Architecting an Ethical, Adaptive, Session-Based Historical Exploration System

Full report

On this page

Introduction

In the contemporary digital landscape, personalized recommendation systems function as the primary engines for information dissemination, profoundly shaping human decision-making and epistemic exposure1. Traditional recommendation architectures, heavily reliant on collaborative filtering and persistent user profiling, are engineered to optimize for commercial engagement metrics, such as click-through rates and session duration4. When applied to the domain of historical exploration and digital humanities, this engagement-centric paradigm introduces profound ethical hazards. The algorithmic pursuit of sustained attention frequently manifests as "filter bubbles"—environments where internet customization isolates individuals from diverse perspectives, thereby reinforcing existing beliefs, attitudes, or misconceptions3. The design of an interactive historical globe necessitates a fundamental paradigm shift from engagement maximization to the ethical facilitation of "knowledge exploration via recommendations," a framework designed to maximize the total amount of diverse knowledge encountered by the user while strictly preserving human autonomy4. Humanities scholars and educational users exhibit distinct information-seeking behaviors that diverge fundamentally from the assumptions inherent in high-volume consumer recommendation models7. These behaviors include context volatility (where research interests shift rapidly based on immediate tasks), a paramount need for epistemic trust tied to verifiable provenance, a desire for strand continuity across long-term investigative threads, and a unique requirement for contrastive seeking—the deliberate pursuit of items that challenge current assumptions7. This report details a comprehensive conceptual model for an ethical, adaptive exploration system tailored for an interactive historical globe. By leveraging advanced session-based recommendation (SBR) techniques2, integrating Maximal Marginal Relevance (MMR) algorithms9, and deploying a strictly account-less, client-side data architecture, the proposed system facilitates the dynamic discovery of relevant events, regions, themes, and historical relationships. Crucially, it accomplishes this adaptive guidance without engaging in covert profiling, demographic inference, or the creation of ideological filter bubbles, ensuring that all adaptive behaviors remain completely explainable and instantly removable through a single, sovereign user control.

Transparent Recommendation Philosophy

The guiding philosophy of this architectural model is anchored in the principles of digital humanism, user-centric knowledge exploration, and uncompromising data privacy7. The system operates under the foundational premise that historical discovery should be expansive, intellectually challenging, and strictly under the conscious control of the user, adhering to the European Union's ethics guidelines for trustworthy AI: respect for human autonomy, prevention of harm, explicability, and fairness1. The system prioritizes epistemic trust over algorithmic omniscience. Recommender systems deployed within the humanities must recognize that relevance is inextricably linked to the verifiable provenance of information7. Consequently, the system transparently maps the causal, geographic, or thematic links between historical nodes, ensuring that recommendations are perceived not as opaque algorithmic decrees, but as structured, logical pathways that users can interrogate. Learning and discovery are understood to be enhanced not merely through similarity and homophily, but through serendipity—defined mathematically and conceptually as the discovery of information that is both unexpectedly novel and structurally relevant9. The system embraces contrastive seeking as a core utility, deliberately surfacing items that challenge current assumptions or highlight opposing historical viewpoints to actively puncture algorithmic filter bubbles3. To definitively eliminate the threats of long-term profiling, identity inference, and demographic targeting3, the system's memory is hermetically sealed within the boundaries of the active session. Once the session terminates, or whenever the user requests a reset, the mathematical state of the recommendation engine returns to absolute neutrality. The user retains complete authority over the algorithms governing their experience. All adaptive behavior is made visible, explainable, and configurable, establishing a sovereign user agency that respects the dynamic and multifaceted nature of historical inquiry.

