Behavior-Based Similarity Matrices

• Behavior-Based Similarity Matrices are structured representations that quantify pairwise similarity based on observable behaviors, flow profiles, or response patterns.
• They are constructed by systematically extracting behavioral features and computing similarities using metrics such as cosine similarity, Procrustes alignment, and kernel methods.
• Applications span semantic shift detection, network role inference, model and brain alignment, user mobility analysis, and recommendation systems.

A behavior-based similarity matrix is a structured representation of pairwise similarity, affinity, or correspondence among entities as defined by their observable behaviors, flow profiles, or response patterns—rather than by static structural or attribute-based features. Such matrices enable unsupervised analysis, clustering, and interpretability across diverse domains, including temporal semantic shift, mobility modeling, network role inference, representational alignment, and recommendation systems.

1. Mathematical Foundations of Behavior-Based Similarity Matrices

Behavior-based similarity matrices are typically constructed by first defining a “behavioral profile” (vector, matrix, or higher-order embedding) for each entity, and then computing similarity (or distance) using a domain-appropriate metric. The precise definition varies by context:

• Temporal Evolution: For semantic trajectories, the diachronic word similarity matrix $S(w)\in\mathbb{R}^{T\times T}$ for word $w$ in $T$ time periods is $S_{ij}(w)=\mathrm{cosine}(e_i(w),e_j(w))$ with $e_t(w)$ the embedding at time $t$ (Kiyama et al., 16 Jan 2025).
• Network Flows and Roles: For nodes in a directed network, the feature vector $f_i$ for node $i$ aggregates the counts of incoming and outgoing walks of different lengths, often scaled by a spectral factor $\beta=\alpha/\lambda_1(A)$, where $A$ is the adjacency matrix. The similarity matrix $S_{ij}=\mathrm{cosine}(f_i,f_j)$ captures role-based equivalence (Cooper et al., 2010, Cooper et al., 2011).
• Representational Alignment: In model comparisons, the similarity matrix $S_{mn}$ encodes pairwise similarity between activation or output patterns of models $m$ and $n$ across standardized stimulus sets, using geometric (e.g., Procrustes, CKA, RSA) or functional (e.g., predictivity) metrics (Bo et al., 2024).
• User Behavior in Mobility and Recommendation: For user $u$, the association matrix $A_u$ summarizes temporal-location behavior, and user-user similarity is computed via weighted overlaps of top singular vectors (“eigen-behaviors”), producing $S_{pq}=\sum_i\sum_jw_{x_i}w_{y_j}|x_i^Ty_j|$ (Thakur et al., 2010).
• Bipartite and Item-based Networks: In recommendation, object-object similarity $s_{\alpha\beta}$ uses functions of shared neighbors, rating overlap, and degree statistics, assembled into a symmetric matrix for further analysis (Liu et al., 2015).

All constructions share the property that similarity is a function of empirical or inferred behavior, and that $S$ is either symmetric (self-similarity) or rectangular (cross-system or node comparison).

2. Construction Pipelines and Algorithmic Workflow

The typical construction of a behavior-based similarity matrix involves several algorithmic steps, adaptable to the particular application domain:

• Feature Extraction: For each entity, extract a behavior vector/matrix: temporal embedding, flow profile, output response, or association signature.
  • Temporal embeddings (e.g., PPMI-SVD across periods) are aligned so that coordinates are directly comparable; this often involves a joint-SVD or Procrustes step (Kiyama et al., 16 Jan 2025, Andreella et al., 2023).
  • In networks, node profiles are computed by path-counting up to length $K$ with scale factors to ensure convergence (Cooper et al., 2010, Cooper et al., 2011).
  • In representational analysis, stimulus-response matrices are whitened and compared by selected metrics (Bo et al., 2024).
  • Mobile user behavior is represented as normalized association matrices, SVD-truncated (Thakur et al., 2010).
• Similarity Computation: Pairwise similarity is typically determined by (a) cosine similarity, (b) inner-product–based kernels (CKA), (c) Procrustes alignment, (d) kernelized RBF transforms on computed distances, or (e) task-specific scoring functions (e.g., common neighbors, Sørensen, Adamic-Adar) (Liu et al., 2015).
• Matrix Assembly: For $N$ entities with behavioral vectors of dimension $d$, assembly costs $O(N^2d)$ for full pairwise computation. Symmetry is enforced where appropriate.
• Clustering and Downstream Analysis: The resulting $S$ can be used for hierarchical clustering (e.g., agglomerative clustering on vectorized $S$, as in semantic shift work (Kiyama et al., 16 Jan 2025)), spectral clustering (e.g., normalized cut (Cooper et al., 2010)), or direct graph algorithms (e.g., community detection for user similarity graphs (Thakur et al., 2010)).

