Towards welfare-oriented recommendations in activity-travel behavior

Every day, people make countless choices about where to go and what to do. From selecting a restaurant for dinner to deciding which park to visit, these decisions are rarely made in a vacuum. Increasingly, they are guided by digital assistants and recommendation systems that suggest options based on what others have liked or what is currently popular. These tools are designed to make life easier, but they operate on a fundamental assumption: that if an option is popular or highly rated, it is a good choice for the individual user. However, this assumption often overlooks a critical, hidden factor in daily life: the cost of getting there. In the world of travel and activity planning, time spent in traffic, money spent on fuel, and the effort required to reach a destination are real expenses that cannot be refunded if the experience turns out to be disappointing. A highly rated restaurant across town might seem like a great idea on a screen, but if the journey takes an hour in heavy traffic and the meal is merely average, the user has lost more than they gained. This disconnect between what algorithms recommend and what actually benefits a person is the central puzzle researchers are now trying to solve.

A team of researchers has developed a new way of thinking about these digital suggestions, shifting the focus from simple popularity to a concept they call "user welfare." Instead of asking which option is the most clicked or the most reviewed, their approach asks a more personal question: will following this suggestion actually leave the user better off than if they had chosen something else on their own? To explore this, the researchers built a complex computer simulation, a virtual city populated by 250 digital travelers. These travelers were not simple robots; they were given distinct personalities, different levels of patience, varying amounts of money, and unique ways of making decisions. Some were impulsive, some were cautious, and some were prone to sticking with their old habits. The simulation allowed these digital people to interact with different types of recommendation systems over the course of sixty simulated days, creating a realistic environment where traffic jams, changing preferences, and the consequences of bad choices could be observed and measured.

The researchers tested five main approaches in their virtual city. The first was a baseline condition where agents selected leisure segments and locations entirely from their own preferences without any algorithmic intervention. The second was a standard system, similar to what exists today, which ranks places based on popularity and general relevance without worrying about the specific cost to the individual. The third and fourth approaches were the new, welfare-oriented systems. One of these used a "positive utility" filter, which acted as a safety check. Before suggesting a place, the system calculated the likelihood that the enjoyment of the activity would outweigh the travel costs. If the math suggested the trip might be a net loss, the system would simply stay silent, letting the user make their own choice. The other new approach used a "regret minimization" filter. This system compared the recommended option against the best possible alternative the user could have found on their own. If the recommendation was likely to leave the user feeling worse off than they would have been with a different choice, the system would withhold the suggestion. The fifth approach was a hypothetical "Oracle" system with perfect knowledge of every user's preferences, serving as an upper bound to measure how much room for improvement exists.

The results of the simulation revealed a clear advantage for the welfare-oriented approach. In the virtual city, the standard system often led users into situations where the travel costs ate up the enjoyment of the activity. While the standard system did improve the average experience slightly compared to having no system at all, it still resulted in a significant number of "bad trips" where users were worse off than if they had never received a suggestion. In contrast, the systems that filtered for welfare dramatically reduced the number of these negative experiences. By refusing to make suggestions when the odds were poor, these new systems protected users from wasting time and money. The simulation showed that when the system held back, the users who did receive recommendations were much happier with their choices. The average satisfaction of the trips increased, and the inequality in how well different users were treated decreased. The system did not just recommend more popular places; it recommended the right places for the right people at the right time.

Perhaps the most surprising finding was how much harm the standard system was causing. The simulation showed that nearly forty percent of the time, a standard recommendation actually made a user's day worse than if they had just chosen a place on their own. This happens because algorithms often prioritize engagement—getting a user to click or visit—over the actual outcome for that user. The new welfare-oriented filters cut this rate of harm significantly, reducing the number of users who ended up worse off to around thirty-one percent. Furthermore, the study found that these filters were particularly helpful for certain types of travelers. People who tend to be more risk-averse or who are easily overwhelmed by too many choices benefited the most from the system's restraint. For these individuals, the algorithm acted less like a salesperson pushing a product and more like a thoughtful friend who knows when to stay quiet.

The researchers also discovered that there is a delicate balance to be struck. If the system becomes too strict and refuses to make suggestions too often, it stops being helpful. The simulation identified specific thresholds where the system was most effective: strict enough to filter out the bad options, but loose enough to still offer useful guidance. When tuned correctly, these welfare-oriented systems could close a large portion of the gap between current technology and a perfect, ideal system that knows exactly what every user wants. While the study was conducted in a simulated environment rather than the real world, the logic holds up under scrutiny. The findings suggest that the next generation of travel and activity apps should not just be smarter at predicting what users will click, but more principled about when to speak up and when to step back. By treating the user's time and energy as valuable resources that cannot be recovered, these systems could transform from tools of mere convenience into genuine partners in daily decision-making.