting companies in leveraging their IT investments for many years to help ensure that IT supports their business activities. The BMC Helix platform, which is a SaaS application, manages service processes, change management, and knowledge assets that the IT department draws upon to support a company’s business activities. “Our set of cognitive capabilities helps produce this outcome,” said Peter Adams, senior director of product management at BMC. BMC Helix operates primarily within the IT department but also has several other user groups such as HR and shared services.
A typical resource delivered by the BMC Helix platform is the development of chatbots that enable responses to IT service issues. “We look at how companies can use the investments they have made, such as existing knowledgebases,” continued Adams. “For example, many companies have a predefined catalog of standard requests, such as filling out a form to request a new computer, or increasing email storage. In this type of narrow domain, the chatbot can easily support that request, leverage the existing service catalog, and convert it into a conversation to trigger the same fulfillment.”
For its natural language processing and other underlying technologies, BMC Helix relies on leading thirdparty technology. “We focus on setting up the chatbot, using prebuilt content, and automating the training,” Adams explained. “Once the chatbot understands the intent of a request, it asks the right questions so it can fire off the transaction for fulfillment. Unlike a service catalog, the chatbot will know about the user’s past transactions and won’t ask again for information that is already in the system.” The BMC Helix platform also offers translations, cognitive searches of external repositories, and sentiment analysis.
Adams concurs with the advice that many others offer when it comes to chatbots—start with the most frequently requested service that is in a narrow domain, and get broader from there. “One higher education institution wanted to develop more self-service options because so many in the younger generation accept or even prefer being able to interact on their mobile device 24/7,” he stated. “They began with a chatbot on how to connect to Wi-Fi in their dorm room, and then expanded from that task to others and to other groups.”
Model validation and explainability
Machine learning is a part of many AI applications. The learning process is iterative; first, part of a dataset is input and analyzed, and a model is developed based on patterns and correlations among the factors in the analyses. The model is then tested on the remaining part of the dataset to see if it produces comparable results. If it does, the application can then go into production, with the assumption that it will be a reliable predictor of outcomes.
In the case of a decision about obtaining a loan, for example, the measured factors might be income, debt-to-income ratio, and a FICO score. Each factor would be weighted and a score calculated. The bank would set a threshold above which the loan would be approved and below which it would be denied. “This type of model is straightforward and explainable,” said Anupam Datta, co-founder, president, and chief scientist of Truera, whose Truera Model Intelligence Platform assists companies in developing models, including validation and monitoring, to help ensure the models are accurate and effective.
Ongoing monitoring is important because even with a simple model, complexities emerge. In the historical data used in the training and testing dataset for loan applications, the outcome metric of greatest concern— default on the loan—would require a sample over time, because the default would not happen right away. Therefore, going forward into production, validity of the model could not be ensured until some time had passed.
In addition, the selected training and testing sample may have sample selection bias. “The Truera platform can identify subsets of data where there is sample selection bias and surface them for data scientists to inspect,” said Datta. The passage of time brings changes that affect model accuracy. For example, the advent of COVID-19 brought uncertainties to financial situations that were previously stable. The model needs to institutions to provide explanations to individuals denied loans,” Datta noted. “The goal is to prevent discrimination and bias both at the individual level and for classes of individuals.” This requirement becomes more challenging as model complexity increases.