FARM: A Fairness-Aware Recommendation Method for High Visibility and Low Visibility Mobile APPs

Qiliang Zhu, Qibo Sun, Zengxiang Li, Shangguang WangIEEE Access 2020 · DOI

Introduction

In the rapidly expanding mobile application market, recommender systems have become indispensable tools for helping users navigate millions of choices. However, these systems often fall prey to a "rich-get-richer" phenomenon. Popular apps with extensive historical data continue to be recommended, while new or niche applications—regardless of their quality—remain buried. This paper introduces FARM (Fairness-Aware Recommendation Method), a novel approach designed to balance recommendation accuracy with market fairness by providing equitable exposure to low-visibility applications.

The Visibility Bias Problem

Traditional recommendation algorithms, such as collaborative filtering and popularity-based models, rely heavily on historical user evaluations. This creates a significant "cold start" problem for new apps. Without initial ratings, these apps are rarely recommended, preventing them from ever acquiring the data needed to improve their visibility. This bias not only stifles innovation for developers but also limits the variety of options available to users. While some existing methods use metadata to predict quality, they often fail to guarantee a fair opportunity for these apps to be discovered by the user base.

The FARM Methodology

The core innovation of FARM is its dual-path processing architecture, which classifies applications into two categories: High Visibility (HV) and Low Visibility (LV), based on a threshold of historical evaluations.

The system architecture of the FARM method, illustrating the separate processing pipelines for high and low visibility applications.

Processing Low Visibility (LV) Apps

For apps in the LV category, FARM introduces a dynamic "fairness factor." This factor is initialized based on baseline reputation metrics (such as safety and absence of intrusive ads) and naturally decays over time to ensure that stagnant, unadopted apps eventually fade out. Crucially, this factor is adjusted in real-time based on the very first user feedback received. To select apps for recommendation, FARM employs a Roulette-Wheel selection algorithm. This probabilistic method ensures that while higher-quality LV apps are prioritized, every new app maintains a mathematical chance of being showcased to a user.

Processing High Visibility (HV) Apps

For established apps with ample data, FARM utilizes a Fuzzy Analytic Hierarchy Process (FAHP). This multi-criteria decision-making model ranks apps by synthesizing various metrics, including overall popularity, average rating, and rating stability.

Hybrid Recommendation

The final recommendation list is a blend of HV and LV apps. The proportion of low-visibility apps is controlled by a tunable parameter ($\gamma$), allowing system administrators to balance the degree of exploration (fairness) versus exploitation (accuracy).

Experimental Evaluation

The authors evaluated FARM using the AnzhiDataSet, a collection of over 17,000 records crawled from a major Chinese Android app store. The study focused on two primary fairness metrics:

  • Qualification Fairness: Measured using Jain’s fairness index, FARM achieved a peak score of approximately 0.827, significantly outperforming traditional baselines like CST and POP.
  • Opportunity Fairness: FARM consistently provided a more equitable distribution of recommendation slots to LV apps as the list size increased.

While FARM showed a slight decrease in traditional precision and recall compared to purely accuracy-focused models, the authors argue this is a mathematical artifact. Traditional metrics treat unrated apps as "not excellent" by default; however, FARM's real-world simulations demonstrated that the system effectively identifies high-quality new apps and increases their exposure while quickly filtering out low-quality ones based on initial feedback.

Conclusion

FARM represents a significant step toward more ethical and balanced recommender systems. By explicitly accounting for visibility bias and employing dynamic fairness adjustments, it ensures that high-quality new applications have a genuine path to success. Future work will focus on integrating personalized user preferences and contextual data to further refine recommendation accuracy without compromising the system's commitment to fairness.

Summary updated

BibTeX
@article{zhu2020farm,
  author = {Qiliang Zhu and Qibo Sun and Zengxiang Li and Shangguang Wang},
  title = {{FARM: A Fairness-Aware Recommendation Method for High Visibility and Low Visibility Mobile APPs}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.3007617},
}
Figure