ViCLAS: Violent Crime Linkage System

• ViCLAS is a system that encodes violent crime data into 446-dimensional binary vectors to systematically link cases based on behavioral, contextual, and geo-temporal factors.
• It employs a Siamese Autoencoder to reduce high-dimensional sparsity and learn compact latent representations by integrating encoder-decoder architecture with geo-temporal cues.
• Empirical evaluations show improved linkage accuracy and operational efficiency, underscoring its potential for real-world forensic and investigative applications.

The Violent Crime Linkage Analysis System (ViCLAS) is a specialized national database maintained by the UK’s Serious Crime Analysis Section (SCAS), a division of the National Crime Agency. Its primary aim is to facilitate the linkage of serious sexual and violent offences by grouping cases committed by the same offender(s), especially where physical or forensic evidence is sparse or absent. ViCLAS accomplishes this through the structured collection and binary encoding of behavioural (modus operandi), contextual, and geographic-temporal data, enabling investigators and machine learning systems to exploit patterns in offender actions and incident circumstances.

1. Structure and Encoding of ViCLAS Data

ViCLAS catalogues offences using an extensive set of categorical variables describing both behavioural and contextual elements. Key behavioural features include attributes such as approach method, level of violence, and weapon use. Contextual variables cover aspects like the location type, time of day, and victim characteristics. For pairwise crime comparison, geographic and temporal measures such as the continuous distance between scenes and the interval between offence dates are computed and log-transformed.

Each offence is represented as a binary vector $x \in \{0,1\}^M$, where $M=446$ in the raw schema. A value of 1 in this vector denotes the presence of a characteristic (e.g., "knife used"), and 0 its absence. Pairwise geo-temporal information $g(x_i,x_j) \in \mathbb{R}^2$ (log distance, log interval) is treated as an auxiliary input during model inference. This binary format addresses the heterogeneity and high dimensionality typical of violent crime case data.

2. Siamese Autoencoder Framework for Crime Linkage

To address the challenges posed by the high-dimensional, sparse, and heterogeneous nature of ViCLAS data, recent work introduced a Siamese Autoencoder architecture. The framework jointly learns a compact latent representation for each case and a similarity metric for case linkage, integrating domain-prioritized reduction strategies and geo-temporal cues to maximize linkage accuracy.

2.1 Encoder and Decoder Architecture

The encoder consists of two fully connected (linear) layers with ReLU activations, transforming the 446-dimensional binary input into an 8-dimensional latent code:

• $$446 \rightarrow \text{ReLU} \rightarrow 128 \rightarrow \text{ReLU} \rightarrow 8$$

The decoder mirrors the encoder in reverse, integrating geographic-temporal signals after the first decoding layer:

• $$8 \rightarrow \text{ReLU} \rightarrow 128 \rightarrow$$ (add geo-temporal embedding) $$\rightarrow \text{ReLU} \rightarrow 446 \rightarrow \text{sigmoid}$$

The geo-temporal vector $g = (\log(\text{distance}), \log(\text{interval}))$ undergoes a linear mapping $W_g \cdot g + b_g \in \mathbb{R}^{128}$, which is added to the 128-dimensional hidden state of the decoder before the second layer:

$$h_d^1 = \text{ReLU}(W_d^1 z + b_d^1 + (W_g g + b_g))$$

This mechanism ensures that geo-temporal variation informs the reconstruction, amplifying behavioral and contextual codes rather than being overshadowed during input embedding.

2.2 Loss Functions

The model is supervised by a composite objective:

$$L = \alpha L_{\mathrm{contrast}} + \beta L_{\mathrm{recon}}$$

Empirically, the weighting parameters are set as $\alpha=1.0$, $\beta=0.2$.

Contrastive Loss:

For two cases $(i,j)$ with latent codes $(z_i, z_j)$ and linkage label $y \in \{0,1\}$ (1 = linked, 0 = unlinked), and a margin $m=5$:

$$d(z_i, z_j) = \sqrt{\sum_k (z_{ik} - z_{jk})^2} + \sum_k |z_{ik} - z_{jk}|$$

$$L_{\mathrm{contrast}} = \mathbb{E}\left[ y \cdot d^2 + (1-y) \cdot \max(m-d, 0)^2 \right]$$

Reconstruction Loss:

Employs cosine similarity between the true ($v$) and reconstructed ($\hat{v}$) binary input vectors:

$$L_{\mathrm{recon}} = -\mathbb{E}\left[ \frac{v^\top \hat{v}}{ \|v\| \|\hat{v}\| } \right]$$

2.3 Inference and Similarity Scoring

After training, each crime is embedded as $z$. The pairwise Euclidean distance $D_{ij} = \|z_i - z_j\|_2$ is converted to a probability-like similarity:

$$S_{ij} = \exp \left( - \frac{D_{ij}}{\beta'} \right)$$

where $\beta' = m / 1.5$ for alignment with the supervised margin. These $S_{ij}$ values are used to rank candidate linkages.

3. Semantic Feature Reduction and Data Sparsity Management

The original ViCLAS binary encoding is highly sparse, with approximately 91% zero entries. Expert-driven data reduction was undertaken in collaboration with operational analysts and forensic psychologists, consolidating semantically related features to mitigate sparsity and focus on investigative utility. Five mapping strategies ("Map 1"–"Map 5") were developed:

Mapping Strategy	Feature Count	Reduction (%)	Basis
No Map (Raw)	446	0	Original schema
Map 1	282	-36.8	Analyst/operational focus
Map 2	384	-13.9	Forensic psychologist input
Map 3	266	-40.4	Hybrid of 1 & 2
Map 4	217	-51.3	Behavioural distinctiveness
Map 5	286	-35.9	Refined Map 4, preserves detail

Category examples include merging distinct car park types and collapsing 16-dimensional weapon usage into 4 categories (firearm, edged, blunt, acquisition). This consolidation increases the density of positive features, improving the reliability of pattern discovery in subsequent learning.

4. Empirical Evaluation and Ablation Analysis

Experiments used 5-fold cross-validation on two ViCLAS datasets: a 1,482-case proof-of-concept and a full set of 11,970–22,282 cases with multiple victims/scenes. Quantitative results highlight the efficacy of model architecture and feature reduction strategy.

4.1 Baseline Comparison

On the full multi-scene dataset, Area Under Curve (AUC), true positive rate at fixed false positive rate, and AUPRC were reported:

Method	AUC (%)	TP@Fixed FP (%)	AUPRC (%)
Logistic Regression	75 ± 2.97	70.43 ± 2.12	10.24
Naive Siamese*	76 ± 2.15	67.53 ± 2.60	13.45
Siamese AE (no map)	77 ± 2.11	68.31 ± 1.92	13.32
Siamese AE (Map 5)	84 ± 2.86	79.38 ± 2.56	15.43

*Naive Siamese denotes prior twin MLP with geo-temporal input-level concatenation. With Map 5, the proposed architecture achieves an absolute AUC gain of 12% over logistic regression and 6.7% over the Naive Siamese, with a 50.7% boost in AUPRC over logistic regression.

4.2 Ablation and Architectural Variants

AUC is sensitive to encoder structure, skip connections, and geo-temporal integration strategy:

Architecture	AUC (%) (geo-temp at decoder)	AUC (%) (geo-temp at input)
MLP (no skip, 2+2)	77.29	76.43
1D-CNN (no skip)	61.74	58.45
SIREN (no skip)	58.28	55.19

Removal of skip connections and decoder-level integration consistently improved performance. A plausible implication is that decoder-level integration prevents geo-temporal features from being masked by high input sparsity.

4.3 Out-of-Time Validation

Evaluation on 1,165 new cases (2021–2025, COVID-era) demonstrates robustness to temporal covariate shift:

Feature Map	Recall (%)	False Positives
Raw (446)	51.5	174,549
Map 1	77.9	399,461
Map 5	42.7	97,013

Under a 15% false positive rate budget, workload can be reduced by up to 80% with moderate recall maintained, indicating operational scalability under shifting offence patterns.

5. Interpretability and Investigative Impact

Embedding analysis and model reconstructions support investigative workflows. The learned 8-dimensional codes can be projected (via t-SNE or UMAP) to visualize clusters corresponding to serially linked offences, aiding in series identification. Decoder-based feature attribution reveals which abstracted behaviours (e.g., “blunt weapon presence”, “surprise approach”) persist in the latent representation, allowing practitioners to relate latent code differences to operational categories.

For unresolved crimes, the model outputs a top-K ranked list of potential prior links via the similarity score $S_{ij}$. Empirical trials report that true linked cases are often found within the top 10–20 ranked cases, significantly narrowing the search scope.

Choice of feature mapping affects the model’s operational footprint. Map 1 is recommended for settings where maximal recall is required (i.e., exhaustive investigative sweeps), whereas Map 5 offers a workload-efficient screening protocol suitable for resource-constrained contexts.

6. Methodological Considerations and Extensions

The approach of combining ViCLAS’s dense behavioural/contextual encoding with a Siamese Autoencoder, decoder-level geo-temporal fusion, and analytically-informed input reduction demonstrates that advanced machine learning can address both structural and operational requirements of violent crime linkage analysis. By optimizing latent code interpretability, recall, and positive predictive value, the system supports practical investigative review and deployment. The methodology is extensible to analogous high-dimensional sparse linkage problems where domain-informed feature reduction and interpretability are paramount.