WikiProfile: Profiling Users & Knowledge

• WikiProfile is a class of analytic systems that profiles users, knowledge, or contributors by leveraging Wikipedia’s structured data such as article entities, categories, and edit histories.
• It employs techniques like entity extraction, category graph traversal, and unsupervised clustering to generate actionable profiles for semantic structuring and information filtering.
• The frameworks extend to evaluating LLM factuality, neural knowledge completion, and OSINT-driven analysis, offering insights into methodological performance and scalability.

WikiProfile is a class of analytic systems and computational benchmarks designed for profiling users, knowledge, or contributors, typically by leveraging Wikipedia's structured information—article entities, category graphs, edit histories—or by adapting “wiki-style” conventions to individual profiling from heterogeneous data. Across distinct research strands, WikiProfile frameworks operationalize user/topic profiling, contributor behavior classification, activity categorization, parametric factuality diagnosis, and infobox completion on large-scale knowledge graphs.

1. Topical Generalization and User Profile Structure

Early work on WikiProfile systems targeted semantic structuring of user interest/expertise profiles by generalizing fine-grained topics to broader semantic clusters using Wikipedia’s category hierarchy (Olieman et al., 2016). The canonical pipeline operates as follows:

• Entity Extraction and Weighting: For a target user, all authored documents are processed with an Entity Recognition and Disambiguation (ERD) system (e.g., DBpedia Spotlight) to yield a bag of weighted DBpedia article URIs, where the weight of topic $e$ is the count of distinct documents linking to it.
• Category Graph Traversal: For each article entity $e$, upward traversal up to $m$ hops ($m=3$) in the DBpedia category graph generates a candidate set $C_e$ of Wikipedia categories. The complete set $C$ is the union across all extracted entities.
• Association and Ranking: A sparse matrix $D$ stores, for each candidate category $c\in C$ and each topic $e$, the shortest path length $d_{c,e}$. AdoptionRank is computed per category:

$e$0

where $e$1 is a penalty inversely proportional to the square of the number of covered topics.

• Clustering Algorithm: Categories are sorted by increasing AdoptionRank. Each serves as a label for a cluster if it covers at least two currently unassigned topics. Remaining topics are treated as unclustered “orphans.”
• Profile Presentation (Layout Modes): UI modes were empirically compared through a user study:
  • Flat: Simple ranked list of topics.
  • Nested: Accordion-style clusters by category labels.
  • Clustered: A hybrid that foregrounds representative topics per cluster for immediate specificity, with the rest collapsible.

Key findings show users rate the nested structure as more useful, but often overlook specific (low-level) topics, motivating the foregrounding hybrid (Olieman et al., 2016).

2. Contributor Behavior Profiling via Clustering

WikiProfile approaches for contributor classification focus on unsupervised clustering of edit history time series (Krishna et al., 2018). The workflow extracts a vectorized behavioral profile per contributor:

• Features (per contributor):
  • Editing intensity: Ratio = edits per day of presence.
  • Temporal gaps: Mean/Max between consecutive edits (in months).
  • Activity regularity: Number of consecutive active months, mean and standard deviation of edits per month.
• Preprocessing: Contributors with fewer than 100 edits or unregistered/bot accounts are excluded.
• Clustering: Ward’s hierarchical clustering followed by Partitioning Around Medoids, with the number of clusters $e$2 selected by maximizing average silhouette score.
• Interpretation: Principal components analysis identifies semantic axes—edit volume, inactivity gap, regularity. Final label types consistently stabilized across languages:
  • On a mission: Burst editing, long subsequent inactivity.
  • Casual: Moderate overall engagement.
  • Regular: Consistently engaged month-on-month.
  • Top: High sustained intensity and regularity.

Cluster assignment can be performed incrementally after each new edit, supporting online, adaptive engagement interventions for community management (Krishna et al., 2018).

3. Interest Profiling on Social Media via Wikipedia Graph Navigation

Interest profiling on user-generated content can map free-form social activity to predefined interest categories by graph navigation over Wikipedia’s category structure (Torrero et al., 2018). The pipeline executes:

• Topic Extraction: Harvests user posts via social login and applies entity extraction (e.g., Dandelion API) to generate a topic map of Wikipedia entities with occurrence counts.
• Category Mapping via Graph Navigation:
  • For each topic $e$3 and sink category $e$4, all upward-pointing paths (via hypernym links) of length at most $e$5 are enumerated.
  • The weight $e$6 is proportional to the count of such paths, inversely scaled by path length and further normalized so that $e$7.
• Category Scoring: Per-user, each sink receives a weighted sum of topic contributions, normalized over all sinks.

