Enhancing Binary Encoded Crime Linkage Analysis Using Siamese Network¶
Conference: AAAI 2026 arXiv: 2511.07651 Code: https://github.com/AlberTgarY/CrimeLinkageSiamese Area: Interpretability Keywords: Crime Linkage Analysis, Siamese Network, Autoencoder, Geo-temporal Feature Fusion, ViCLAS Database
TL;DR¶
This paper proposes a Siamese Autoencoder-based crime linkage analysis framework that integrates geo-temporal features at the decoder stage and employs a domain expert-driven dimensionality reduction strategy. Evaluated on the real-world ViCLAS database from the UK National Crime Agency (NCA), the method achieves up to 9% AUC improvement, providing an effective machine learning solution for high-dimensional sparse binary-encoded crime data.
Background & Motivation¶
Problem Setting¶
Crime Linkage (CL) is a critical task in modern law enforcement, aiming to identify serial crimes by analyzing offenders' Modus Operandi (MO). Accurate CL helps optimize investigative resource allocation and enhance public safety. Its theoretical foundations are behavioral consistency (the same offender exhibits similar behavior across crimes) and behavioral distinctiveness (behavioral patterns of different offenders are distinguishable).
Limitations of Prior Work¶
Limitations of traditional statistical methods: Logistic regression assumes linear feature relationships, and decision trees impose rigid hierarchical structures—neither is suited to capturing nonlinear associations in criminal behavior.
High-dimensional sparse data challenges: Real-world crime databases (e.g., ViCLAS) contain 446-dimensional binary-encoded features, with approximately 91% zero values, resulting in extremely sparse signals.
Underutilization of geo-temporal information: Existing methods either ignore geo-temporal information or naively concatenate it to the input layer, where the 2-dimensional geo-temporal signal accounts for less than 1% of the 217–446 dimensional behavioral features and is severely diluted.
Limited dataset scale: Prior research has focused primarily on small-scale, geographically restricted datasets, without validation on large-scale real-world databases such as ViCLAS.
Core Innovation Motivation¶
- Signal dilution problem: Geo-temporal features are shifted from the input layer to the decoder stage for fusion, allowing them to modulate the latent representation after behavioral abstraction is complete, avoiding submersion by high-dimensional behavioral features.
- Domain knowledge-driven dimensionality reduction: Five data reduction mapping strategies were designed in collaboration with NCA domain experts, preserving behavioral semantics while reducing dimensionality.
Method¶
Overall Architecture¶
The framework comprises three core components: 1. Siamese Autoencoder network: Dual branches with shared weights; encoder compression + decoder reconstruction + geo-temporal fusion at the decoder stage 2. Joint contrastive and reconstruction loss: Pulling linked crimes together and pushing unlinked crimes apart 3. Inference pipeline: Similarity probability scores computed from latent space distances
Key Designs¶
1. Siamese Autoencoder Network¶
- Encoder: Two linear layers + ReLU (446→128→8), compressing input to an 8-dimensional latent space
- Decoder: Mirror structure (8→128→446), reconstructing the original feature space
- Geo-temporal fusion: After the first decoder layer output, log-transformed spatial-temporal features are mapped via a linear layer and fused additively
- Parameter count: 21,740 parameters (vs. 22,981 for Naive Siamese)—fewer parameters, better performance
- Design Motivation:
- The reconstruction constraint of the autoencoder ensures the latent representation retains structural information
- The 8-dimensional bottleneck forces the network to learn the most compact behavioral representation
2. Decoder-Stage Geo-temporal Fusion¶
- Core Idea: Geo-temporal data inherently reflects pairwise relationships (distance and time interval between two crimes); fusing it after encoding individual behavioral features is more consistent with investigative logic
- Implementation:
- Log-transform spatial distance and temporal interval
- Map 2D geo-temporal features to 128 dimensions via a linear layer
- Additively fuse with the first decoder layer output
- Advantage over alternatives:
- Input-layer concatenation: 2D signal in 446D accounts for <1%, nearly no effect
- Decoder fusion: Introduced after behavioral abstraction is complete, yielding significant signal amplification
- Experimental Validation: Consistently improves AUC by 0.86–3.29% across network variants (Table 4)
3. Domain Expert-Driven Dimensionality Reduction¶
Five reduction strategies were designed (Table 1):
| Strategy | Remaining Features | Reduction Rate | Designer |
|---|---|---|---|
| No Map | 446 (original) | 0% | — |
| Map 1 | 282 | 36.8% | NCA analyst (20+ years experience) |
| Map 2 | 384 | 13.9% | Map 1 + forensic psychologist consultation |
| Map 3 | 266 | 40.4% | Hybrid of Map 1 and Map 2 |
| Map 4 | 217 | 51.3% | Forensic psychology expert |
| Map 5 | 286 | 35.9% | Refinement of Map 4 (reduced abstraction) |
- Reduction method: Semantically similar binary features are merged into more abstract categories (e.g., "shopping center parking lot" and "stadium parking lot" merged into "parking lot")
- Design Motivation: Analysts in practice typically link crimes via thematic matching rather than exact behavioral matching
Loss & Training¶
Joint loss: \(\mathcal{L} = \alpha \mathcal{L}_{\text{contrast}} + \beta \mathcal{L}_{\text{recon}}\), with \(\alpha=1.0\), \(\beta=0.2\)
Contrastive loss: Uses a hybrid Euclidean–Manhattan distance metric $\(\mathcal{L}_{\text{contrast}} = \mathbb{E}[y \cdot d^2 + (1-y) \cdot \max(m-d, 0)^2]\)$ where \(y\) indicates linkage (1 = linked), and \(m=5\) is the margin parameter.
