⚡ electrical engineering

Deep Learning-Assisted Multicast Subgrouping in Massive MIMO

This paper proposes a deep learning-assisted framework that utilizes PCA and LSTM networks to infer optimal multicast subgrouping configurations from spatial channel statistics, enhanced by transfer learning to efficiently maximize spectral efficiency in massive MIMO systems while overcoming pilot overhead and the worst-user effect.

Original authors: Alejandro de la Fuente, Leopoldo Carro-Calvo, Giovanni Interdonato

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alejandro de la Fuente, Leopoldo Carro-Calvo, Giovanni Interdonato

Original paper licensed under CC BY 4.0 (http://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

Imagine a massive concert hall (the Base Station) with hundreds of speakers (the Antennas) trying to play the same song for a crowd of thousands of people (the Users).

In a traditional setup, the sound engineer has two bad options:

The "Unicorn" Approach: Give every single person their own private speaker and a unique song version. This sounds great for everyone, but it requires thousands of microphones to talk to the engineer first (too much Pilot Overhead), and the system gets overwhelmed.
The "One-Size-Fits-All" Approach: Play one giant broadcast to the whole room. This is efficient, but the volume is limited by the person sitting in the worst spot (the Worst-User Effect). If one person is far away or has a bad view, the whole room has to turn the volume down so they can hear it.

The Problem: The "Crowd" is Messy

The paper argues that the crowd isn't random. People naturally cluster together. Some groups are sitting close to the stage with a clear view; others are in the back with a bad angle. If you treat the whole crowd as one group, you miss the opportunity to give the front row a louder, clearer signal without hurting the back row.

The goal is to split the crowd into smaller, smarter groups (Subgroups) based on where they are sitting and how they hear the sound. But here's the catch: How many groups should you make?

If you make 1 group, you get the "bad view" problem.
If you make 1,000 groups, you need 1,000 microphones (too expensive).
If you guess randomly, you might end up with a messy mix that doesn't work well.

The Solution: An AI "Crowd Manager"

The authors propose a Deep Learning system (specifically using a type of AI called LSTM) that acts like an expert crowd manager. Instead of asking every single person how they hear the music, the AI looks at the "spatial statistics"—basically, a map of where people are sitting and how the sound waves bounce around them.

Here is how the AI works, step-by-step:

1. The "Snapshot" and the "Compression" (PCA)

Imagine taking a photo of the crowd. The photo is huge and full of details. The AI first uses a technique called PCA (Principal Component Analysis) to shrink this giant photo into a small, manageable "thumbnail" that still keeps the most important details (like "there's a dense crowd on the left" and "it's empty on the right").

2. The "Memory" (LSTM)

The AI then feeds these thumbnails into a Long Short-Term Memory (LSTM) network. Think of this as a manager who reads the crowd one person at a time, remembering the patterns.

"Okay, Person A is here. Person B is nearby. They look similar. Let's group them."
"Person C is far away. They need a different group."
"Wait, Person D is in a totally different spot. Maybe we need a third group."

The magic of this AI is that it doesn't need to know exactly how many people are in the room beforehand. It can handle a crowd of 40 or 120 people and still figure out the best number of groups.

3. The "Transfer Learning" Upgrade

The authors added a special trick called Transfer Learning.

First, the AI learns to count: It gets trained to guess, "Based on this crowd map, we probably need 4 groups."
Then, the AI learns to predict the result: Instead of just guessing the number, the AI is fine-tuned to predict, "If we use 4 groups, the total sound quality (Spectral Efficiency) will be X. If we use 5, it will be Y."

This allows the AI to skip the trial-and-error phase. It doesn't have to test every possible combination of groups to find the best one; it just looks at the map and says, "4 groups is the sweet spot."

The Results: Why It Matters

The paper tested this AI against other methods:

Random Guessing: The AI was much better.
DBSCAN (A standard clustering tool): The AI was better, especially when the crowd was spread out or messy. DBSCAN is like a rigid rulebook; the AI is flexible.
Unicasting (One group per person): The AI was much more efficient, saving massive amounts of "microphone" resources.

The Big Win: The AI achieved 85% of the theoretical perfect performance without needing to know the exact layout of the crowd in advance. It works even if the data about the crowd is a little bit fuzzy or imperfect.

The "Real World" Analogy

Think of a teacher trying to teach a class of 100 students.

Old Way: The teacher speaks to the whole class at once. The smartest kids get bored, and the kids in the back can't hear.
New Way (The Paper): The teacher uses an AI assistant. The assistant looks at the room, sees that 30 kids are in the front row, 20 are in the back, and 50 are scattered. The AI tells the teacher: "Split the class into 3 groups. Group A gets a projector, Group B gets a whiteboard, and Group C gets a handout."
The Result: Everyone learns better, and the teacher doesn't have to write 100 individual lesson plans.

Summary

This paper presents a smart, AI-driven way to organize massive wireless networks. Instead of forcing everyone into one group or giving everyone their own private channel, the system uses deep learning to dynamically figure out the perfect number of groups based on where people are standing. It saves resources, improves speed, and works even when the data isn't perfect.

