Enhancing Binary Encoded Crime Linkage Analysis Using Siamese Network

AI summary

Overview

Research area: Machine learning for behavioural crime linkage analysis, applied to real-world law-enforcement data from the UK's Violent Crime Linkage Analysis System (ViCLAS).

Technical level: Intermediate. The framework is a Siamese Autoencoder, and the evaluation relies on AUC, TPFP and AUPRC metrics; readers with basic familiarity with neural networks and classification metrics will follow it, and the paper explains its design choices in accessible terms.

Scope: The paper proposes and evaluates a Siamese Autoencoder, with decoder-stage geographic-temporal integration and expert-designed data reduction, for predicting whether pairs of offences were committed by the same offender.

What This Paper Is About

Crime linkage tries to determine whether two or more offences were committed by the same person, based on the offender's modus operandi (MO). Real crime databases are high-dimensional, sparse and geographically and temporally heterogeneous, and traditional statistical methods struggle with these conditions.

The paper introduces a Siamese Autoencoder that learns compact latent representations from binary-encoded ViCLAS features, and adds geographic and temporal information at the decoder stage rather than at the input, where it would otherwise be drowned out by hundreds of behavioural features. The goal is to produce a probability score for any pair of cases indicating whether they are linked.

Key Contributions

1. A novel application of a Siamese Autoencoder to crime linkage, evaluated on ViCLAS, described by the authors as the largest real-world dataset on serious sexual offences analysed in published research.
2. Decoder-stage geographic-temporal integration, which fuses 2-dimensional distance/time signals after behavioural abstraction instead of concatenating them with the input features, mitigating signal dilution.
3. An exploration of domain-expert-defined data reduction strategies (five mappings, plus the unmapped baseline) and their effect on model performance on sparse, high-dimensional data.
4. Insights into dataset properties and network sensitivity, including findings that shallower architectures and no skip connections outperform deeper or skip-connected alternatives on this data.

Main Findings

• Single Victim-Offender-Scene Series results: On the 1,482 cases across 493 series, the proposed method achieved a mean AUC of 85% (std 1.98) and 77.73% TP Fixed FP (std 4.32), compared with logistic regression at 86% AUC (std 2.14) and 77.19% TP Fixed FP (std 5.98), and PCA at 82% AUC (std 4.02) and 64.97% TP Fixed FP (std 4.00). AUPRC is not reported for this dataset.
• Multiple Victim-Offender-Scene Series results: On the main dataset, the model without reduction achieved 77% AUC (std 2.11), 68.31% TP Fixed FP (std 1.92) and 13.32% AUPRC. With Map 5 it reached 84% AUC (std 2.86), 79.38% TP Fixed FP (std 2.56) and 15.43% AUPRC.
• Improvements over baselines with Map 5: Relative improvements over logistic regression were 12.0%, 12.71% and 50.68% for AUC, TP Fixed FP and AUPRC respectively; over the Naive Siamese baseline they were 6.67%, 8.37% and 9.36%. The abstract reports AUC improvements of up to 9% over traditional methods.
• Two comparable-parameter networks: The proposed network has 21,740 parameters versus 22,981 for the Naive Siamese baseline, yet performed consistently better across metrics.
• Data reduction matters: Across the five reduction strategies, AUC exhibited a range of 8.00 (std 3.54), TP Fixed FP a range of 10.26 (std 4.07) and AUPRC a range of 1.46 (std 0.66). On average, the method improved by 2.30%, 3.02% and 3.48% in AUC, TP Fixed FP and AUPRC over the Naive Siamese baseline. Map 5 was identified as optimal, achieving 84% AUC; the paper's text states it reduces features by 44.8%, while Table 1 lists Map 5 at 286 features and a 35.9% reduction rate.
• Architecture comparisons: The MLP design achieved a mean AUC of 65.30%, outperforming 1D CNN (57.10%) and SIREN (52.16%). Omitting skip connections yielded a mean AUC of 63.03% versus 56.48% with skip connections, a difference of 6.55%.
• Depth analysis: A 2-encoder/2-decoder configuration achieved the highest AUC of 77.29%; 1+1 reached 52.49%, 3+3 reached 70.85% and 4+4 reached 63.57%.
• Geographic-temporal integration effect: Decoder-level integration raised MLP AUC from 76.43% to 77.29% (an absolute increase of 0.86%). For 1D CNN without skip connections it improved from 58.45% to 61.74% (+3.29%), and for SIREN from 55.19% to 58.28% (+3.09%). The authors report consistent improvements of 0.86–3.29% AUC across network variants.
• Out-of-time generalisation: On 1,165 solved cases from 2021 to 2025, Map 1 achieved the highest recall (77.94%) with 53 of 68 true positives, while Map 5 produced only 97,013 false positives out of the 678,030 possible case pairs and maintained 42.65% recall. Performance was significantly lower than in 5-fold cross-validation, but the system reduced manual review workload by up to 80% while preserving over half of genuine connections.
• Bias assessment: Comparing victim and suspect demographics (age and profession) between unit reports and ViCLAS revealed no significant differences for the available variables.
• Sparsity: The 446 binary dimensions contain approximately 91% zero values.

