Entity Profiling & Uniqueness Analysis

Updated 9 December 2025

Entity profiling and uniqueness analysis is a methodology that identifies, characterizes, and quantifies distinguishing features of entities to enable robust matching and privacy risk assessments.
The approach integrates techniques from information theory, machine learning, and network science, including ACID properties, Shapley value metrics, and embedding models like HAS.
Empirical applications in social networks, knowledge graphs, and population-scale data demonstrate improved accuracy and speed while highlighting challenges like feature redundancy and impersonation risks.

Entity profiling and uniqueness analysis encompass the systematic identification, characterization, and quantification of features that distinguish entities—such as individuals, user accounts, or real-world objects—within a collection or domain. Accurate profiling enables robust matching, search, authorship attribution, and de-anonymization, while quantifying uniqueness informs privacy risk, re-identification probability, and the selection of informative variables. This area synthesizes methods from information theory, machine learning, and network science, with domain-specific modeling tailored for structured data (e.g., knowledge graphs), text corpora, and social networks.

1. Theoretical Foundations of Uniqueness and Attribute Discriminability

The quantification of uniqueness is central to entity profiling. In the context of public attributes, the ACID framework (Goga et al., 2015) formalizes the reliability of an attribute for matching and profiling via four properties:

Availability (A): The probability the attribute is present on both profiles of a true match.
Consistency (C): The likelihood that matching profiles report similar or identical values.
Non-Impersonability (nI): The difficulty for adversaries to create convincing forgeries.
Discriminability (D): The probability that a true profile’s attribute does not collide with any non-match above a chosen similarity threshold.

The discriminability $D$ (and its effective variant $\tilde{D}$ ) operationalizes attribute-level uniqueness—the rate of false matches is inversely related to $D$ . In typical large-scale settings, only attributes with $A \approx C \approx nI \approx D \approx 1$ can provide highly reliable entity resolution; otherwise, ambiguity and collision rates constrain recall at high precision.

The Uniqueness Shapley measure (Seiler et al., 2021) offers a principled, order-invariant allocation of identification power to variables. For a subject $t$ and variable $i$ among categorical variables $V=\{1,\ldots,d\}$ , the Shapley value $\varphi_i(t)$ is the expected marginal log-reduction in cohort size due to revealing $x_{t,i}$ , averaged over all reveal orders:

$\varphi_i(t) = \sum_{S \subseteq V \setminus \{i\}} \frac{|S|! (d - |S| - 1)!}{d!} [\log|C_t(S)| - \log|C_t(S \cup \{i\})|]$

Summing over subjects provides global or subgroup uniqueness contributions, linking directly to (conditional and cross-) entropy:

$\varphi_i^{1:n} = \frac{1}{d} \sum_{S \subseteq V \setminus \{i\}} \binom{d-1}{|S|}^{-1} H(X_i \mid X_S)$

2. Algorithmic Approaches to Entity Profiling

2.1. Feature Embedding and Pattern Profiling

Knowledge graphs (KGs) demand techniques that account for both attributive and relational information. The HAS model (Zhang et al., 2020) leverages homophily-, attributive-, and structural-equivalency–based random walk path strategies to embed entities in a latent space reflecting multi-pattern similarity. Distinctiveness scores for features (labels) are then computed as:

$d(l) = \frac{1}{|E_t^l|^2} \sum_{i,j \in E_t^l} \operatorname{sim}(i,j) - \frac{1}{|E_t^l||E_t^c|} \sum_{i \in E_t^l, j \in E_t^c} \operatorname{sim}(i,j)$

where $E_t^l$ are entities with label $l$ , and $\operatorname{sim}$ is the cosine similarity of HAS embeddings. High $d(l)$ indicates that the feature both internally clusters positives and externally separates them from negatives, marking it as distinctive and suitable for profiling.

2.2. Statistical Profiling via Divergence

Latent Personal Analysis (LPA) (Mokryn et al., 2020) defines the domain-based distance $d(e,D)$ between entity $e$ ’s usage vector and the global domain distribution $D$ using Kullback–Leibler divergence, with $\varepsilon$ -padding to handle zero counts:

$d(e,D) = D_{KL}(V_e' \| D) = \sum_t (V_e'(t) - D(t)) \log\frac{V_e'(t)}{D(t)}$

A personal signature $\sigma_e$ is constructed by selecting the top $N$ terms whose contributions to $d(e, D)$ are maximal in magnitude, with sign indicating over/under-usage.

3. Practical Computation and Scalability

Direct calculation of uniqueness measures can be computationally prohibitive. The Uniqueness Shapley metric's evaluation scales as $O(n^2 d 2^d)$ naively, but an all-dimension tree (ADTree) (Seiler et al., 2021) reduces query time for matching cohort sizes from $O(n)$ to $O(|S|)$ via efficient indexing on value conjunctions, yielding speed-ups of up to $\sim$ 2000 $\times$ on real datasets.

