Investigating Longitudinal Effects of Online Recommendations with Agent-Based Simulation

Automated recommendations are nowadays a central part of our online user experience. Today, many online services – including e-commerce sites, media streaming platforms, and social networks – use recommendations to achieve organizational goals, such as increasing sales or customer retention. However, the existing literature on the design of recommendation algorithms mostly does not consider this business-oriented perspective but focuses almost exclusively on consumers’ value. The underlying assumption for doing so is that recommendations that optimally satisfy consumers’ information needs are ultimately also the most valuable ones from a business perspective. In reality, however, this assumption is a severe oversimplification of the underlying, much more complex situation, since multiple stakeholders’ goals must be taken into account. Consider, for example, an online hotel booking platform that charges commissions to partnering property owners whenever bookings are made through the site. When recommending hotels to consumers, an underlying machine learning algorithm might, first of all, focus on identifying hotels that match the consumer’s preferences, thereby increasing the probability of booking in the first place. At the same time, however, to maximize the platform provider’s profit, the algorithm might consider each viable option’s profitability when ranking the items recommended to the user. Finally, the algorithm might also want to ensure that all property owners’ items are recommended from time to time, to keep them partnering with the site. Ultimately, the booking platform’s problem is to design a recommendation strategy that balances the potentially conflicting goals of multiple stakeholders and assures the business’s profitability in the long run.
Despite their practical relevance, multi-stakeholder recommendation scenarios and longitudinal effects of recommender systems still represent two major research gaps. In our research, we employ agent-based simulation, a novel approach in this area, to model and analyze the alternative recommendation strategies effects on different stakeholders’ behaviors and objectives over time. Specifically, the agents in our model include consumers, recommendation service providers that determine the strategy, and item providers, i.e., property owners in the above example. Furthermore, we include social reputation mechanisms (e.g., posts on social media sites), which may amplify consumers’ publicly shared quality experiences. Our proposed agent-based model ultimately aims to capture the subtle dependencies between the objectively selected items which are recommended to users, the consumers’ related quality expectations, and the feedback and transactional behavior that might arise from not matching the consumers’ quality expectations over time. Moreover, we model how the actual consumer behavior, which is influenced by the perceived quality of the provided recommendations, may influence the profitability of the service providers and the item providers in the long run. Complex dynamics can emerge from the mutual interdependencies between the stakeholders, which were not adequately analyzed so far.
The first set of simulation experiments based on real-world consumer feedback data confirms the viability of our simulation approach. The experiments show that the proposed model is capable of reflecting real-world dynamics. It turns out that recommending the items with the highest expected value for consumers may not be optimal in terms of business value. At the same time, however, the model realistically encodes that optimizing only for provider profitability may lead to missed business opportunities. Our current and future works aim to better understand the extent to which social reputation mechanisms can amplify and accelerate the observed dynamics.