Methodology in Plain English

The researchers treated crime linkage as a pairwise comparison problem. Each incident is a binary vector of 446 features (1 = attribute observed, 0 = unobserved or unrecorded), and the goal is a function that takes two incidents plus their geographic distance and time interval and returns a probability between 0 and 1 that they are linked.

The model is a Siamese Autoencoder: two identical sub-networks process the two cases of a pair in parallel. Each encoder compresses the 446 input features through two linear layers with ReLU activations (446→128→8), and each decoder mirrors this (8→128→446) to reconstruct the original features. Geographic and temporal data are logarithmically transformed, passed through a linear layer, and added to the output of the first decoder layer, so they modulate the learned behavioural representation rather than competing with it at the input.

Training combines two losses with weights α = 1.0 for contrastive loss and β = 0.2 for reconstruction loss. The contrastive loss clusters cases by the same offender and separates unlinked ones using a hybrid Euclidean-Manhattan distance, with a margin of 5 chosen empirically from values between 1 and 10. The reconstruction loss uses cosine similarity to keep the latent codes informative.

For inference, the Euclidean distance between two cases' latent codes is converted to a bounded similarity score with an exponential decay function, where the scaling parameter is set to the margin divided by 1.5.

Data preparation involved merging duplicate entries within Excel sheets (encoding 1 if any part of the incident showed the attribute) and unifying overlapping categorical labels, tagging each dimension as behavioural, contextual or both (177 behavioural, 158 contextual, 11 both). Five expert-driven reduction maps consolidated similar features into broader categories: Map 1 was created by an NCA analyst with 20+ years of experience, Map 2 refined Map 1 with forensic psychologists, Map 3 merged Maps 1 and 2, Map 4 was developed by forensic psychology experts from crime linkage literature, and Map 5 was a less abstract version of Map 4.

Training used PyTorch on an Intel 12700k CPU for 2 epochs with batch size 128, taking 6 hours in total, using the Adam optimizer (β1 = 0.9, β2 = 0.999), a learning rate of 0.001, Cosine Annealing learning rate scheduling, Gaussian noise added to inputs, and 5-fold cross-validation.

Why This Matters

Impact on research: The paper shows that architecture and preprocessing choices that normally help in machine learning (deeper networks, skip connections, input-level feature concatenation) can actually hurt on sparse binary behavioural crime data. It also contributes a reproducible expert-driven feature reduction methodology and the largest published analysis of ViCLAS for linkage research.

Real-world applications:

• Triage of offence series: Screening large numbers of case pairs to surface likely linked offences for analyst review, with the out-of-time test suggesting manual review workload could be reduced by up to 80%.
• Prioritising investigative resources: Ranked similarity lists let analysts and managers allocate effort toward the most promising candidate series.
• Supporting justice outcomes: Linking previously unconnected cases can offer closure for victims and accountability for offenders, and can help detect serial offenders earlier.
• Cross-jurisdictional adaptation: The binary encoding is deliberately language-agnostic, so the pipeline could be adapted to other legal systems, with local experts developing culturally appropriate feature consolidations.

Industry relevance: Law enforcement agencies and national crime databases in the UK and comparable systems in Europe, Canada and New Zealand could adopt the approach. The paper explicitly frames deployment around human-in-the-loop review, routine bias audits, transparent evaluation and periodic retraining, aligned with the National Police Chiefs' Council Covenant for Using AI in Policing.

Future Directions

• Extending the method to additional crime types to assess generalisability across offence categories.
• Incorporating natural-language offence descriptions to recover complex patterns lost in binary encoding.
• Evaluating the approach in audited operational settings, examining both effectiveness and ethical implications for real-world triage.
• Investigating periodic retraining schedules to counter temporal distribution shift, given the performance degradation on post-2021 data coinciding with COVID-19 aftermath.
• Addressing the acknowledged gap in research on behavioural differences across victim groups and offender demographics, for which no studies currently exist.

Target Audience

Researchers working on crime linkage, behavioural crime analysis and applied machine learning for public safety; data scientists and analysts in law enforcement agencies considering ML support tools; forensic psychologists and criminologists interested in how expert domain knowledge can be encoded as feature reduction; and machine learning practitioners interested in Siamese architectures, sparse high-dimensional binary data, and the design trade-offs of feature fusion and network depth.

Authors’ abstract

Effective crime linkage analysis is crucial for identifying serial offenders and enhancing public safety. To address limitations of traditional crime linkage methods in handling high-dimensional, sparse, and heterogeneous data, we propose a Siamese Autoencoder framework that learns meaningful latent representations and uncovers correlations in complex crime data. Using data from the Violent Crime Linkage Analysis System (ViCLAS), maintained by the Serious Crime Analysis Section of the UK's National Crime Agency, our approach mitigates signal dilution in sparse feature spaces by integrating geographic-temporal features at the decoder stage. This design amplifies behavioral representations rather than allowing them to be overshadowed at the input level, yielding consistent improvements across multiple evaluation metrics. We further analyze how different domain-informed data reduction strategies influence model performance, providing practical guidance for preprocessing in crime linkage contexts. Our results show that advanced machine learning approaches can substantially enhance linkage accuracy, improving AUC by up to 9% over traditional methods while offering interpretable insights to support investigative decision-making.