Technical Summary: Deep Learning-Assisted Multicast Subgrouping in Massive MIMO

Problem Statement
Efficient content delivery in massive multiple-input multiple-output (mMIMO) multicasting is constrained by two primary factors: pilot overhead and the "worst-user effect." In conventional multicast, a single transmission rate must serve all users, limiting performance to the user with the poorest channel conditions. Furthermore, assigning orthogonal pilots to every user (unicast) creates prohibitive overhead in large-scale systems. While multicast subgrouping—partitioning users into subsets with similar spatial channel statistics to apply distinct precoders—offers a solution, determining the optimal number of subgroups ( $G$ ) remains a challenging, largely unresolved problem. Existing heuristic methods (e.g., K-means with fixed $G$ ) or density-based clustering (e.g., DBSCAN) often fail to adapt to heterogeneous user distributions or require exhaustive search to find the $G$ that maximizes sum spectral efficiency (SE).

Methodology
The paper proposes a deep learning-assisted framework to infer the optimal number of multicast subgroups directly from users' spatial channel statistics (covariance matrices) without requiring instantaneous channel state information (CSI). The approach operates under a scalable CSI acquisition model where one pilot is reused per subgroup.

The core methodology consists of three main stages:

Input Preprocessing (Snapshot-Specific PCA): Since the number of users ( $K$ ) varies across scenarios, the authors employ a per-snapshot Principal Component Analysis (PCA). This transforms variable-size spatial covariance matrices into a fixed-length representation of Principal Component (PC) scores. If the number of users is less than the maximum system capacity ( $K_{max}$ ), the input vectors are zero-padded to maintain a consistent dimension for the neural network.
Sequential Encoding (LSTM): A Long Short-Term Memory (LSTM) network processes the sequence of user PC scores. The LSTM is chosen for its ability to handle variable-length sequences and capture latent spatial dependencies among users. The architecture includes a "data replication" enhancement where input data is replicated to always fill $K_{max}$ time steps, refining the model's internal state and improving robustness to variable user counts.
Output and Transfer Learning (TL):
- Base Model: The initial model treats subgroup inference as a multi-output classification task. It predicts the probability distribution of the number of subgroups ( $G$ ) for progressively larger user subsets within a single forward pass.
- Transfer Learning Extension (LSTM-TL-SE): To further optimize performance, a transfer learning approach is introduced. The pre-trained LSTM encoder (frozen) extracts spatial features, while a lightweight dense head is fine-tuned to directly predict the normalized sum SE for different candidate subgroup counts. The optimal $G$ is then selected as the one maximizing the predicted SE, bypassing the need for exhaustive search.

Key Contributions

LSTM-Assisted Subgrouping Mechanism: A novel framework that infers the number of multicast subgroups from spatial covariance structures, adapting to variable user set sizes without heuristic assumptions.
Snapshot-Specific PCA Pipeline: A preprocessing technique that converts variable-size covariance matrices into fixed-length feature representations tailored to each specific user distribution snapshot.
Transfer Learning for SE Estimation: An extension that reuses the spatial feature extractor to directly predict sum spectral efficiency, enabling near-optimal subgroup selection without exhaustive evaluation.
Intra-Service Subgrouping Strategy: A method that preserves multicast semantics (single content stream) while improving physical-layer efficiency through multiple precoders tailored to statistically homogeneous subsets.
Comprehensive Evaluation: Extensive simulations validating the approach against benchmarks including unicast, conventional multicast, random subgrouping, and DBSCAN clustering across clustered and sparse user distributions.

Results
Simulation results in a single-cell mMIMO scenario (64 antennas, up to 120 users) demonstrate the following:

Performance: The proposed LSTM (TL-SE) model achieves up to 85.2% of the theoretical maximum sum spectral efficiency (the upper bound found via exhaustive search). This outperforms the best practical benchmark, DBSCAN (79.2%), and significantly exceeds unicast (67.2%) and conventional multicast (71.4%) in diverse scenarios.
Robustness: The model maintains high performance in both clustered scenarios (where subgrouping is beneficial) and sparse scenarios (where it naturally reverts to unicast-like behavior, $G \approx K$ ).
Imperfect CSI: Under perturbed covariance information (simulating estimation errors), the transfer learning model shows the highest resilience, retaining approximately 80.1% of the upper-bound performance even with significant mismatch ( $\alpha=0.5$ ), outperforming DBSCAN in these conditions.
Generalization: When applied to Cell-Free (CF-mMIMO) scenarios without retraining, the models trained on single-cell data maintain small deviations from the optimal subgroup configuration, indicating that the learned spatial structure inference generalizes across different network architectures.
Complexity: The proposed method scales polynomially with the number of users and antennas, offering a computationally efficient alternative to optimization-based beamforming methods that require full instantaneous CSI.

Significance and Claims
The paper claims to present the first framework that jointly combines snapshot-specific dimensionality reduction, sequential modeling of user covariance structures, and transfer learning for spectral efficiency prediction in multicast subgrouping.

The authors emphasize that their approach is specifically designed for the pilot-limited mMIMO regime, where full per-user instantaneous CSI is unavailable or too costly to acquire. By inferring the subgroup structure from large-scale spatial statistics, the method enables adaptive subgroup configurations that mitigate the worst-user effect without incurring linear pilot overhead. The work highlights that the model learns the underlying spatial structure governing subgrouping, allowing it to correctly identify when subgrouping is beneficial (clustered users) and when it is not (sparse users), providing a robust solution for dynamic and unpredictable user topologies. The authors modestly note that while the framework generalizes well structurally, the mapping from subgroup configuration to specific spectral efficiency remains system-dependent, and future work should explore real-world measurement campaigns and multi-cell interference.