Configuration systems play an important role in achieving the sustainable development goals (SDGs) defined by the United Nations. As decision support systems, configurators help users to decide which components or features to include in or exclude from a configuration. An important task of configurators is the provision of explanations which help to achieve goals such as increasing configuration understandability, increasing a user's trust, and persuading users/customers to include specific configuration components. Our goal in this paper is to introduce the concept of „sustainability-aware explanations“ which can help to support the sustainable development goals. The type of explanations we propose in this context are somehow orthogonal to typical explanations used in industrial configuration environments. A major objective in this context is to follow a „less-is-more“ principle focusing on diferent aspects of the idea of „responsible configuration“ which refers to configuration techniques explicitly supporting the mentioned sustainability goals. We report the initial results of an evaluation that provide insights on potential impacts of the proposed explanations.
The 17 sustainable development goals (SDGs) defined by
the United Nations (UN) provide a blueprint for peace and
prosperity on our planet.1 Examples of such goals are
good health and well-being (e.g., in terms of fostering the
consumption of healthy food), responsible consumption and
production (e.g., in terms of reduced energy consumption),
and sustainable cities and communities (e.g., in the context
of tourism, avoiding negative environmental impacts and
taking into account the local communities and cultural
heritage) [
Knowledge-based configuration [
In knowledge-based systems, explanations can be applied
for diferent purposes [
In this paper, we focus on the persuasion aspect of
explanations [
Sustainability-aware explanations have to focus on
argumentations including sustainability aspects. Our
formulation of such explanations is based on large
language model (LLM) prompts [
Positive impacts of such sustainability-aware
explanations can be, for example, higher-quality
configuration decisions, a lower amount of unneeded
components, and components with less negative
2The categorizations of [
The contributions of this paper are the following. First, we propose the concept of sustainability-aware explanations for configurations. Second, we provide reference examples of such explanations in the automotive domain. Third, we present initial results of a corresponding evaluation.
The remainder of this paper is organized as follows. In Section 2, we provide diferent examples of LLM-generated sustainability-aware explanations in the car configuration domain. Thereafter, we discuss initial results of a related evaluation (Section 3). In Section 4, we discuss threats to validity. Finally, we conclude the paper with Section 5.
In the following, we discuss scenarios where sustainabilityaware explanations can have an impact on user decisions. All scenarios are related to car configuration where users receive explanations of current configurations. The major goal of such explanations is to make users think about their current configuration settings and to potentially adapt their articulated preferences. Consequently, our explanations are not in the line of why or why not explanations but focus more on indicating potential alternatives to the current configuration, i.e., a kind of why not choose something else explanation. All example explanations in this paper have been generated on the basis of the LLM ChatGPT 3.5.3
persons (configurator users) who intend to purchase an SUV more aware of sustainability aspects of smaller cars. To support this, we have generated LLM-based explanations using the following (example) LLM prompt: Assume the following scenario: person A wants to purchase a car and is interested in an SUV. Please provide persuasive explanations against purchasing an SUV following the six persuasion principles of Cialdini. The resulting explanations are depicted in Table 2. 3https://chat.openai.com Scenario 2: Long vs. standard range battery. The idea is to make configurator users interested in purchasing a car with a long-range battery aware of the sustainability aspects of standard-range batteries. To support such explanations, we have generated LLM-based explanations using the following LLM prompt: person A wants to purchase an electric car and is interested in a long-range battery. Please provide persuasive arguments against purchasing a long range battery following the six persuasion principles of Cialdini. The corresponding LLM-generated sustainability-aware explanations are depicted in Table 3.
Configurator users should think about the advantages of not having a car when living in the city center. We have generated related LLM-based persuasive explanations using the following LLM prompt: person A who lives directly in the city center with various connections to public transportation wants to purchase a car. Please provide persuasive arguments against purchasing a car following the six persuasion principles of Cialdini. The corresponding sustainability-aware explanations are depicted in Table 4.
The idea is making configurator users with limited financial resources intending to purchase an expensive car to change their mind and purchase a less expensive car. To support such explanations, we have generated LLMbased explanations using the following (example) LLM prompt formulation: person A with very limited financial resources and a family with three children wants to purchase an expensive car. Please provide persuasive arguments against purchasing an expensive car following the six persuasion principles of Cialdini. The related LLM-generated explanations are depicted in Table 5.
Properties of LLM-based explanations. In Table 6, we summarize the diferent argumentation lines generated by the large language model (LLM). (1) In the context of the persuasion dimension reciprocity, LLM-generated explanations refer to the aspect of „giving something back to the community“, for example, purchasing an eco-friendly vehicle can be a way of giving back to the environment. (2) Explanations related to the persuasion dimension scarcity on the one hand refer to decreasing incentives for sustainable equipment (e.g., cars), on the other hand to limited resources (e.g., financial resources). (3) In the context of the persuasion dimension authority, diferent experts such as environmental experts are used as representatives of authorities. (4) Explanations related to the persuasion dimension commitment assume an existing commitment of the current user, for example, already proved in previous configuration sessions. (5) In the context of the persuasion dimension liking, LLM-generated explanations refer to a user’s family, friends, and neighbors (e.g., your family will like your decision). (6) Explanations related to the persuasion dimension social proof refer to trends of peers, the wider community, and similar families.
which of these explanations would have the highest persuasive
impact on a user? Please provide a ranking., we asked the
LLM also for a ranking of the impact of the generated
explanations following the idea of LLM self-evaluation
[
Feedback of study participants. For a very first evaluation beyond the LLM feedback, we conducted a small user study in our research group with N=10 participants (computer science students at our university). The participants were asked to rank the diferent explanations according to their potential persuasive impact. The results are included in Table 7. Overall, there exists an overlap between the explanations ranked highest by the LLM and those selected by the study participants. In this context, the authority principle has been regarded as relevant in all example configuration scenarios. This initial result indicates basic LLM capabilities to recommend persuasion strategies.
The explanation concepts presented in this paper are based on one selected configuration scenario (car configuration). These explanations are just high-level examples and many further (also more detailed ones) can be envisioned for car configuration (e.g., the sustainability aspects of a less powerful car engine) and beyond. In our work, we did not focus on a specific phase of a configuration process, i.e., the mentioned explanations could even be used before the configuration process has been started (e.g., as explanations in wish lists or product information). Gaining more related insights is a major focus of our future research.
The presented impact ranking of explanations has been
primarily discussed on the basis of an LLM-generated
ranking [
In this paper, we have introduced the concept of sustainability-aware explanations of configurations. Using the example of car configuration, we have explained and exemplified this type of explanation. Following a set of persuasion dimensions, we have analyzed the LLM-generated explanations with regard to the used argumentation lines and analyzed the impact of the generated explanations on the user. In this context, LLMs show to be applicable in terms of generating explanations in a flexible fashion but also to recommend explanations in specific configuration contexts. Our future work will include detailed studies with real users with the goal to compare LLM-based rankings with the perception of explanations by real users. Further research will include an analysis of the efects of combining explanations (e.g., integrating authority-based with commitment-based explanations).