The Principles

In this article, we present the principles of human-machine collaboration. These principles should be considered when developing data-driven decision-making whereby a human decision maker collaborates with data science algorithms.

It should be noted that the relationship between challenges and design principles is many-to-many. That is, the first challenge of human-machine collaboration is addressed via the human-in-the-loop and collaborative rationality design principles.

Moreover, the challenges of defensive decision-making, accountability and algorithmic aversion can be addressed via the ability to explain and open design principles. The challenge of opacity is in turn addressed via the no intermediaries design principle.

Furthermore, there are certain classes of non-functional design principles that do not map directly to a specific challenge but rather govern the way the platform is to be designed and implemented, such as the data integration and governance design principles.

Let’s take a closer look at these principles.

Data integration and governance design principles

P1 – Human in the loop

While machines can mitigate human error in a wide range of applications, cases from organizations and governmental agencies have recently shown severely negative consequences of eliminating human decision-making entirely. This is especially amplified where the decision-maker is what is known as a knowledge worker (where their main contribution to their job is their knowledge) or when the decision has some critical implications. Thus, data science for decision-making will always maintain the human decision-maker in the loop.

P2 – Collaborative rationality

A substantial part of the decision practices supports the notion of bounded rationality. Bounded rationality is the opposite of unbounded rationality. The concept of bounded rationality was developed by Herbert Simon. The opposite, unbounded, means someone who knows everything, can compute everything, and can remember everything. Which is of course not possible for a human. Many models in the social sciences are classified under bounded rationality. We advocate an approach whereby humans and machines (analytics) collaborate through so called collaborative rationality, making decisions in order to reduce the impact of bias and human-alone bounded rationality.

P3 – No intermediaries

Use of data science without intermediaries is expected to foster accountability. In order to align with the objective of serving a wide range of organizations and decisions, the data science tools need to have a user interface that does not rely on specific technical competencies which are owned often by only a few managers. The tools need to be designed so that they can be later adopted by decision-makers from various domains given the fact that competence levels vary, for example self-service or democratized.

P4 – Explainability

Machine learning and AI algorithms have tremendous success in interpreting knowledge from data to support decision makers’ insights relevant to the decision situation, albeit often with a lack of explainability and interpretability which leaves users unsure as to why a certain recommendation is made. Accordingly, this principle is aimed at ensuring that data science tools should be equipped with the right techniques to make analytics explainable and interpretable to its intended users, since analytic techniques are not only expected to be good, but also explainable.

P5 – Multi-species analytics

Data science should maintain a neutral position towards statistical, machine learning, and other AI techniques in order to serve different decision-making purposes. Towards that end, we aim to adopt a wider definition of analytics techniques which is proposed to offer business intelligence, big data, machine learning and AI enabled via a data science platform. The platform is a type of information system which ought to process the data and the analytics

P6 – Evaluative nature of outcomes

Decisions need to be evaluated, no matter how they have been taken, but many organizations find it hard to evaluate decisions that impair their ability to learn from decision-making mistakes. Therefore, the organizations ought to develop a range of metrics to facilitate evaluating decisions, based on contextual multi-criteria such as efficiency, cost and impact.

P7 – Openness

Openness is defined as the provisioning of technical means for others to access the core functions of the data science tools. The openness of the tools supports the ability of other service providers to connect to them. Open systems are characterized by the fact that their key interfaces are published, and accordingly, the interface to connect to the data science tools should be published. Tools need to open up their business to external users, like suppliers and distributors, to enable external transactions between them.

P8 – Data integration

The data science tools should be able to collect, cleanse, and integrate data of all varieties (structured, semi/unstructured), sources and formats – in order to provide the foundation for further analysis. It is to be noted that in a situation where data is not integrated, or not within reach, users tend to copy and aggregate without any standard procedure for doing so, which will influence the analytics outcomes and affect the decision quality.

P9 – Governance

In the era of big data, governance continues to represent one of the key pillars. Therefore, the tools used need to ensure that roles and decision as well as rights and responsibilities for efficient and compliant use of data and analytics techniques are preserved. This includes data ownership and data stewardship, amongst others.