ICANEWS

Engineering & Technology

Causally Interpretable Wi-Fi CSI-Based Human Activity Recognition with Discrete Latent Compression

Human Activity Recognition (HAR) using Wi-Fi Channel State Information (CSI) has emerged as a significant area of research due to its potential for non-invasive

Introduction: Bridging the Gap in Human Activity Recognition

Human Activity Recognition (HAR) using Wi-Fi Channel State Information (CSI) has emerged as a significant area of research due to its potential for non-invasive and pervasive monitoring. While deep neural networks have shown strong predictive capabilities in CSI-based HAR (CHAR), they often suffer from a lack of interpretability, relying on continuous latent representations that are opaque and challenging to modify. Conversely, purely symbolic methods, while interpretable, struggle to directly process the raw, high-dimensional CSI streams. A new study, detailed in the arXiv preprint arXiv:2604.22979v1, addresses these challenges by proposing a novel, fully automatic and strictly decoupled pipeline specifically designed for causally interpretable Wi-Fi CSI-based Human Activity Recognition. The research, titled "Towards Causally Interpretable Wi-Fi CSI-Based Human Activity Recognition with Discrete Latent Compression and LTL Rule Extraction," introduces a method that combines the ability to process raw CSI data with the benefits of symbolic interpretability, aiming to provide a viable alternative to existing 'black-box' models.

The core objective of this research is to meet specific requirements: causal interpretability, symbolic controllability, and direct operation on high-dimensional raw signals. By fusing these elements, the proposed system, referred to as CHAR Latent Temporal Rule Extraction (CHARL-TRE), seeks to overcome the limitations of current approaches in the field. This addresses a critical need for models that not only perform well but also provide insights into their decision-making process, which is particularly crucial in sensitive applications of HAR.

The Challenge of Interpretability in CSI-based HAR

The landscape of CSI-based Human Activity Recognition is currently dominated by deep neural models. These models are adept at extracting complex patterns from the vast and high-dimensional data provided by Wi-Fi CSI, leading to impressive predictive performance. However, their internal workings often remain obscure. The 'continuous latent representations' upon which these models operate are inherently difficult to deconstruct or alter, making it challenging to understand why a particular activity was identified. This lack of transparency, known as the 'black-box' problem, can hinder trust and adoption in applications where accountability and explainability are paramount.

“Deep neural models achieve strong predictive performance on CSI-based HAR (CHAR), yet rely on continuous latent representations that are opaque and difficult to modify.”

In contrast, purely symbolic approaches offer inherent interpretability, as their decision-making processes can be directly understood through symbolic rules. However, their primary drawback lies in their inability to directly process raw CSI streams. These symbolic systems often require a pre-processing step to convert the raw, continuous data into a symbolic form, which can lead to information loss or introduce biases. The research aims to bridge this gap, allowing for direct processing of raw CSI signals while maintaining a symbolic, interpretable output.

Research Goal: Achieving Causal Interpretability with Symbolic Control

The central research objective is to develop a system for Human Activity Recognition using Wi-Fi Channel State Information (CSI) that simultaneously fulfills three stringent requirements: causal interpretability, symbolic controllability, and direct operation on high-dimensional raw signals. The study explicitly states its aim to address these joint requirements, which are often at odds in traditional HAR methodologies.

Addressing Joint Requirements for Advanced HAR

The first requirement, 'causal interpretability', refers to the ability to understand the underlying causes or reasons behind the model's predictions. This goes beyond mere pattern recognition, aiming to establish clear cause-and-effect relationships within the data that lead to an activity classification. For instance, rather than simply stating an activity occurred, the goal is to explain which specific changes in CSI indicate that activity.

The second requirement, 'symbolic controllability', emphasizes the ability to modify or influence the model's logic through interpretable symbols. This means that if a rule is identified as incorrect or needing adjustment, it can be directly altered at a symbolic level, rather than requiring extensive retraining of a complex deep learning model. This offers a level of human oversight and intervention that is often absent in 'black-box' systems.

Finally, 'direct operation on high-dimensional raw signals' ensures that the proposed method can process the raw, unprocessed Wi-Fi CSI data without relying on extensive feature engineering or manual pre-processing steps. This preserves the richness of the original data and avoids potential information loss or biases introduced by intermediate data transformations. The combined fulfillment of these requirements defines the novelty and ambition of the research, moving towards more transparent and adaptable HAR systems.

Key Findings: A Decoupled Pipeline for Interpretable Classification

The research presents several key findings, demonstrating the viability of their proposed approach, CHARL-TRE. The primary finding is the successful development and evaluation of a fully automatic and strictly decoupled pipeline that can process CSI magnitude windows, compress them into discrete representations, infer causal relationships, and extract Linear Temporal Logic (LTL) rules for symbolic classification. This pipeline achieves competitive performance while explicitly preserving temporal and causal structure.

The Decoupled Pipeline Architecture

The proposed pipeline operates in a series of distinct stages. Firstly, CSI magnitude windows are processed. These windows represent segments of the raw Wi-Fi CSI data over time. The core of the initial processing involves compression using a categorical variational autoencoder (VAE) specifically designed with Gumbel-Softmax latent variables. This VAE operates under a capacity-controlled objective, which is crucial for yielding a compact and discrete representation of the high-dimensional input. The use of Gumbel-Softmax allows for a continuous approximation of discrete variables, facilitating training while ultimately resulting in a truly discrete output. This discrete representation is a fundamental step towards symbolic interpretability, as it transforms continuous, complex data into manageable, distinct categories.

Once the encoder part of this VAE is trained, it becomes frozen and is subsequently used as a deterministic mapping. This mapping transforms raw CSI windows into one-hot latent trajectories. These trajectories represent sequences of discrete latent states, effectively converting the continuous temporal evolution of CSI into a symbolic sequence. This step is critical because these 'one-hot latent trajectories' serve as the basis for the subsequent causal discovery process.

Causal Discovery and LTL Rule Extraction

Following the generation of these discrete latent trajectories, causal discovery techniques are applied. Specifically, the research states that causal discovery is performed on these trajectories to estimate class-conditional temporal dependency graphs. These graphs aim to uncover statistical relationships indicating cause and effect between the discrete latent variables over time, conditioned on the specific human activity being performed. The focus on 'class-conditional' graphs implies that the causal dependencies might differ depending on the activity (e.g., walking might have different causal patterns than sitting).

The statistically supported lagged dependencies identified from these temporal dependency graphs are then translated into Linear Temporal Logic (LTL) rules. LTL is a formal system for reasoning about sequences of events, making it ideal for capturing temporal relationships in a symbolic and machine-interpretable format. These LTL rules form the basis of the classifier. The resulting system is described as a 'fully symbolic and deterministic classifier' that relies solely on rule evaluation and aggregation, explicitly without any learned discriminative head. This absence of a traditional discriminative head underlines the symbolic nature of the classification and its reliance on interpretable rules rather than learned weights within an opaque neural network.

Competitive Performance with Explicit Structure

A significant finding is that CHARL-TRE indicates competitive performance when compared to existing methods. This is particularly noteworthy because it achieves this performance while concurrently preserving explicit temporal and causal structure. Many high-performing CHAR models achieve their accuracy at the expense of interpretability, obscuring the temporal and causal relationships within the data. CHARL-TRE demonstrates that it is possible to maintain both high performance and a transparent understanding of the model's reasoning. The phrase 'explicit temporal and causal structure' highlights that the model doesn't just infer an activity; it provides a structured explanation of *when* and *how* different latent states (derived from CSI) lead to that activity.

Structured Multi-Antenna Fusion

Another important finding relates to the potential for structured multi-antenna fusion. Because the extracted rules are defined over discrete latent variables, the research suggests that antenna-specific rule sets can, in principle, be combined at the symbolic level. This capability enables 'structured multi-antenna fusion without retraining the encoder'. In typical multi-antenna systems, integrating data from multiple antennas often requires retraining the entire model, which can be computationally intensive and time-consuming. The symbolic nature of CHARL-TRE's rules offers a more efficient and flexible way to integrate information from multiple sources, potentially leading to more robust and accurate HAR systems without the overhead of complete model retraining.