Evaluation on manually labeled Twitter accounts shows the framework correctly identifies the top category for $e$880% of cases—a result robust to the absence of supervised learning and domain-specific tuning (Torrero et al., 2018).

4. Information Filtering with TF·IDF and ESA

A distinct thread leverages Wikipedia at the indexing level for TF·IDF-driven information filtering and user profiling (0804.2354):

• Indexing Pipeline: Wikipedia articles are preprocessed to produce a lemma-article posting table with TF·IDF weights, supporting scalable inverted search.
• User Document Mapping: Each user document is represented in the Wikipedia concept space via Explicit Semantic Analysis (ESA), matching with top-K most similar WP articles.
• Category Clustering: Extracted article categories are weighted by similarity scores and greedily clustered, excluding parent–child category pairs to create semantically coherent user topics.
• Personalized Filtering: For each user topic, a subcorpus index is constructed and used to classify new incoming documents (e.g., email, news) by computing aggregate relevance scores and filtering at a user- or system-defined threshold.

Key statistics demonstrate the scalability and coverage attainable with this technique, even on multi-million-article corpora (0804.2354).

5. Profiling Factual Knowledge in LLMs

WikiProfile is also instantiated as a benchmark to diagnose LLM factuality, distinguishing encoding (fact is present) from recall (fact can be retrieved) (Calderon et al., 15 Feb 2026):

• Benchmark Construction:
  • Automated extraction of facts from Wikipedia, annotated via NER and context constraints.
  • Generation of a suite of probing tasks per fact: direct/reverse closed-book QA, proposition completion, context-primed rephrasings, and multiple-choice formats.
  • Fact selection and filtering is partially LLM- and Google Search-grounded for uniqueness and answerability.
• Profiling Taxonomy:
  • Encoding: Whether a fact is present in the model; operationalized as at least one correct response in encoding-related tasks above a threshold.
  • Recall: Whether the fact can be produced consistently under various question patterns; requiring correct response across all recall probes.
  • Classification of fact–model pairs into profiles: encoding failure, direct recall, recall only with inference (chain-of-thought), recall failure, or inference-only knowledge.
• Empirical Findings: SOTA LLMs (GPT-5, Gemini-3 Pro) encode nearly all facts (>95%), but recall is bottlenecked (only 65–75% direct recall). Long-tail facts and reverse association forms exhibit higher recall failure. Inference-time reasoning (CoT) recovers a substantial fraction of missed facts.

This analysis reframes the “knowledge bottleneck” not as parametric storage, but as an accessibility limitation, with strategic implications for benchmark design and model improvement (Calderon et al., 15 Feb 2026).

6. Neural Profiling for Knowledge Completion

Profiling in the sense of expectation generation and type inference leverages neural architectures over knowledge graphs (Wikidata/DBpedia) to fill in missing infobox fields (Ilievski et al., 2018):

• Formalism: Given partial facet–value information for an entity, the profiling task predicts a full probability distribution over plausible values for each missing facet.
• Autoencoder (AE) Architecture: Facet embeddings are concatenated and projected via a dense layer to per-facet softmax output heads; masking enforces generalization.
• Embedding-based Predictor (EMB): Pretrained entity embeddings serve as input; decoded as with the AE.
• Evaluation: Both AE and EMB outperform baselines (most-frequent-value, Naïve Bayes), especially for high-entropy facets such as “award received”.
• Crowdsourced Comparison: System-generated profiles more closely match human expectation distributions (measured by J-S divergence) than frequency-based models, especially as context size increases.

The system can operate efficiently in real time and enables population-scale infobox completion, anomaly detection, and expectation formation (Ilievski et al., 2018).

7. OSINT-driven Individual Profiling Frameworks

WikiProfile can also denote OSINT-style profiles constructed from heterogeneous public data (V et al., 2020):

• System Architecture: Modular design integrating input normalization, API-based data collection (WebMii, Pipl, Twitter, etc.), aggregation, and stylometric authorship analysis.
• Relevance Scoring: Each data item combined from multiple sources is scored by reliability, visibility, match similarity, and recency, then thresholded for inclusion.
• Authorship Clustering: High-dimensional stylometric feature vectors support attribution of social/posts/blog content to a single author, using cosine similarity and cluster confidence.
• Report Generation: The final profile assembles factual and inferred data into a wiki-format output, with inline source attributions and confidence annotations.

This design allows rapid aggregation of public data into structured, wiki-style profiles for applied domains (e.g., professional, legal, matrimonial screening), leveraging principles of information integration and authorship validation (V et al., 2020).