The LPA framework (Mokryn et al., 2020) uses batch computation of (padded) entity-domain divergences and term-wise sorting to efficiently extract signatures even for thousands of entities. HAS embedding for profiling in KGs (Zhang et al., 2020) scales via negative sampling, allowing application to datasets with millions of nodes and hundreds of types.

4. Empirical Evidence and Application Domains

ACID-based evaluation on cross-network social profile matching (Goga et al., 2015, Halimi et al., 2017) reveals inherent limits: for real names, 60% of Facebook users share at least one exact name, sharply bounding discriminability. Full-scale experiments show that while small-sample precision/recall can reach 90–95%, real-world, high-recall matching at similar precision rarely exceeds 30–40%. Attributes such as “friends” lists can improve discriminability only if linkability across sites is high.

In LPA, signatures effectively flag sockpuppets and front-users; precision/recall for sockpuppet detection can reach F₁ ≈ 0.93–0.99 under optimal thresholds, outperforming standard TF-IDF+cosine strategies (Mokryn et al., 2020).

4.2. Structured Knowledge Graphs

The HAS+r model in knowledge graphs consistently outperforms random, TF-IDF, and single-view baselines on both intrinsic measures (MAP@5, F-M@10) and human-evaluated profile quality (Zhang et al., 2020). Concise sets of high-distinctiveness labels yield accurate, interpretable profiles, with extrinsic tasks (e.g., “spot the difference”) showing profile-aided identification yields 16% higher accuracy and %%%%29 $D$ 30%%%% speed-up for humans.

4.3. Population-Scale Profiling

Application of the Uniqueness Shapley measure to North Carolina voter registration (n≈7.5 million) reveals that zip code (8.6 bits) and age (5.4 bits) are far more identifying than race (1.2), party (1.5), or gender (1.2), with all measures exceeding marginal entropy due to variable dependence. Subgroup analysis demonstrates identification power shifts: race's contribution rises from 0.53 (white) to 5.1–5.4 (small minorities), and coarser binning of variables predictably reduces uniqueness (Seiler et al., 2021).

5. Risks, Limitations, and Privacy Countermeasures

Collisions and impersonation attacks present fundamental limits to uniqueness and reliability. In cross-OSN matching, about 1% of user accounts have near-perfect impersonators, and 7% of true matches have no consistent public attribute (Goga et al., 2015). Even under optimal multi-stage workflows, recall above 40% is unattainable at 95%+ precision for large populations.

LPA and linear attribute-weighted classifiers illustrate that reducing the utility of highly discriminative attributes (e.g., images) can halve the attacker's success rate with only minor functional loss (Halimi et al., 2017). The framework models privacy–utility trade-offs as constrained optimization, limiting maximum similarity while maximizing retained functionality.

Uniqueness profiling relies on sufficient entity coverage and the availability of domain-wide or population reference statistics. All methods are constrained by the missingness, noise, and idiosyncrasies of real-world data. For example, LPA loses performance on short texts; signature interpretability and coverage depend on the choice of $N$ and domain specificity (Mokryn et al., 2020); HAS feature selection efficacy is sensitive to label frequency and redundancy (Zhang et al., 2020).

6. Methodological Best Practices and Extensions

Robust entity profiling demands selection of features with jointly high availability, consistency, discriminability, and non-impersonability (Goga et al., 2015). Evaluation should employ “reliability-preserving” negative sampling to accurately estimate uniqueness and expected false-match rates at population scale. Multi-stage workflows (candidate generation, disambiguation, and “Guard” steps) help mediate recall–precision trade-offs and mitigate impersonation risks.

Shapley-based and entropy-linked uniqueness metrics generalize across entity types, modalities, and domains (Seiler et al., 2021), but require appropriate variable selection and binning for non-categorical data. LPA applies to any domain exhibiting a pronounced head–tail distribution, including product ratings and consumption patterns, provided entity sample size is sufficient and signatures are domain-calibrated (Mokryn et al., 2020). HAS and similar embedding models are applicable wherever neighborhood structure and attributes can be captured as walks or context pairs (Zhang et al., 2020).

7. Comparative Summary of Approaches

Method/Model	Profile Type	Uniqueness Metric
ACID (Goga et al., 2015)	Social profiles	Discriminability ( $D$ , $\tilde D$ ), ACID properties
Uniqueness Shapley (Seiler et al., 2021)	Categorical variable records	Shapley value (marginal log-reduction in cohort size), link to entropy
HAS (Zhang et al., 2020)	Knowledge graphs	Embedding-based distinctiveness $d(l)$
LPA (Mokryn et al., 2020)	Discrete distributions (text, consumption)	Domain-based KLD, entity signature $\sigma_e$
Profile Matching ML (Halimi et al., 2017)	Attribute sets (OSN accounts)	Weighted similarity, feature importance, privacy–utility curves

These paradigms each target aspects of entity uniqueness: ACID and Shapley quantify identification power at the attribute or variable level; LPA and HAS synthesize composite profiles with interpretable unique features; machine learning approaches optimize global similarity for matching and privacy control.

References: