PaperlayerView exact revision on arXiv

Source-linked AI summary

Human Activity Recognition via Ultra-Wideband Data: A Framework for Dimensionality Reduction, Pattern Discovery, and Predictive Modeling

Nahid Sahel Gozin, Reza Sedaghat, Prathap Siddavaatam

arXiv:2609.32008v1cs.LG

TL;DR

UWB activity recognition is challenged by data complexity and by reduction methods that may be difficult to interpret or too resource-intensive for real-time use. The paper combines two dimensionality-reduction approaches with rule mining and classification, reporting high accuracy across six activities.

  • Problem

    UWB HAR still faces difficulty extracting complex features, interpreting model decisions, and using methods in real time, while traditional reduction can be hard to interpret or resource-intensive.

    Source: §I · p. 1, §II · p. 2

  • Method

    The framework combines CPE-IPCA and PPSA dimensionality reduction with association-rule features and Random Forest or Vector Space Model classification.

    Source: §I · p. 1, §V · p. 11

  • Results

    CPE-IPCA achieved 100% accuracy with RF and F1-scores of 0.99–1.00 with VSM, while PPSA-NLD attained 99% accuracy with RF and perfect precision, recall, and F1 with VSM across six activities.

    Source: §VI · p. 12

  • Takeaways & Limitations

    The authors describe the framework as a highly accurate, interpretable, and scalable approach for real-time HAR in smart environments, healthcare, and assistive technologies.

    Source: §VI · p. 12

Abstract

from arXiv · show

Recent advances in sensor technology have enabled more effective human activity recognition (HAR), particularly in real-time systems with limited computational resources. However, Ultra-Wideband (UWB) radar data remain challenging due to high dimensionality, noise, complexity, and nonlinear characteristics. This research proposes a framework to efficiently reduce data size, uncover significant patterns, and classify six activity types (Standff, Liedown, Noactivity, Sit, Stand, and Walk) from UWB signals with high accuracy. Two novel dimensionality reduction techniques are introduced in this paper. The first, Clustered Polynomial Expansion with Incremental PCA (CPE-IPCA), combines clustering and polynomial feature expansion with Incremental PCA, preserving 100% of the variance in only 50 components. The second, Post-PCA Standardization Approach (PPSA), standardizes data after PCA and retains 99.1% of the variance in 80 components, achieving superior compression and computational efficiency compared to conventional nonlinear methods. Frequent patterns are identified using Apriori and FP-Growth, which are then classified with Random Forest and a Vector Space Model (VSM). The framework achieves 100% accuracy with Random Forest on CPE-IPCA and 99% on PPSA, while VSM attains 100% precision, recall, and F1 on PPSA and near-perfect performance on CPE-IPCA (precision 1.00, recall 0.98-1.00, F1 0.99-1.00), demonstrating a fast, interpretable, and robust HAR system suitable for healthcare, assisted living, and smart environments.

I. INTRODUCTION

The introduction presents UWB activity recognition as challenging because its data are high-dimensional, noisy, and nonlinear, while existing reduction methods may be difficult to interpret or costly for real-time use. It proposes two reduction methods and connects dimensionality reduction, rule mining, and classification against related work that still faces interpretability and real-world deployment challenges.

I. INTRODUCTION

  • UWB data are high-dimensional, noisy, and nonlinear, while traditional reduction methods may produce hard-to-interpret features or demand too many resources for real-time use.The introduction frames these constraints as especially relevant to systems with limited computing power.Source: §I · p. 1
  • CPE-IPCA combines feature grouping, polynomial expansion, and Incremental PCA, while PPSA applies normalization after PCA to make results easier to understand without losing key details.CPE-IPCA is presented as capturing nonlinear patterns; PPSA as improving interpretability.Source: §I · p. 1
  • The framework links dimensionality reduction with Apriori and FP-Growth rule discovery and Random Forest and Vector Space Model classifiers to support accurate and interpretable activity recognition.The authors say this combination makes the reasoning behind classification decisions easier to understand.Source: §I · p. 1

II. RELATED WORK

  • Prior HAR studies report strong classifier results, but Random Forest methods can be slower to train, harder to interpret, and less suited to real-time use.A cited Random Forest study reported 98% test accuracy and 95.5% cross-validation accuracy, alongside training-time and interpretability concerns.Source: §II · p. 2
  • Related work identifies continued difficulty extracting complex UWB features, understanding model decisions, and recognizing activities reliably in real-world, multi-sensor settings.UWB research has often focused on laboratory settings and single actions, while real-world activities can vary in duration and overlap.Source: §II · pp. 1–2

III. SYSTEM MODEL AND PROPOSED FRAMEWORK

The framework uses UWB signals to recognize six activities, applying dimensionality reduction before association-rule mining and classification. CPE-IPCA combines correlation-based feature clustering, polynomial expansion, and Incremental PCA to produce compact representations.

A. UWB Data Acquisition and Preprocessing

  • UWB measurements from synchronized anchor nodes capture spatiotemporal information in controlled indoor settings with multiple participants performing diverse movements.Preprocessing removes missing and duplicate entries, separates complex CIR values into real and imaginary components, and labels samples with six activity classes.Source: A. UWB Data Acquisition and… · p. 2

B. Dimensionality Reduction Techniques

  • CPE-IPCA captures nonlinear dependencies, while PPSA standardizes PCA-projected components to improve stability and interpretability.Both methods are described as producing compact, information-rich representations that improve computational efficiency and classification accuracy.Source: B. Dimensionality Reduction Techniques · p. 3

1) Clustered Polynomial Expansion with Incremental PCA:

  • CPE-IPCA integrates correlation-based clustering, cluster representatives, polynomial expansion, Incremental PCA, and post-IPCA standardization into a hierarchical reduction framework.Polynomial expansion is applied to cluster representatives, capturing nonlinear inter-cluster relationships while limiting dimensional growth; Incremental PCA supports memory-efficient computation.Source: 1) Clustered Polynomial Expansion with… · pp. 3–4
  • Correlation-based clustering groups features with similar profiles, and each cluster’s representative features are polynomially expanded to capture nonlinear dependencies.Applying expansion to cluster representatives preserves inter-cluster relationships while reducing dimensional growth and computational cost compared with conventional PCA and nonlinear variants.Source: 1) Clustered Polynomial Expansion with… · pp. 3–4
  • CPE-IPCA is described as faster than Kernel PCA, Isomap, and t-SNE for large-scale, high-dimensional datasets such as UWB signals.The comparison passage gives Kernel PCA and Isomap complexity as O(m^3) and t-SNE as O(m^2).Source: 1) Clustered Polynomial Expansion with… · pp. 4–5

2) Post-PCA Standardization Approach (PPSA) Overview:

PPSA enhances PCA for high-dimensional UWB data by standardizing projected components. It combines centering, PCA projection, and post-projection standardization in a reduced feature matrix.

2) Post-PCA Standardization Approach (PPSA) Overview:

  • PPSA centers the original data, selects the top r PCA components, projects the centered data, then standardizes each projected component.The returned matrix is the standardized reduced feature matrix X′.Source: 2) Post-PCA Standardization Approach… · p. 5
  • PPSA standardizes each PCA-projected component to zero mean and unit variance, so all components contribute equally.The approach is described as improving interpretability and numerical stability.Source: 2) Post-PCA Standardization Approach… · p. 5
  • For m ≫ n and r ≪ n, standardization O(mr) is negligible and PCA complexity O(min(mn^2, m^2n)) = O(mn^2) dominates.The PCA cost selects the more efficient path according to the relative sizes of m and n.Source: 2) Post-PCA Standardization Approach… · p. 5

C. Pattern Mining via Association Rule Mining

Association Rule Mining identifies dependencies among reduced-dimensional features using Apriori and FP-Growth. The rules are assessed by support, confidence, and lift, with thresholds determining which patterns are retained.

C. Pattern Mining via Association Rule Mining

  • Apriori iteratively expands frequent itemsets, while FP-Growth uses an FP-tree to avoid repeated candidate generation.Association Rule Mining was selected for interpretability and its ability to uncover feature relationships not revealed through dimensionality reduction alone.Source: C. Pattern Mining via Association Rule… · p. 5
  • Each rule A ⇒ C links subsets of co-occurring features, and support, confidence, and lift assess its significance.Support measures statistical relevance, confidence predictive strength, and lift non-random association; rules exceeding defined thresholds are retained.Source: C. Pattern Mining via Association Rule… · pp. 5–6

D. Human Activity Classification

The classification phase applies Random Forest and VSM to feature sets refined through dimensionality reduction and pattern discovery. The framework is reported to achieve robust, computationally efficient HAR performance across diverse activity scenarios.

D. Human Activity Classification

  • Random Forest and VSM classify feature sets refined through dimensionality reduction and pattern discovery using a stratified 70% training and 30% testing split.Performance is evaluated with precision, recall, F1-score, and accuracy.Source: D. Human Activity Classification · p. 6
  • The framework integrates dimensionality reduction, pattern mining, and classification to achieve robust, computationally efficient HAR performance across diverse activity scenarios.The framework uses Random Forest and VSM on refined feature sets; the dataset is stratified into 70% training and 30% testing subsets.Source: D. Human Activity Classification · p. 6

IV. EXPERIMENTAL RESULTS

The experimental evaluation examines dataset characteristics and linearity, then evaluates dimensionality reduction, pattern discovery, and classification. The reported results demonstrate effectiveness and robustness across all stages of the HAR pipeline.

IV. EXPERIMENTAL RESULTS

  • The evaluation covers dataset characteristics, linearity, dimensionality reduction, pattern discovery, and classification, with effectiveness and robustness reported across all HAR pipeline stages.The analysis proceeds from dataset and linearity assessment to evaluations of the three pipeline stages.Source: §IV · p. 6

A. Data Characterization and Linearity Analysis

The UWB dataset contains 17,063 observations and 2,040 standardized numeric features. Regression analysis indicates that its feature relationships are predominantly nonlinear.

A. Data Characterization and Linearity Analysis

B. Proposed Dimensionality Reduction Techniques

The section compares PCA, CPE-IPCA, and PPSA on a nonlinear, high-dimensional dataset, then examines frequent patterns and threshold effects in the reduced PPSA data.

1) Traditional Principal Component Analysis:

2) Clustered Polynomial Expansion with Incremental PCA:

C. Association Rule Mining Applied to the Post-PCA Dataset

  • PPSA's 9,913-row binary dataset yielded top itemsets with supports from 0.291 to 0.356, and FP-Growth produced nearly identical results faster.[PC1], [PC2], [PC4], and [PC3] had the highest supports.Source: C. Association Rule Mining Applied to… · p. 7

1) Frequent Rule Patterns and Key Components:

2) Rule Counts and Metrics by Thresholds (Post-PCA):

3) Network Visualization of Apriori and FP-Growth:

The section reports association patterns in the CPE-IPCA dataset and compares rule-network structure and search strategies across Apriori and FP-Growth.

3) Network Visualization of Apriori and FP-Growth:

  • The Figure 6 network contains 2,892 nodes and 45,150 edges; both algorithms generated the same rules, with strongly connected regions and consistent patterns.Apriori uses breadth-first search, while FP-Growth uses depth-first search, producing distinct layouts.Source: 3) Network Visualization of Apriori… · p. 7, Figure 6 · p. 9

D. Association Rule Mining Applied to CPE-IPCA Dataset

  • CPE-IPCA's top 20 itemsets had supports of 0.239–0.272, with [PC50] highest at 0.272; FP-Growth produced nearly identical results faster.The analysis used 10,733 records and minimum support of 0.0181.Source: D. Association Rule Mining Applied to… · pp. 7, 9

1) CPE-IPCA Key Rules and Feature Correlations:

  • The rule PC47, PC9 → PC1, PC48, PC7 had the highest lift (9.35), while noactivity → PC50 had the highest support (0.12).Other noactivity-related rules involving PCs 39, 12, and 15 had support around 0.11.Source: 1) CPE-IPCA Key Rules and Feature… · p. 9

2) Rule Extraction and Threshold Effects:

  • At support 0.05, confidence 0.2, and lift 0.4, rule extraction balanced reliability and coverage; meaningful rules appeared only at support 0.05.Raising confidence from 0.05 to 0.4 raised average confidence from 0.22 to 0.46 while lift stayed at 1.42.Source: 2) Rule Extraction and Threshold Effects · p. 9, Table V · p. 9

3) Network Structure and Algorithmic Differences:

  • The Figure 9 CPE-IPCA network has 1,695 nodes and 59,166 edges, and Apriori and FP-Growth produce similar networks with structural differences.The differences are associated with their breadth-first versus depth-first search strategies.Source: 3) Network Structure and Algorithmic… · p. 9

E. Activity Classification Using Rule-Based Features

The paper evaluates Random Forest and VSM on rule-based features from PPSA and CPE-IPCA. Both classifiers perform strongly, and the authors describe the framework as interpretable.

E. Activity Classification Using Rule-Based Features

1) Activity Classification via Random Forest:

2) Activity Classification via Vector Space Model:

V. DISCUSSION

  • The study concludes that rule-based features provide an effective, interpretable framework for recognizing human activities from UWB signals.Source: §V · p. 11

A. Implications of Data Structure and Linearity

The framework combines dimensionality reduction, association-rule mining, and classification to address nonlinear UWB data and support interpretable activity recognition. Its reported results include high classification performance using compact rule-based features.

A. Implications of Data Structure and Linearity

  • A large Linear Regression–Random Forest performance gap indicates strong nonlinear relationships, motivating nonlinear reduction such as CPE-IPCA; PPSA provides a stable linear baseline for feature analysis and classification.Preprocessing simplified and standardized the signals, and variance filtering highlighted informative features.Source: A. Implications of Data Structure and… · p. 11

B. Impact of Dimensionality Reduction Techniques

  • CPE-IPCA combines correlation-based clustering, polynomial expansion, and Incremental PCA for full variance retention with fewer components; PPSA yields zero-centered, unit-variance components but captures nonlinear dependencies less effectively.The passages describe CPE-IPCA as computationally efficient and suitable for real-time environments.Source: B. Impact of Dimensionality Reduction… · p. 11

C. Association Rule Mining: Interpretability and Insights

  • For PPSA, Apriori and FP-Growth found PC43, PC32, and PC7 in high-lift rules, while PC1 dominated high-support rules; FP-Growth was faster and equally informative.The network visualization showed structured relationships among components.Source: C. Association Rule Mining:… · p. 11
  • CPE-IPCA’s high-lift rules included PC47, PC9 → PC1, PC48, PC7, while PCs 50, 39, 33, and 19 were linked to inactivity states.Network visualization showed structural consistency, and Table V illustrated how thresholds affected rule precision.Source: C. Association Rule Mining:… · p. 11

D. Discussion: Activity Classification Using Rule-Based Features

  • Rule-based binary features use only 10 features per sample; CPE-IPCA improves discrimination for low-support Liedown, while sparse rules support rapid, interpretable VSM decisions.RF performed strongly on both reduced datasets, with confusion matrices showing negligible misclassifications.Source: D. Discussion: Activity Classification… · pp. 11–12

E. Comparison with Existing Methods

  • Compared with fall-detection and PCA/t-SNE/UMAP approaches, the authors report coverage of six activities; CPE-IPCA supports incremental learning, while PPSA maintains interpretability and low computational cost.The compared fall-detection method distinguishes falls from basic activities and uses a small number of participants; the PCA/t-SNE approach is described as computationally costly.Source: E. Comparison with Existing Methods · p. 12

VI. CONCLUSION

  • The study applies CPE-IPCA and PPSA-NLD to 17,063 observations and 2,040 attributes, then uses Apriori and FP-Growth patterns with RF and VSM classifiers.CPE-IPCA combines feature clustering, polynomial expansion, and incremental PCA; PPSA-NLD standardizes PCA components.Source: §VI · p. 12
  • 100% accuracy with RF on CPE-IPCA and 99% with RF on PPSA-NLD; VSM attained F1-scores of 0.99–1.00 on CPE-IPCA and perfect precision, recall, and F1 on PPSA-NLD.The framework classified six activities.Source: §VI · p. 12
Loading 2609.32008v1…