Prohibited Practices in Historical Recommendation

Constructing a genuinely ethical exploration system requires the explicit definition and technological prohibition of practices that exploit user psychology or degrade the pedagogical value of historical data. The architecture structurally prevents the following outcomes through its reliance on session-only data and mathematically enforced diversity algorithms. The system strictly forbids partisan personalization, political profiling, and demographic inference. Traditional recommenders often infer a user's race, gender, nationality, or contemporary political alignment from their metadata or interaction history, subsequently serving content assumed to match those identities3. In the context of history, this practice is reductive and deterministic, projecting modern ideological biases onto the past and assuming that a user's demographic profile dictates their intellectual interests. The SBR architecture utilized here is entirely blind to user identity, isolating recommendations from any external demographic assumptions. Furthermore, the architecture algorithmically prohibits engagement-maximizing outrage loops. Commercial platforms frequently construct "rabbit holes" that trap users in continuous loops of emotionally volatile content to maximize session duration4. In historical exploration, this algorithmic tendency often manifests as the sequential chaining of wars, atrocities, plagues, or natural disasters, cultivating a skewed, trauma-centric view of human development. The underlying algorithms must penalize the over-representation of conflict or disaster nodes, ensuring that exploration encompasses periods of peace, scientific advancement, institutional development, and cultural flourishing. The system is also barred from treating repeated content as corroboration or reinforcing historical misconceptions. In historical research, the popularity or search frequency of a narrative does not equate to its factual accuracy13. The recommendation vectors are insulated from crowd-sourced popularity metrics that might artificially elevate historical myths. Additionally, the system must not rank regions by tragedy or geopolitical importance. Eurocentric assumptions or the disproportionate surfacing of the Global North are mitigated through a strict geographic-diversity balancing mechanism that governs the candidate selection pool, ensuring equitable exposure across all regions of the globe6.

Comparative Analysis of Recommendation and Discovery Systems

To inform the optimal architecture of the historical globe, fifteen distinct recommendation and discovery systems across gaming, streaming media, archives, and academic databases were evaluated. This analysis highlights effective mechanisms for serendipity and metadata utilization, alongside critical anti-patterns that the ethical globe must avoid.

Session-Only Data Model and Explicit-Interest Interface

To completely eliminate identity inference while maintaining a highly adaptive and personalized exploration experience, the historical globe relies exclusively on an ephemeral, client-side data model8. The architecture necessitates that no backend database records, stores, or analyzes the user's trajectory.

The Session State Vector Architecture

Upon accessing the historical globe, the user's browser initializes a JSON-based Session State Vector (SSV), stored securely and exclusively within sessionStorage. The SSV serves as the dynamic context for all algorithmic calculations, capturing real-time interaction metrics across both explicit user commands and implicit behavioral signals. The SSV ingests explicit parameters defined directly by the user. At launch, the user declares a learning\_goal, which dictates the system's overall information density. "Beginner Mode" optimizes for plain-language summaries and broad thematic overviews; "Classroom Mode" aligns with pedagogical standards and curates accessible primary sources; and "Expert Mode" maximizes data density, surfacing historiographical debates, raw archival data, and extensive citations. The user also defines accessibility\_settings, such as screen-reader optimization or high-contrast visuals, which the algorithm uses to prioritize nodes rich in audio descriptions or clear structural hierarchies. Furthermore, users manipulate an evidence\_strength\_filter, toggling between "Consensus History," "Primary Sources Only," or "Historiographical Debates." The user's explicitly\_selected\_interests are stored as a multidimensional array (e.g., Technology, Economics, Military, Arts, Politics), adjusted dynamically via interface sliders. Concurrently, the SSV captures implicit session-derived parameters. The selected\_regions parameter logs a recency-weighted array of geographic bounding boxes the user has interacted with, while selected\_dates maintains a chronologically weighted array of viewed timeframes. The system monitors the viewed\_event\_category to understand thematic leanings. Crucially, the system tracks dwell\_time—the milliseconds spent actively reading or viewing specific event nodes. This acts as a localized relevance multiplier, indicating genuine interest, though it is mathematically capped to prevent runaway feedback loops that could inadvertently create a filter bubble3. Finally, the SSV distinguishes between "Manual" and "Passive" modes. In manual mode, the user actively clicks through nodes; in passive mode, the globe auto-rotates, generating a curated, audio-visual narrative trajectory based purely on the SBR's next-best predictions.

Explicit-Interest Selection Interface

The interface deliberately eschews hidden profiling algorithms in favor of a transparent "Control Deck." Users are presented with a dynamic radar chart or a series of physical sliders representing the current algorithmic weightings governing their session. If the system observes the user lingering on the structural architecture of ancient Rome, it will visibly and automatically slide the "Architecture" and "Antiquity" sliders upward. The user has the sovereign authority to immediately drag the "Architecture" slider back down to zero and push "Economics" to maximum, instantly re-ranking the surrounding geographic nodes. This provides absolute algorithmic transparency, demystifying the recommendation process and granting the user total control over their intellectual trajectory1.

Shareable, Non-Identifying Journeys

Because the entire recommendation context is encapsulated locally within the SSV, users can share their exact historical journey without compromising their identity or transmitting personal data. The system serializes the core weights of the SSV and the current node ID into an encrypted Base64 URL parameter (e.g., globe.edu/explore?state=eyJoaXN0...). When another user, such as a student receiving a link from an educator, clicks the URL, their client deserializes the state, placing them at the exact point in the journey with the identical algorithmic weights. To prevent any theoretical deanonymization, the system strips out hyper-specific accessibility settings, hardware data, and granular dwell-time micro-metrics during serialization, encoding only the macro-level thematic and geographic weights11.

Algorithm Specification: Diversity, Serendipity, and Discovery

The fundamental mathematical challenge in SBR architectures is navigating the trade-off between relevance and diversity; optimizing solely for relevance inevitably leads to content homogenization, over-specialization, and the formation of filter bubbles2. To facilitate ethical and expansive historical exploration, the system utilizes a highly modified Maximal Marginal Relevance (MMR) architecture5, operating in tandem with Determinantal Point Processes (DPP) to ensure intra-batch diversity28.

The Historical MMR Equation

Standard MMR algorithms iteratively select items by maximizing their relevance to the user's query while simultaneously minimizing their similarity to items already selected or viewed9. The proposed Historical MMR adapts this logic specifically to the multi-dimensional SSV. The algorithm begins with Candidate Generation, where the system queries the knowledge graph to retrieve a global candidate set of historical nodes situated within a reasonable temporal, geographic, or semantic radius of the user's current node. These candidates then undergo Relevance Scoring ([Figure omitted from source export]), where they are evaluated against the SSV based on overlapping themes, geographic proximity, causal links, and alignment with the user-defined learning\_goal. To prevent homogenization, the algorithm applies a strict Diversity Penalty ([Figure omitted from source export]). The system computes the semantic and geographic distance of the candidate from the user's recent interaction history contained in the SSV. Candidates that are too similar to the immediate past are heavily penalized, forcing the algorithm to look further afield4. Concurrently, the system applies a Serendipity/Unexpectedness Boost ([Figure omitted from source export]). Utilizing the ExpSerendipity formulation, the algorithm actively rewards items that introduce a new categorical dimension not yet present in the SSV9. For example, if a user's session history consists exclusively of "European Military" nodes, a candidate node tagged "Asian Economics" receives a massive serendipity multiplier, encouraging a cross-disciplinary leap. The final ranking score determining which nodes are suggested next on the globe is computed as: [Figure omitted from source export] The variables within this equation are directly tied to user interface controls. Lambda ([Figure omitted from source export]) represents the user-controlled "Focus vs. Exploration" slider. A high [Figure omitted from source export] yields highly relevant, localized, and chronologically tight events. A low [Figure omitted from source export] aggressively increases the diversity penalty, forcing expansive exploration28. Gamma ([Figure omitted from source export]) represents the user-controlled "Serendipity" dial, which, when elevated, pushes the algorithm to find structural analogies in entirely different eras or regions. Finally, Delta ([Figure omitted from source export]) functions as the Underrepresented Region Modifier. The system maintains a macro-level index of regional visibility in digital history (e.g., noting that pre-colonial Africa or indigenous Americas are statistically underrepresented). If a candidate node belongs to one of these historically underrepresented regions and passes the minimum relevance threshold, it receives a flat priority boost ([Figure omitted from source export]), ensuring equitable geographic surfacing without compromising baseline relevance14.

Semantic and Structural Linkages

Unlike standard e-commerce SBRs that rely on collaborative user matrices (recommending items because "users like you also clicked this")2, this system uses a deterministic historical knowledge graph driven by the Art Institute of Chicago API and Wikidata methodologies20. Transitions are mapped along explicitly defined vectors rather than behavioral approximations. The algorithm moves users geographically ("What was happening across the border at this exact time?"), chronologically ("What happened next in this specific location?"), causally ("What were the upstream causes of this event?"), or via shared structural patterns ("Where else in history did a similar sociopolitical dynamic occur?").

Explainability and User Controls

To satisfy the ethical requirement for explicability, every adaptive recommendation generated on the historical globe is accompanied by a prominent, interactive "Why am I seeing this?" indicator. Clicking this indicator opens a transparent, plain-language explanation generated directly from the variables in the MMR algorithm, demystifying the mathematical weighting process1.

