Program

Bringing LLMs into the RecSys has sparked interest, partly due to beliefs that these models can simplify the creation of personalized recommendations across various contexts. However, a critical question has been overlooked: how do LLM-driven recommendation processes affect “non-traditional” users—those who may not fit the typical “profiles” for which LLMs and recommenders were originally designed? In this brief talk, we will explore the potential (or lack thereof) of LLMs for RecSys when non-traditional users are (directly or indirectly) at the receiving end of the suggestions. By placing the human at the center, we will examine the complexities surrounding AI and LLMs—from the benefits they could bring to the recommendation ecosystem to the need to carefully consider the risks and unknowns these models pose for diverse user groups.

In the pre-LLM era, various cognitive models have been used to improve recommender systems. For example, emotions have been used as context and implicit feedback, personality has been used as additional user characteristics. Although this research showed potential, the main downside that prevented the uptake in production was the difficulty of obtaining user data along with ethical issues. In this discussion, I would like to explore the opportunities that LLMs can bring for applying cognitive models in recommender system. In particular, mapping digital traces onto user cognitive models (need for cognition, need for affect, values) and their usage for improving recommendations.

In our presentation at the “Collaborating with AI” workshop, we will introduce the concept of integrating local knowledge with large language models (LLMs) to enhance recommender systems. This will be done by providing a brief overview about the concept itself and by showing current developments in this field. While LLMs like GPT and LLaMA have significantly advanced the field of natural language processing, their application in recommendation systems is not without challenges, especially when it comes to leveraging context-specific, local data alongside the vast global knowledge embedded in these models. Our talk will serve as an introduction to this emerging area, raising several key questions that we believe are crucial for advancing the field:

• How can LLMs be effectively combined with local knowledge to improve the accuracy and relevance of recommendations?
• What ethical considerations and potential biases should be taken into account when using LLMs in conjunction with local data?
• What strategies can be employed to mitigate the risks of hallucination and ensure that recommendations are factually accurate?
• How can users and items be represented in such systems?
• In what ways can LLMs help address the cold start problem in recommendation systems, and what limitations might arise?
• How can we harness the dynamic capabilities of LLMs to refine recommendations based on real-time user feedback?

We will discuss important tasks for group collaboration with AI and why they are challenging. We will mention where we are now and what is the vision for future systems, including functionality and technologies.

Item creators / providers are essential to recommender systems: without items to recommend, such systems have no purpose. But providers are typically an afterthought in the design of recommender systems. When providers are considered at all, it is often in an adversarial way, seeking to prevent providers from misrepresenting their content or manipulating the recommendation algorithm. The result is that providers often feel alienated from the systems that they depend on to distribute their content and to make a living. An alternative perspective is to consider the value that providers can bring as stakeholders in a recommender system. They understand their content better than an algorithm might; they understand their audiences / customers and who they are trying to reach. Multistakeholder recommendation provides a framework in which we can envision new roles for providers as collaborators in building and operating recommender systems.

Recommender systems serve an intended purpose that is not always desirable. In this talk, we discuss our view on principles for responsible recommender systems as a starting point for discussion. These are multistakeholder fairness, user control and transparency, diversity and filter bubbles, and privacy. We briefly present preliminary examples for user interaction for, e.g., sustainable tourism recommendations.

On most platforms, recommendations are continuously and repeatedly served to consumers. It is thus important to measure in a longitudinal way if the recommendations are leading to the desired long-term effects, both in terms of customer-related quality aspects as well as in terms of business metrics. In academic settings, longitudinal effects are however barely examined. Most of the published research is based on one-time measurements, and this applies both to offline evaluation settings and to user studies. To overcome current limitations, researchers need to explore alternative methodological approaches. In this talk, we will outline the importance and existing challenges of longitudinal recommender research and review the potential value of simulation approaches in this context.

This presentation discusses the evolution and evaluation of recommendation systems, focusing on the integration of generative models such as large language models (LLMs). It highlights how recommendation systems are shifting from simple rankers and retrieval models to more complex, generative systems capable of offering curated experiences and nuanced recommendations. However, as these systems evolve, concerns about fairness and trust have become increasingly important. Traditional metrics and frameworks fall short in capturing the complexities introduced by LLMs, which can inherit biases from vast datasets.  The presentation addresses issues related to recent sensitive attribute-based recommendations and advocates the need for “normative” frameworks to evaluate fairness in these systems, ensuring that fairness is assessed based on principled standards rather than arbitrary assumptions.