“Results from CHAR Latent Temporal Rule Extraction (CHARL-TRE) indicate competitive performance while preserving explicit temporal and causal structure, showing that deterministic symbolic classification grounded in unsupervised discrete latent representations constitutes a viable alternative to end-to-end black-box models for wireless HAR.”

Methodology: A Decoupled Pipeline for Causal HAR

The methodology employed in this research is characterized by its fully automatic and strictly decoupled pipeline, designed to transform raw Wi-Fi CSI signals into interpretable, symbolic classifications of human activities. This section details the distinct stages of this pipeline as described in the source material.

Stage 1: CSI Magnitude Window Compression

The initial stage of the methodology involves processing Wi-Fi CSI. Specifically, 'CSI magnitude windows' are the input. These windows represent segments of the continuous CSI data over a certain period. The core technique here is compression using a 'categorical variational autoencoder (VAE) with Gumbel-Softmax latent variables'. This VAE is trained under a 'capacity-controlled objective'. The purpose of this specific VAE architecture is to transform the high-dimensional, continuous CSI magnitude windows into a 'compact discrete representation'. The Gumbel-Softmax reparameterization trick allows for gradient estimation through discrete variables during training, effectively enabling the VAE to learn discrete latent codes. The 'capacity-controlled objective' helps in ensuring that the latent space is sufficiently compact and meaningful, preventing the VAE from learning an overly diffuse or complex representation.

The output of this stage is a set of discrete latent variables, each representing a compressed summary of a CSI magnitude window. This transformation from high-dimensional continuous data to low-dimensional discrete symbols is a critical enabler for the subsequent symbolic processing.

Stage 2: Deterministic Mapping to One-Hot Latent Trajectories

Once the categorical VAE's encoder is trained and its parameters are learned, it is 'frozen'. This means its weights are no longer updated. In this frozen state, the encoder acts as a 'deterministic mapping'. This fixed mapping is then used to convert new CSI magnitude windows into 'one-hot latent trajectories'. A one-hot representation for a discrete variable means that for a variable with $N$ possible states, its value is represented by a vector of length $N$ where one element is 1 and all others are 0, indicating the active state. Therefore, a 'one-hot latent trajectory' is a sequence of these one-hot vectors, representing the temporal evolution of the discrete latent states derived from the CSI signals.

This deterministic mapping ensures consistency and allows the subsequent stages to operate on a stable, symbolic representation of the temporal dynamics embedded within the CSI data. The 'trajectories' are sequences of these discrete states over time, which are essential for uncovering temporal dependencies.

Stage 3: Causal Discovery on Latent Trajectories

With the one-hot latent trajectories now available, the next methodological step is 'causal discovery'. This process is performed specifically on these trajectories. The objective is to 'estimate class-conditional temporal dependency graphs'. These graphs are designed to identify statistical dependencies that suggest a causal relationship between the discrete latent variables over time. The 'class-conditional' aspect implies that these dependencies are analyzed specifically within the context of different human activities. For example, the causal flow of latent states when someone is 'walking' might be distinct from when they are 'sitting'.

This stage moves beyond mere correlation, attempting to establish the direction and strength of influence between the symbolic latent states. The outcome is a graphical representation of how these discrete symbolic events evolve and influence one another, conditioned on the activity in question. The source does not specify the exact causal discovery algorithms used.

Stage 4: LTL Rule Extraction and Symbolic Classification

The final stage translates the findings from causal discovery into a directly usable classification mechanism. 'Statistically supported lagged dependencies' identified in the temporal dependency graphs are converted into 'Linear Temporal Logic (LTL) rules'. LTL is a formal logic used for describing how a system behaves over time, making it particularly suitable for encoding temporal patterns and causal relationships. These rules are inherently symbolic and human-readable, fulfilling the interpretability requirement.

The collection of these LTL rules forms a 'fully symbolic and deterministic classifier'. Crucially, the source states that this classifier operates 'based solely on rule evaluation and aggregation, without any learned discriminative head'. This means the classification decision is made directly by evaluating whether the observed sequence of latent states matches the defined LTL rules, and then aggregating the results from multiple rules. This design choice fundamentally differentiates CHARL-TRE from traditional neural network classifiers that rely on complex, non-linear functions within a 'discriminative head' to make predictions. The absence of such a head reinforces the system's transparency and symbolic controllability.