Representative pseudocode for Procrustes-based matrix construction and RBF-kernelization is given in (Andreella et al., 2023).

3. Applications Across Domains

Behavior-based similarity matrices are scalable and interpretable tools across scientific disciplines:

• Semantic Shift Detection: In diachronic linguistics, $S(w)$ enables identification of stable periods (“blocks”), change points (block boundaries), and ephemeral events (off-diagonal spikes). Clustering $S(w)$ across vocabulary isolates groups with similar semantic-change trajectories, with applications to English (COHA, COCA) and Japanese corpora (Kiyama et al., 16 Jan 2025).
• Functional Role Discovery: In directed networks, such as metabolic, trade, or ecological flow systems, role-based similarity groups nodes by global flow patterns rather than local density, revealing stratification (e.g., “core” vs. “periphery” vs. “intermediate” in trade and food-webs) (Cooper et al., 2010, Cooper et al., 2011).
• Model and Brain Alignment (NeuroAI): Model–model and model–brain similarity matrices assess functional correspondence across trained/untrained states, architectures, or fMRI subjects. Geometry-preserving metrics (Procrustes, CKA) show the strongest behavioral and group-level discriminability (Bo et al., 2024, Andreella et al., 2023).
• Mobile User Clustering and Mobility Models: Pairwise similarity between behavioral profiles recovers modular groupings in real wireless trace data, highlighting the insufficiency of legacy mobility models that fail to capture behavioral diversity (Thakur et al., 2010).
• Recommendation and Bipartite Networks: Item–item similarity matrices are foundational for recommendations, where the stability (resilience under data subsampling) and form of the metric matter critically. Unstable metrics recommend more false positives; robust metrics (CN, AA, RA) enable more consistent outputs (Liu et al., 2015).
• Anomaly and Intrusion Detection in Graphs: In BS-GAT, a three-tiered similarity matrix among network flows feeds into attention-weighted message passing for intrusion detection; the design ensures graph construction with uniform node degree and performance gains over alternative methods (Wang et al., 2023).

4. Methodological Considerations: Normalization, Stability, and Multiscale Analysis

Robustness and interpretability of behavior-based similarity matrices depend on several factors:

• Normalization and Alignment: Joint SVD, normalization of feature vectors (e.g., $z$-scoring dimensionwise), or orthogonal Procrustes alignment are required for coherent comparisons across periods, systems, or individuals (Kiyama et al., 16 Jan 2025, Andreella et al., 2023).
• Stability: In sparse, incomplete, or randomly subsampled data, the stability of the similarity metric can be quantified using bias ($\mu$), standard deviation ($\sigma$), and Pearson correlation ($\rho$) across multiple samples. Indices cluster by stability characteristics, and practical “top-$n$-stability” filtering improves recommendation robustness (Liu et al., 2015).
• Multi-scale and Parameter Selection: In role-based similarity, the scale parameter $\alpha$ tunes the locality versus globality of role assignment. $\alpha\to0$ recovers local (degree-based) similarities; $\alpha\to1$ emphasizes global flow structure, potentially at the cost of numerical conditioning and interpretability (Cooper et al., 2010, Cooper et al., 2011).
• Dimensionality Reduction: For high-dimensional matrices (e.g., fMRI), efficient SVDs and rank-reduction tricks accelerate Procrustes alignment and make subsequent clustering feasible (Andreella et al., 2023).

