Traditional recommendation and advertising systems often operate under a naive assumption: if a user has a high baseline probability of clicking an item, showing it to them repeatedly will yield consistently high engagement. In reality, human attention is a depletable resource. When exposed to the same or similar content multiple times, users rapidly desensitize—a psychological cliff where engagement drops precipitously. As Jiakai Wang notes, performance can plummet by over a third within just a few weeks of high-frequency exposure.
To build systems that accurately forecast and optimize user experiences, researchers have shifted from static click-prediction models to frameworks that explicitly model this decay. This paradigm treats user attention not as a fixed constant, but as a dynamic state heavily influenced by sequential externalities and creative saturation.
Modeling Fatigue as a Sequential Externality
Why does a user stop clicking an ad they previously found interesting? In sequential environments, every ad shown acts as a "negative externality" on future slots. Nicola Gatti et al. formalized this in Mechanism Design for Mobile Geo-Location Advertising by introducing a count-based discount factor. Instead of tracking the exact identity of every past exposure—which quickly becomes computationally intractable in real-time auctions—the system scales down the expected click-through rate (CTR) based purely on the number of prior exposures.
This count-based abstraction makes optimization tractable. For instance, in Yahoo's native ad engine, Michal Aharon et al. implemented "Soft Frequency Capping" in Soft Frequency Capping for Improved Ad Click Prediction in Yahoo Gemini Native. Rather than enforcing hard, rule-based limits on ad repetitions, they integrated historical exposure counts directly into a machine-learned click prediction model. By explicitly penalizing the predicted CTR of repeated content, the model avoids over-serving stale ads, leading to a substantial lift in platform revenue.
The Value of What is Ignored: Negative Feedback Loops
Most sequential recommendation systems focus heavily on positive interactions like clicks and purchases. However, to capture the true onset of user saturation, models must listen to what the user doesn't do.
When an item is displayed but unclicked, it provides a strong signal of diminishing marginal utility. In Know in AdVance: Linear-Complexity Forecasting of Ad Campaign Performance with Evolving User Interest, Xiaoyu Wang et al. introduced a Conditional State Space Model (SSM) that explicitly ingests these negative feedback signals. By tracking displayed-but-ignored items alongside positive history, the model prevents the system from overestimating future campaign yields. Similarly, in the context of sequential recommendations, Nian Li et al. developed the FRec framework in Modeling User Fatigue for Sequential Recommendation to balance relevance with variety, preventing the recommendation loop from trapping users in a state of over-saturation.
This shift is also crucial for causal bidding. As Martin Bompaire et al. demonstrate in Causal Models for Real Time Bidding with Repeated User Interactions, as a user's interaction frequency increases, the incremental impact of showing another ad actually decreases. Traditional models that ignore this tend to over-invest in highly active users who would have converted anyway, wasting budget on saturated attention.
Strategic Scheduling: Familiarity vs. Satiation
If repeated exposure causes decay, what is the optimal way to distribute ads across a user session? Intuition might suggest uniform spacing—spreading ads evenly to give the user a "break." However, behavioral psychology reveals a fundamental tension: a user needs some exposure to build brand familiarity, but too much triggers annoyance and satiation.
In Ads that Stick: Near-Optimal Ad Optimization through Psychological Behavior Models, Kailash Gopal Darmasubramanian et al. modeled this tension mathematically. They discovered that when users are highly sensitive to overexposure, the mathematically optimal strategy is actually to cluster ads at the very beginning and the very end of a session. This non-uniform scheduling maximizes the psychological impact of familiarity while confining the interruption cost to specific boundaries, rather than continuously draining user attention throughout the entire session.
This boundary is highly economic: as shown by Haoran Yu et al. in Monetizing Mobile Data via Data Rewards, this wear-out effect creates a hard ceiling on advertising-based revenue models. Pushing past this ceiling by increasing ad volume simply drives down the price advertisers are willing to pay, as user engagement hits an inescapable psychological wall.
Algorithmic Countermeasures: Creative Diversification
While scheduling and frequency capping help manage the volume of exposure, the ultimate tool to combat attention decay is content variety. The earliest insights into this date back to Charu C. Aggarwal et al. in A Framework for the Optimizing of WWW Advertising, who noted that pages with high user recurrence require rapid advertisement rotation to overcome user habituation and "ad blindness."
Modern approaches automate this rotation by generating diverse creative variations from the same underlying product or source material. For example, Akihiko Kato et al. introduced the CAMERA³ dataset in CAMERA³: An Evaluation Dataset for Controllable Ad Text Generation in Japanese to train models that can dynamically pivot between different psychological appeals (such as emphasizing price versus quality). By changing the angle of the message, the system resets the user's desensitization curve.
To scale this process, Shaunak Mishra et al. in Learning to Create Better Ads: Generation and Ranking Approaches for Ad Creative Refinement and Yichao Zhou et al. in Recommending Themes for Ad Creative Design via Visual-Linguistic Representations leverage multimodal frameworks to automatically recommend new visual and textual themes. By continuously injecting fresh, targeted variations into the pipeline, platforms can sustain consumer attention and maintain long-term campaign effectiveness without causing user burnout.
Key Foundational Work
Personalized Ad Delivery When Ads Fatigue: An Approximation Algorithm (2007) — contributed 3 concept elements to this direction.
A Framework for the Optimizing of WWW Advertising (1998) — contributed 2 concept elements to this direction.