Implications: Enhancing Interpretability and Multi-Antenna Fusion

The successful development and validation of CHARL-TRE carry significant implications for the field of Wi-Fi CSI-based Human Activity Recognition. The research highlights two primary implications: providing a viable alternative to existing 'black-box' models and enabling structured multi-antenna fusion.

A Viable Alternative to 'Black-Box' Models

The most prominent implication is that CHARL-TRE constitutes a 'viable alternative to end-to-end black-box models for wireless HAR'. This is particularly important in domains where transparency, trust, and the ability to explain decisions are paramount. While deep learning models offer high performance, their opaque nature can be a barrier to adoption in critical applications like healthcare monitoring, elder care, or security systems. By achieving 'competitive performance' while 'preserving explicit temporal and causal structure', CHARL-TRE offers a pathway to deploy HAR systems that are not only accurate but also understandable and debuggable.

The phrase 'deterministic symbolic classification grounded in unsupervised discrete latent representations' encapsulates this implication. It means that the model's decisions are based on clear, logical rules derived from an unsupervised understanding of the underlying patterns in the CSI data, rather than being an outcome of an uninterpretable, high-dimensional calculation. This interpretability allows practitioners to understand *why* a certain activity is detected, facilitating fault diagnosis, model improvement, and regulatory compliance.

Enabling Structured Multi-Antenna Fusion

Another significant implication relates to the potential for integrating information from multiple Wi-Fi antennas. The research points out that 'because rules are defined over discrete latent variables, antenna-specific rule sets can in principle be combined at the symbolic level'. This is a powerful capability because it enables 'structured multi-antenna fusion without retraining the encoder'. In many multi-antenna HAR systems, combining data from different antennas often involves concatenating raw signals or features, which then necessitates retraining the entire predictive model. This can be computationally expensive and time-consuming, especially when hardware configurations change or new antennas are added.

The CHARL-TRE approach suggests a more flexible and efficient method. If each antenna's CSI data can be processed by its own encoder to generate discrete latent rules, these *rules* can then be combined symbolically. This 'symbolic level' combination means that new information from additional antennas or different sensing geometries might be integrated by simply adding or modifying rules, rather than by entirely relearning the underlying representations. This could lead to more scalable, adaptable, and robust multi-antenna HAR systems, where the integration logic is as transparent as the activity classification itself.

What's Next: Advancing Explainable Wireless HAR

While the source material does not explicitly outline a 'What's Next' section in the future research steps, the implications inherently suggest directions for further investigation and application. The success of CHARL-TRE in providing causally interpretable, symbolically controllable, and direct operation on high-dimensional raw signals opens avenues for expanding its utility and robustness.

Expanding the Scope of Interpretable HAR

Future work could explore the application of CHARL-TRE to a wider range of human activities, including more complex or subtle movements that require fine-grained temporal and causal understanding. The framework's ability to preserve explicit temporal and causal structure could be leveraged to differentiate between activities that are superficially similar but differ in their underlying sequence of movements or physiological impacts, for which a 'black-box' model might struggle to provide distinguishing rationale. This could lead to HAR systems capable of detecting not just *what* happened, but also *how* it happened, which is crucial for applications such as rehabilitation monitoring or athletic performance analysis.

Refining and Validating Multi-Antenna Fusion Strategies

The proposed 'structured multi-antenna fusion' at the symbolic level offers a compelling direction. Further research would likely involve empirically validating this concept across diverse real-world scenarios, exploring different strategies for combining these antenna-specific rule sets. This could include investigating different aggregation methods for symbolic rules or developing formal frameworks for conflict resolution when rules from various antennas provide contradictory insights. Demonstrating the practical benefits of this approach—such as improved accuracy, robustness to antenna placement variations, or reduced computational overhead for system expansion—would be a significant next step. This would solidify CHARL-TRE's position as a scalable and adaptable solution for complex wireless HAR deployments.

Key Findings