Emotional dynamics and engagement cycles in swiping dating apps: An agent-based modeling approach

Nobody likes rejection. But if you're actively using dating apps... you better. Because rejection doesn't happen just once, it can happen dozens or even hundreds of times a day. Every unrequited right-swipe, every match that never replies, every conversation that ends abruptly, every first date that ghosts you afterwards, each one is logged in your brain as a tiny emotional loss. Those losses sting, and over time they accumulate into patterns of emotional decline and recovery. If you're one of the 50% of young-adults that have tried a dating app in the past, you've almost certainly experienced this. Here's how it typically plays out:

• You download Tinder or Bumble or Hinge.
• You're excited at first, you swipe through a bunch of profiles, maybe you get a few matches, maybe you don't.
• Then after a few days or weeks, something shifts. The excitement fades. You start feeling frustrated, maybe a little defeated.
• You delete the app.
• Then a month later you re-download it, telling yourself you'll try again.

Sound familiar?

Well, today's paper is looking deeply at this phenomenon. That is: the "emotional dynamics" and "engagement cycles" of dating apps. Before we dive into the authors' methodology and findings, let's establish what we're actually talking about here. When I say "emotional dynamics," I'm referring to how your feelings about an app change over time based on your experiences. When I say "engagement cycles," I mean the pattern I just covered: using the app intensively, then backing off or deleting it, then returning later. The researchers wanted to understand a key question:

What's driving these cycles? And do they affect different types of users differently?

Let's dive in.

The core idea here is that 'swiping apps' expose users to high volumes of what the authors call "subtle/implicit rejection". Unlike face-to-face dating scenarios where rejection might be rarer and more explicit, swiping apps deliver rejection constantly but ambiguously. Previous research has shown that rejection is psychologically distressing (as you'd expect). But swiping apps turn up the velocity. You get rejected, then immediately have to evaluate someone else, then potentially face more rejection, all within a few seconds.

So how do you model something like this computationally? There are a few ways, but the authors settled on a giant simulation, using an agent-based model, constructed using control theory. They created a series of virtual agents (think of them as characters), each with several characteristics: a gender, a score representing perceived-attractiveness (drawn from a normal distribution), and a relationship-seeking strategy that's either casual, serious, or hybrid. Each of these strategies translate into a different mathematical approach for signaling interest in other agents.

• Agents following a casual strategy use what's called the "good genes hypothesis" from evolutionary psychology. They prioritize absolute attractiveness, essentially trying to match with the most attractive available partners regardless of compatibility. This approach calculates the probability of swiping right based primarily on the target's attractiveness level.
• Agents with a serious strategy follow the "matching hypothesis," which suggests people prefer partners whose attractiveness is similar to their own. These agents calculate swiping probability based on how closely the target's attractiveness matches their own, favoring compatibility over maximum attractiveness.
• The hybrid strategy is more sophisticated. It combines both approaches by adapting the casual strategy based on the agent's own attractiveness level. Agents with higher self-perceived attractiveness become more selective when using the hybrid approach, while those with lower attractiveness become less selective. This creates a dynamic that (it's argued) reflects real-world behavior, where people adjust their standards based on their own position.

The heart of their model is the "emotional regulation system". Each agent has goals and expectations about how many matches they should get based on their perceived attractiveness and past performance. Their goal at any given time is calculated as the average between the agent's quality score and their previous environmental feedback. This means expectations adapt somewhat to experience, but slowly.

Environmental feedback represents the benefit derived from platform exposure. It's calculated as the ratio of appreciative signals received to the sum of appreciative signals received and sent. In other words: if you're getting more likes than you're giving out, your feedback score is high. If you're giving out more likes than you're receiving, it's low. When reality doesn't match expectations, a discrepancy emerges. This discrepancy drives emotional updates. The progress rate captures how quickly the agent is closing the gap between expectations and reality using a mathematical function that keeps values bounded between negative one and positive one.

Agents 'leaving' and 'rejoining' the platform is modeled as switching between "inactive" and "active" states, respectively. The transitions between active and inactive states are governed probabilistically by emotional state and output level. For inactive agents, the probability of reactivation depends on their current emotional state. For active agents, the probability of disengagement depends on both low emotional state and high activity rate, meaning frustrated agents with high swiping activity are most likely to quit.

When an agent is active, its emotional state is updated as a weighted average of current progress and previous emotional state. There's also a "forgetting" factor that controls how much weight is given to past emotional states versus current progress. This was calibrated through systematic testing to find values that produced realistic behavior patterns. Output is inversely related to emotional state, meaning agents swipe more when their emotional state is lower and less when it's higher. The researchers also built in gender differences based on data from real dating apps. Specifically, female agents are modeled as significantly more selective than male agents, with women about three times more selective in their swiping behavior.

For inactive agents, emotional recovery follows a different process. The simulation assumes a gradual automatic return to a positive state during periods of disengagement. This is designed to reflect the psychological recovery that happens when people take breaks from stressful activities.

Each "day" in the simulation consists of one hundred time steps. This is meant to represent approximately one interaction every fifteen minutes during waking hours. Agents are limited to fifty swipes per day, reflecting typical platform constraints (like Tinder's daily like limit for free accounts). The simulation ran for thirty "days" and repeated this 'run' 100 times, to account for random variation.

So what were the results?

First, as I discussed at the top: there's a cyclical emotional trajectory across all groups. Emotional state starts relatively high, drops sharply in the early days, and then partially rebounds over time. Males tend to sustain higher emotional states throughout the simulation, while females plateau at lower levels. The discrepancy between expectations and reality persists, but rather than a steady monotonic decline, the model shows oscillations with partial recoveries.

Second, there are clear gender differences in behavioral patterns. Male agents show higher rates of both disengagement and re-engagement, averaging significantly more daily disengagements compared to females. They also show more behavioral volatility, with much broader variation around their average trajectories. Female agents show more stable engagement patterns but consistently lower overall emotional states.

The disengagement tendency analysis revealed that while agents became less likely to disengage overall as time progressed, disengagement followed an oscillatory pattern with multiple peaks rather than a simple downward slope. Males demonstrated much higher disengagement tendencies than females across all time periods. And both the hybrid and serious relationship-seeking strategies were associated with lower disengagement tendencies compared to casual strategies.

Third, the relationship-seeking strategy effects were more complex than initially expected. While simple averages suggested casual users disengaged more, the statistical models showed that serious seekers actually disengaged more frequently than casual seekers. Hybrid seekers did not differ significantly in main effects, though interaction terms revealed further nuance. Males adopting hybrid strategies disengaged more frequently than expected based on the main effects alone, while males with serious strategies disengaged less often than the casual baseline. This suggests that the combination of gender and relationship goals creates emergent behavioral patterns that wouldn't be visible from studying either factor in isolation.

For daily re-engagements, similar patterns emerged. Time had a consistent negative effect across all groups. Males with hybrid strategies re-engaged more often than other groups, while male serious seekers showed faster decline in re-engagement over time compared to other combinations.

The most interesting finding is what happens when you look at the interaction between gender and relationship strategy over time. At the end of the simulation, female agents showed lower and more tightly clustered emotional states, particularly when pursuing casual strategies, while male agents exhibited broader and higher distributions, especially when using casual relationship-seeking approaches. This suggests that the platform environment affects different user types very differently. Female agents converge toward similar emotional outcomes regardless of their relationship goals, while male agents show much more variation depending on the strategy they're using.

The authors also found that disengagement tends to be temporary rather than permanent. The simulation tracks transitions between active and inactive states, showing that agents who become inactive due to low emotional states gradually recover while offline, then re-engage when their emotional state improves sufficiently. This creates the cyclical pattern that many real users report experiencing (the one we talked about at the beginning of this episode).

What makes this study methodologically interesting is that it's using computational modeling to test theories about human behavior. Agent-based models like this one allow researchers to isolate specific mechanisms and test how they interact under specific conditions. The control theory framework treats emotional states not as random fluctuations but as the output of a feedback system that continuously monitors progress toward goals and adjusts behavior accordingly. As a whole, this paper demonstrates that computational modeling can provide insights into complex social and psychological phenomena that might otherwise be difficult or impossible to study experimentally. While you can't ethically manipulate people's emotional states or control their dating experiences, you can build simulations that test different theories about how these systems work.

If you want to dive deeper into the mathematical formulations behind their framework, see their statistical analysis, or explore their future plans, you'll definitely want to download the paper. Whether you're building dating apps yourself, or gamifying a different type of user-experience, the methodology they used here (and the analysis they produced) should give you a meaningful structure to apply to your own work.