💻 computer science

Leakage-Aware 5G Infill-Site Prioritization via Robust Submodular Portfolio Optimization

This paper proposes ARGO-5G, a leakage-aware robust submodular optimization algorithm that prioritizes 5G infill sites by balancing future demand capture, regional service equity, and engineering constraints without relying on future labels, demonstrating superior portfolio balance and deployability on a dataset of over 14,000 macro sites.

Original authors: Runzhe Liu, Dabin Wang, Yun He

Published 2026-09-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Runzhe Liu, Dabin Wang, Yun He

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Mobile network operators are currently navigating a complex phase of expansion known as "brownfield" planning. Unlike building a network from scratch on empty land, this stage involves upgrading thousands of existing cell towers that are already in service. The central challenge is not simply predicting where traffic will be heavy next month, but deciding which specific towers to upgrade first when resources are limited. If operators rely solely on current traffic data, they risk pouring all their budget into a few already crowded districts while neglecting other areas that need service, or they might select sites that are technically difficult and expensive to upgrade. The goal is to find a balanced portfolio of upgrades that captures future demand, spreads service evenly across different cities, and respects the physical difficulty of the construction work, all without knowing the exact traffic patterns of the future.

Researchers at Yunnan Communications Vocational and Technical College have addressed this dilemma by developing a new method called ARGO-5G. Rather than trying to predict the future with perfect accuracy, the team treated the selection of upgrade sites as a puzzle of balancing multiple competing needs. They analyzed data from 14,129 physical macro sites across 16 cities in China. The core of their approach involves a "leakage-aware" strategy, which means they strictly separated the data used to make decisions from the data used to test them. They used information available in December to build their list of recommended sites, but they kept the actual traffic data from January completely sealed and hidden until the very end. This ensured that their method was not accidentally using future knowledge to make past decisions.

The researchers found that simply ranking sites by their current traffic volume was insufficient. While a standard ranking based on December traffic could predict January traffic with reasonable accuracy, it resulted in a list of upgrades that was unbalanced. It tended to select too many sites in a few high-traffic cities and ignored others, leading to a situation where some cities received no new upgrades at all. Furthermore, the standard list included many sites that were technically complex and difficult to build, which could slow down the entire project. The new ARGO-5G method, however, successfully reshaped the list. By using a mathematical framework that values diversity and balance, the algorithm selected a portfolio of 100 sites that captured nearly the same amount of future traffic as the standard method but spread that coverage across 16 cities instead of just 10.

Crucially, this improvement in balance did not come at the cost of performance. The new method captured 1.334% of the total future traffic burden, a figure almost identical to the 1.343% captured by the standard traffic-based ranking. However, the composition of the list changed significantly. The new approach reduced the number of highly complex, difficult-to-build sites from 49% of the list down to 32%, making the project more feasible for engineers to execute. It also ensured that every city in the region received at least some attention, raising the minimum service level for the least-served city from zero to a small but meaningful fraction. The researchers verified these results through rigorous testing, including simulations that introduced random shifts in traffic patterns and stress tests that varied the priorities of the network planners. In every scenario, the new method maintained its ability to capture future demand while providing a more equitable and practical list of upgrades.

This work demonstrates that the best way to manage a network upgrade is not just to predict where the traffic will be, but to carefully construct a list that balances demand with geography and engineering reality. The study explicitly rules out the idea that a more complex traffic prediction model is the solution; in fact, their results showed that a sophisticated predictor did not solve the problem of unbalanced city coverage. Instead, the solution lay in the selection process itself. By treating the upgrade list as a portfolio that must satisfy multiple constraints simultaneously, the researchers created a system that is both robust and fair. The findings suggest that mobile operators can achieve better network performance and more efficient use of their construction budgets by adopting this balanced approach, ensuring that the benefits of 5G technology are distributed more evenly across the regions they serve.

Technical Summary: Leakage-Aware 5G Infill-Site Prioritization via Robust Submodular Portfolio Optimization

Problem Statement
The paper addresses the challenge of "brownfield" 5G network expansion, where mobile operators must select a fixed-size portfolio of existing physical macro sites for upgrade before future congestion is observed. The core difficulty lies in balancing multiple, often conflicting, objectives: capturing future traffic demand, ensuring regional service balance across cities, managing engineering workload (complexity), and avoiding spatial redundancy. Traditional approaches that rely solely on traffic forecasting or simple ranking by current backflow often concentrate resources in high-traffic districts, ignore regional equity, or select sites with high engineering costs but low marginal deployment value. The authors argue that site selection is a set decision problem where the value of a site depends on the other sites already selected, a characteristic not captured by standard point-wise prediction models.