5. Evaluation Metrics and Validation Procedures

Assessment of behavior-based similarity matrices and their utility involves both intrinsic and extrinsic validation:

• Clustering Validity: Silhouette scores on the clustering of behavior-based feature vectors provide unsupervised validation of discovered group structure (Kiyama et al., 16 Jan 2025).
• Classification Accuracy: On pseudo-labeled data, the ability of $S$ to distinguish among schema types (e.g., patterns of semantic change) is quantified (e.g., 72.1% accuracy for best matrix/clustering choices) (Kiyama et al., 16 Jan 2025).
• Functional Alignment and Discriminability: Pearson correlations between representational and behavioral similarity matrices, and group-separation statistics (e.g., $d'$ for trained/untrained) provide direct functional interpretation for model/brain alignment (Bo et al., 2024).
• Graph Statistics: Modularity, clustering coefficient, characteristic path length, and number of communities in similarity graphs distinguish real-world behavioral diversity from model artifacts (Thakur et al., 2010).
• Recommendation Stability: Mean ranking position $\langle R\rangle$ under cross-sample agreement connects similarity-matrix stability to practical system reliability (Liu et al., 2015).

6. Limitations, Extensions, and Open Challenges

Existing research highlights several limitations and avenues for development:

• Choice of Metric and Parameterization: The appropriateness of similarity metrics is data- and task-dependent. Metrics emphasizing global geometry (e.g., Procrustes, CKA, RSA) have been shown to better correspond with functional and behavioral distinctions, but may be less interpretable or computationally efficient in specific cases (Bo et al., 2024).
• Computational Complexity: Matrix assembly and clustering scale quadratically or cubically in the number of entities; however, efficient implementations (joint SVD, sparse matrix ops, effective rank reduction) render practical analyses feasible for vocabularies and subject counts on the order of $10^3$–$10^4$ (Kiyama et al., 16 Jan 2025, Andreella et al., 2023).
• Generalization Across Systems: Role-based similarity definitions can in principle be extended to time-varying, weighted, or motif-enriched networks, with potential computational and modeling trade-offs (Cooper et al., 2011).
• Sensitivity to Sampling and Data Loss: Instability under partial observation remains a challenge for metrics involving higher-order or quadratic degree terms. Structures designed to preserve only the top $n$ most stable similarities mitigate this to an extent (Liu et al., 2015).
• Interpretability in High-Dimensional or Multi-Modal Systems: Direct mapping from similarity-matrix entries or clusters to interpretable groupings (e.g., semantic senses, behavioral subgroups) is nontrivial and may call for auxiliary weighting, visualization, or regression methods (Andreella et al., 2023).

7. Representative Matrices and Domain-Specific Constructions

The following table summarizes key behavior-based similarity matrix constructions across domains:

Domain	Entity	Feature Construction	Similarity Metric
Semantic Shift	Word	Temporal embeddings (joint-SVD)	Cosine
Directed Networks	Node	Scaled in/out walk counts	Cosine/EUCLID
Model Comparison	Model	Activations on benchmark stimuli	Procrustes, CKA, RSA, etc.
User Mobility	User	Spatio-temporal association SVD	Weighted eigen-behavior overlap
Recommendation	Item	User overlap, rating statistics	CN, AA, RA, COS, etc.
Intrusion Detection	NetFlow	Rule-based network flow feature vectors	Scalar similarity function

These schema illustrate the general principle of representing richly structured, high-dimensional behavioral patterns as similarity matrices, facilitating unsupervised discovery of functional units, change-points, or clusterings.

---

Behavior-based similarity matrices provide an explicit, interpretable, and versatile representation for quantifying functional similarity based on observable behaviors or response profiles. Their continued evolution—as evidenced in temporal semantics, role discovery, behavioral neuroscience, anomaly detection, and recommendation—depends on methodological innovations for scale, stability, and appropriate metric selection (Kiyama et al., 16 Jan 2025, Cooper et al., 2010, Cooper et al., 2011, Bo et al., 2024, Andreella et al., 2023, Wang et al., 2023, Thakur et al., 2010, Liu et al., 2015).