The prospects for i* modelling to meet the emerging challenges poses by RE for AI are reviewed. In this talk I will outline some ideas about how i* can be applied to the concerns of fairness in AI [1] and how this suggests new roles in data modelling requirements, extensions to softgoals and the concepts of dependency in i*. The implications of RE for AI have been explored in the AIRE workshop series as well as modelling using i* to explore roles with AI development to illuminate how data science, AI and software engineers are all necessary to contribute to the overall requirements picture [2]. System-level modelling with i* has also been explored in software ecosystems [3]. Combinations of models ranging from goals to i* and other system models have been used to explore different viewpoints in RE [4]. In multiple modelling languages, including i*, study of chronic disease management systems has demonstrated how multiple layers of models could map influences from the socio, political and business levels to operational requirements [5]. I will take up the role viewpoints theme in AI and propose how i* might be used to model the emerging problem space of AI applications. First, the case study to anchor my arguments. A call centre in a service engineering domain (see Figure 1a) of telecommunications logs data on customer calls, customer satisfaction, call duration, repeat calls, etc. The company aims are to improve customer satisfaction and call-centre agents' performance. Metrics are calculated on mean call duration, satisfaction ratios, call load/agent, etc. Satisfaction is recorded as a simple 1-5 scale rating at the end of each call. A machine learning classifier has been trained on customer call data sets to identify profiles of dissatisfied customers and underperforming agents, with the objective of improving customer relationship management. The company is concerned with fairness in any treatment of underperforming agents and the completeness of identifying satisfied customers. The application poses two requirements problems: (i) how to identify 'fairness' and (ii) an accuracy/completeness of classification for satisfaction, as well as the AI problem of conformance with ethical standards for fairness to avoid legal challenges from unfair treatment of employees. The problem overview is summarised in Figure1(a), showing the dependencies between agents for the service company with the softgoals of efficiency for the company and customer satisfaction as a shared goal. The approach to this problem develops awareness requirements concepts [6]: how can monitors be developed to evaluate classified output for fairness? This begs the question of defining fairness and operationalising definitions as metrics and measures. Fairness can be evaluated against training set metadata, e.g. ethnicity, gender, age and other demographic variables for call-centre agents.
Awareness requirements (Figure 1b) are implied to check the distribution of underperforming
operators against their demographics to ensure there is no statistically significant bias by age,
gender, ethnicity, etc., using standard tests such as analysis of variance and regression.
Developing a monitoring process to detect digression in classifier output with respect to
attributes distribution in the subject population becomes the responsibility of requirements and
software engineers, while defining tolerance limits and selecting statistical tests necessitates the
expertise of data analysts. Requirements are thus extended into the realm of data science;
furthermore, any digressions which are detected trigger a further class of requirements for
investigation and adaptation, e.g. call duration can be longer for female operators who tend to be
more patient and thorough than males, while older operators have longer response times from
normal ageing effects. i* models were extended from strategic dependencies between agents,
goals and softgoals implicit in the baseline system, to model dependencies between attributes of
agents in input data models, fairness values elaborated from softgoals and the original system
goals specified for the classifier’s requirements. More detailed i* dependency models describing
the training data domain were compared with models describing the classifier output so
interrelated (training-output) models dependencies could be investigated with the addition of
softgoals representing fairness values (i.e. Equality, Social responsibility, Diversity) and other
‘soft issues’ [
While these problems may become apparent during system operation, it is desirable to
discover them at design (or classifier training) time and then implement counter measures either
through awareness requirements and monitors where management can be alerted to take
appropriate action, such as improving classifier training with better selection of training data, use
of supervised learning to mark up relevant bias examples, transfer learning from selected training
sets, and ultimately classifier self-correction by learning bias controls, etc. The implications of
corrective requirements involve communication with another group of experts: ML classifier
developers and operators [
While i* model inspection helped to identify many potential biases and consequent
awareness requirements, a more systematic, reusable process is desirable. Knowledge gathered
during this application was gleaned from data scientists and ML experts, which indicated issues
for developing a ‘lingua franca’ among the contributing communities from AI, data science and
RE, developing insights from [