Methodology: ARGO-5G
The authors propose ARGO-5G, a robust submodular portfolio optimization framework designed to be "leakage-aware" (i.e., it does not use future data for decision-making). The methodology consists of three main components:

Preference-Robust Utility Construction:
Instead of relying on a single fixed weighting of criteria (users, backflow, traffic intensity), the model treats planning preferences as uncertain. It uses a Dirichlet distribution to sample preference weights, calculating a risk-adjusted utility for each site. This utility ( $\rho_i$ ) is defined as the mean preference score minus a risk term proportional to the variance of the score ( $\gamma \sigma_i$ ). The final deployable score ( $a_i$ ) combines this risk-adjusted utility with direct evidence from the current period (December burden), weighted by a parameter $\omega$ . Crucially, future labels (January data) are strictly sealed and used only for evaluation.
Submodular Set Objective:
The selection problem is formulated as maximizing a set function $F(S)$ subject to a cardinality constraint $K$ . The objective function combines three terms:
- Demand-Complexity Utility ( $U(S)$ ): The sum of risk-adjusted scores minus a penalty for engineering complexity ( $\eta c_i$ ).
- Spatial Diversity ( $D(S)$ ): A term based on Haversine distance to geographic landmarks, modeled to capture diminishing returns (submodularity) as sites near landmarks are selected.
- City-Service Saturation ( $Q(S)$ ): A term that rewards coverage of different cities but saturates the contribution once a city-specific threshold ( $\tau_c$ ) is reached, preventing over-concentration in a few cities.
The authors prove that under non-negative weights, this objective is monotone submodular. This property allows the use of a greedy algorithm with a theoretical approximation guarantee of $(1 - 1/e)$ relative to the optimal solution.
Temporal Validation Protocol:
The study employs a strict temporal holdout. December data is used for scoring and optimization, while January data is sealed. A separate "retrospective audit" uses city-grouped Gradient Boosting models to forecast January burden, but this predictor is explicitly not used in the deployable portfolio construction to avoid data leakage.

Key Contributions

Formalization: The paper formalizes 5G infill-site prioritization as a temporally separated, leakage-aware portfolio selection task, distinguishing between predictive signal generation and the actual set decision.
Auditable Objective: It derives a closed-form objective function that couples preference-risk utility (via Dirichlet sampling), spatial landmark coverage, city-service saturation, and engineering complexity calibration.
Algorithmic Framework: It introduces ARGO-5G, an auditable greedy algorithm that provides a theoretical guarantee for the surrogate objective while optimizing for real-world deployment constraints.
Comprehensive Evaluation: The framework is evaluated on a dataset of 14,129 physical macro sites across 16 cities and 130 counties/districts. The evaluation includes comparisons against traffic-only, multi-criteria, and quota-based baselines, as well as extensive ablation studies, Pareto landscape analysis, and stress tests (pseudo-windows and preference stress).

Results
At a budget of $K=100$ sites, the ARGO-5G-C variant (incorporating complexity penalties) achieved the following results compared to a December-proxy ranking baseline:

Future Burden Capture: Captured 1.334% of the total January burden, nearly identical to the baseline (1.343%) and close to the theoretical Oracle (1.385%).
Regional Balance: Improved the Jain city-service index from 0.417 to 0.695 and increased the number of cities covered from 10 to 16.
Engineering Efficiency: Reduced the number of selected cell records from 216 to 190 and lowered the share of high-complexity sites from 49.0% to 32.0%.
Robustness: The algorithm maintained high capture rates across various sensitivity sweeps, preference stress tests, and calibrated drift simulations, demonstrating that the gains in balance and efficiency did not come at the cost of demand capture.
Ablation: Component analysis confirmed that city saturation terms drove the balance improvement, while complexity penalties reduced workload, and preference robustness stabilized utility without altering aggregate balance at this specific budget.

Significance and Claims
The paper claims that ARGO-5G provides a reproducible bridge between mobile traffic analytics and operator-facing 5G upgrade planning. Its primary significance lies in demonstrating that predictive accuracy alone is insufficient for network planning; a disciplined conversion of predictive signals into a constrained set decision is necessary to achieve operational goals like regional equity and engineering feasibility.

The authors emphasize that the value of the work is not in a new traffic predictor (the forecasting audit showed strong signal quality but was not the decision engine) but in the portfolio optimization layer. This layer transforms a ranked list into a deployable portfolio that is:

Leakage-aware: Strictly avoiding the use of future data for selection.
Auditable: Decomposing the decision into interpretable components (demand, space, city balance, complexity).
Balanced: Achieving a Pareto-optimal trade-off between capturing future demand and ensuring a geographically and operationally balanced upgrade list.

The work aligns with emerging trends in O-RAN and digital twins by providing explainable decision systems that separate prediction from selection, allowing operators to manage the trade-offs between demand capture, service equity, and engineering constraints without relying on opaque "black box" rankings.

Technical Summary: Leakage-Aware 5G Infill-Site Prioritization via Robust Submodular Portfolio Optimization

More like this