Reconstruction loss: Based on cosine similarity $\(\mathcal{L}_{\text{recon}} = \mathbb{E}\left[\frac{v_1^\top \hat{v_1}}{\|v_1\| \|\hat{v_1}\|} + \frac{v_2^\top \hat{v_2}}{\|v_2\| \|\hat{v_2}\|}\right]\)$
Inference stage: \(S_{ij} = \exp(-D_{ij}/\beta)\), \(\beta = m/1.5\), converting latent space distances to probability scores in \((0, 1]\).
Training details: Adam optimizer, lr=0.001, batch_size=128, 2 epochs, Cosine Annealing learning rate schedule, 5-fold cross-validation.
Key Experimental Results¶
Datasets¶
- Single Victim-Offender-Scene Series: 1,482 cases / 493 series / recorded in 2014; no geo-temporal data
- Multiple Victim-Offender-Scene Series: 22,282 crimes (1990–2021), 446 features, 11,970 used for analysis (the largest ViCLAS dataset in published research)
Main Results¶
Small dataset (Single Victim-Offender-Scene):
| Method | AUC | TP Fixed FP |
|---|---|---|
| Ours | 85 ± 1.98 | 77.73 ± 4.32 |
| Logistic Regression | 86 ± 2.14 | 77.19 ± 5.98 |
| PCA | 82 ± 4.02 | 64.97 ± 4.00 |
Large dataset (Multiple Victim-Offender-Scene, using Map 5):
| Method | AUC | TP Fixed FP | AUPRC |
|---|---|---|---|
| Ours (map 5) | 84 ± 2.86 | 79.38 ± 2.56 | 15.43 |
| Naive Siamese (map 5) | 83 ± 2.72 | 76.20 ± 2.69 | 15.09 |
| Logistic Regression | 75 ± 2.97 | 70.43 ± 2.12 | 10.24 |
| Ours (no map) | 77 ± 2.11 | 68.31 ± 1.92 | 13.32 |
Relative gain over LR: +12.0% (AUC), +12.71% (TPFP), +50.68% (AUPRC).
Ablation Study¶
Impact of architectural choices (RQ3, Table 4 excerpt):
| Configuration | AUC (%) | Note |
|---|---|---|
| MLP, no skip, 2+2, Decoder | 77.29 | Best configuration |
| MLP, no skip, 2+2, Concat | 76.43 | Decoder fusion +0.86% |
| MLP, skip, 2+2, Decoder | 67.07 | Skip connections harmful |
| 1D CNN, no skip, 2+2, Decoder | 61.74 | MLP significantly better than CNN |
| SIREN, no skip, 2+2, Decoder | 58.28 | Periodic activation unsuitable |
| MLP, no skip, 1+1 | 52.49 | Too shallow |
| MLP, no skip, 4+4 | 63.57 | Too deep, overfitting |
Key findings: - 2+2 layer depth is optimal; both shallower and deeper architectures degrade performance - Omitting skip connections is beneficial (+6.55%), likely because direct feature propagation interferes with the abstraction of subtle crime patterns
Key Findings¶
- Decoder fusion consistently outperforms input-layer concatenation: MLP +0.86%, 1D CNN +3.29%, SIREN +3.09%
- Moderate dimensionality reduction improves performance: Map 5 (35.9% reduction) achieves the best AUC of 84%; both excessive and insufficient abstraction underperform
- Temporal OOD challenge: Post-COVID data from 2021–2025 shows notable performance degradation, necessitating periodic retraining
- System can reduce manual review by 80%: Substantially reduces false positives while retaining over half of true crime linkages
Highlights & Insights¶
- Real-world validation: The UK NCA's ViCLAS database—the largest publicly studied sexual offense dataset to date—is used instead of synthetic data, enhancing practical significance.
- Generalizability of decoder fusion: When auxiliary information dimensionality is far lower than that of primary features, the late fusion strategy merits broader adoption in other domains.
- Domain expert-driven feature engineering: The five reduction strategies were designed by experts from different professional backgrounds, exemplifying best practices in combining ML with domain knowledge.
- Thorough ethical consideration: Detailed deployment safeguards, bias auditing plans, and human-in-the-loop design are provided.
Limitations & Future Work¶
- Information loss from binary encoding: Reducing complex criminal behavior to 0/1 encoding may discard important frequency and intensity information.
- Temporal distribution shift: Model performance degrades notably on post-2021 data, requiring a continuous retraining mechanism.
- Data confidentiality: ViCLAS data is not publicly available, limiting reproducibility.
- 8-dimensional bottleneck only: The bottleneck may be overly compact; analysis of the performance impact of varying bottleneck dimensions is limited.
- Limited to sexual offenses: Generalizability to other crime types such as burglary and robbery has not been evaluated.
Related Work & Insights¶
- Solomon et al. (2020): Applied Siamese networks to burglary linkage (40-dimensional TF-IDF); the present work extends this to 446-dimensional binary encoding with added reconstruction constraints.
- Tonkin et al. (2017): Source of the hybrid Euclidean–Manhattan distance metric.
- Insight: For high-dimensional sparse data combined with low-dimensional auxiliary information, signal dilution is an underappreciated problem; late fusion strategies warrant systematic investigation.
Rating¶
- Novelty: ⭐⭐⭐ — Core techniques (Siamese + autoencoder) are not novel; primary contributions lie in the application and fusion strategy
- Experimental Thoroughness: ⭐⭐⭐⭐ — Real-world data, extensive ablations, temporal OOD testing; lacks evaluation on other crime types
- Writing Quality: ⭐⭐⭐⭐ — Well-structured, research-question-driven, with thorough ethical discussion
- Value: ⭐⭐⭐⭐ — Direct practical value for law enforcement; methodological contribution is moderate