An Adaptive Anxiety Intervention Recommender System Using DSW-UCB Contextual Multi-Armed Bandit with Relevance Filtering

In the quiet, often invisible landscape of mental health, anxiety is a condition that refuses to stay still. It shifts with the time of day, the weather, a sudden stressor at work, or a quiet moment of reflection. Because these feelings are so fluid, the tools designed to help manage them often struggle to keep pace. A coping strategy that calms a person on a Tuesday morning might feel useless or even irritating by Thursday evening. For years, digital health applications have tried to solve this by offering personalized suggestions, but many of these systems rely on a fundamental flaw: they treat every past interaction as if it were happening right now. They remember a user's preferences from months ago with the same intensity as their feelings from five minutes ago, creating a digital memory that is too heavy with history to adapt to the present moment.

This is where a new approach, developed by researchers at Usmanu Danfodiyo University in Nigeria, offers a different path. The team set out to build a smarter system for recommending anxiety interventions, one that understands that human emotions are non-stationary, meaning they change over time and do not follow a fixed pattern. Instead of trying to learn from a user's entire history, their new system focuses on what matters most: the recent past. By combining two specific techniques—forgetting old data quickly and filtering out unhelpful feedback—they created a model that can pivot almost instantly when a user's emotional state shifts. The result is a digital assistant that does not just remember, but truly adapts, learning to offer the right breathing exercise or mindfulness activity at the exact moment it is needed.

The researchers began by acknowledging a critical limitation in existing technology. Most current recommendation systems operate on a "full-history" basis. Imagine a student who studies for a test by reading every single textbook they have ever owned, from first grade to the present, without ever closing a book. While this student has a vast amount of information, they are likely to be confused by outdated facts and unable to focus on what is relevant to the test happening today. In the same way, a mental health app that weighs a user's calm state from six months ago equally with their current panic attack will make poor recommendations. The system becomes slow to react, stuck in a loop of suggestions that no longer fit the user's reality. The researchers argued that for anxiety interventions to be effective, the system must prioritize recent interactions and discard the noise of the distant past.

To solve this, the team designed a system that uses a "sliding window" of memory. Instead of looking at the entire history of a user's life, the system only pays attention to the last hundred interactions. As a new interaction occurs, the oldest one falls out of view, ensuring the system is always learning from the most current data. However, simply looking at the last hundred moments is not enough; some of those moments might be confusing or unhelpful. Perhaps a user clicked on a suggestion but immediately closed the app, or perhaps they gave a rating that didn't match their actual feelings. To handle this, the researchers added a "relevance filter." This mechanism acts like a quality control inspector, assigning a score to every interaction based on how engaged the user seemed. If a session was short and disengaged, the system gives it very little weight in its learning process. If the user spent time with the suggestion and seemed to benefit, that interaction counts for much more. This prevents the system from being misled by accidental clicks or moments of distraction.

The team tested their new system, which they called a Discounted Sliding-Window Upper Confidence Bound model, in a controlled simulation. They created a digital environment where a virtual user's anxiety levels and context changed abruptly, mimicking the real-world unpredictability of human emotion. They pitted their new system against several older models, including the traditional full-history approach and other variations that used only a sliding window or only a discount factor. The results were clear and significant. In a scenario where the user's needs changed suddenly, the traditional system took over 600 steps to figure out what the new best intervention was. In contrast, the new system with the relevance filter recovered in just 235 steps, getting back on track more than twice as fast.

The performance gap was even more striking when looking at the total success of the recommendations. Over the course of the simulation, the new system accumulated 20.9% more positive outcomes than the traditional full-history model. It also reduced the "regret"—a measure of how many times the system made a suboptimal choice—by 61.1%. This means the system was not only faster at adapting but also made better decisions consistently. The researchers found that the size of the memory window mattered, too. They tested windows ranging from 50 to 500 interactions and discovered that a window of 100 interactions offered the best balance. It was large enough to provide stable learning but small enough to react quickly to changes. When they combined this specific window size with their relevance filter, the system achieved the highest possible reward and the lowest regret in their tests.

The study also revealed that the way the system handles bad data is crucial. When the researchers tested a version of the system that aggressively discarded any low-quality interaction, the results became unstable. The system swung wildly between good and bad recommendations. However, the version that used a "soft" filter—giving low-quality interactions a tiny bit of weight rather than ignoring them completely—maintained a steady, reliable performance. This suggests that even noisy or confusing feedback has some value, as long as it is not allowed to dominate the learning process. The system learned to be robust, ignoring the noise without losing the signal.

While these results are promising, the researchers are careful to note that they were achieved in a simulated environment. The study did not involve real human participants or clinical trials; it was a rigorous test of the algorithm's logic and adaptability within a computer model. The authors emphasize that the next step is to test this framework with real users under ethical supervision, incorporating actual physiological signals like heart rate or sleep patterns to make the context even richer. They also plan to explore how to make the system more transparent, ensuring users understand why a specific suggestion was made and giving them control to override it if needed.

The core achievement of this work is the demonstration that a recommendation system can be both fast and accurate in a chaotic environment. By letting go of the past and carefully weighing the present, the new model avoids the trap of historical bias that plagues many digital health tools. It offers a blueprint for a future where mental health support is not a static library of advice, but a dynamic conversation that evolves with the user. In a world where anxiety is often defined by its unpredictability, a system that can move with that unpredictability represents a significant step forward in making digital care truly personal and effective.