The Explanation Template

The system utilizes a deterministic natural language template that translates the highest-scoring vector from the MMR equation into readable text. A typical explanation reads as follows: "We recommended \[Node Title\] because it shares a structural pattern of \[Category/Theme\] with your previous interest in \[Previous Node\]. Additionally, it offers a contrasting perspective from \[Underrepresented Region\]. Your current Control Deck slider for \[Explicit Interest\] is set high, which prioritized this event over geographically closer options. \[Click here to adjust these settings\]."

The Sovereign Control Architecture

The Control Deck houses the mechanisms for sovereign user agency. It includes user-controlled recommendation criteria, allowing users to explicitly view and manipulate how Geography, Chronology, and Theme are currently weighted in the algorithm. A literal dial labeled "Surprise Me" acts as the Serendipity Control, increasing the [Figure omitted from source export] variable in the MMR equation, which relaxes geographic and temporal constraints to surface deep thematic analogies. Crucially, the system features a Geographic-Diversity Balancing and Contrasting-Case toggle labeled "Challenge My View." When a user activates this toggle, the algorithm deliberately inverses standard relevance logic. It explicitly seeks historical events that contradict or complicate the socioeconomic or geopolitical narrative of the currently viewed node, accommodating the contrastive seeking behavior essential for rigorous humanities research7. Furthermore, a single, global toggle switch labeled "Disable All Adaptation" allows the user to instantly strip all SBR logic, reverting the globe to a purely manual, unweighted geographic encyclopedia.

Ten Worked Recommendation Journeys

To demonstrate the capability of the algorithm to move users meaningfully without trapping them in engagement loops, the following ten journeys simulate the SBR routing across various vectors.

Journey 1: Structural Patterns (The Industrial Pivot)

The user initiates their session by viewing "The Industrial Revolution in Manchester" (UK, 1840s, Theme: Economics, Technology). Seeking a broader perspective, the user manually increases the "Serendipity" control ([Figure omitted from source export]) and requests a structural leap rather than a geographic one. The algorithm calculates structural similarities across different eras and spins the globe to Japan in the 1870s, surfacing the "Meiji Restoration." The explanation reads: "We are showing you the Meiji Restoration because it represents a rapid, state-sponsored parallel to the industrialization you viewed in Britain, occurring in a profoundly contrasting cultural and geographic context."

Journey 2: Causal Epidemiological Links (The Microbial Exchange)

The user explores "The Black Death" (Europe, 1347, Theme: Biology, Tragedy). The SBR vector detects a high dwell-time as the user reads the epidemiological data; consequently, the algorithm automatically boosts the "Science/Disease" weight. Searching for the next massive epidemiological shift, the system recommends "The Columbian Exchange" (Americas, 1492). The explanation reads: "Because you focused on the spread of pathogens, this node shows the next major epoch of global microbial exchange, shifting focus to the Americas."

Journey 3: Contrasting Cases (Emancipation vs. Revolution)

The user is viewing "The Emancipation Proclamation" (USA, 1863, Theme: Politics, Slavery). Wanting to avoid a narrow, US-centric view of abolition, the user engages the "Challenge My View" (Contrastive Seeking) toggle7. The algorithm penalizes North American geography and searches for contrasting structural events, recommending "The Haitian Revolution" (Haiti, 1791). The explanation reads: "We recommended this to provide a contrasting historical case: whereas US emancipation was enacted by a state authority during a civil war, the Haitian Revolution was a successful self-liberation by enslaved people."

Journey 4: Institutional Knowledge Transfer (The Preservation of Math)

Operating in "Beginner Mode" (lower information density, broader themes), the user researches "The Renaissance" (Italy, 1500s, Theme: Art, Science). The user clicks the evidence-strength filter to view the "Origins of Mathematics." The knowledge graph traces the institutional provenance of the Renaissance texts backward, recommending "The House of Wisdom" (Abbasid Caliphate, Baghdad, 800s). The explanation reads: "This node is surfaced to show the institutional origin of the classical texts that fueled the European Renaissance you were just viewing."

Journey 5: Underrepresented Region Surfacing (The Global Cold War)

The user is viewing "The Cuban Missile Crisis" (Cuba/US/USSR, 1962). The SBR algorithm detects a heavy session focus on the US/USSR binary, risking a geopolitical filter bubble. The [Figure omitted from source export] (underrepresented modifier) is triggered automatically. The system recommends "The Bandung Conference" (Indonesia, 1955). The explanation reads: "You've been exploring superpower conflicts. We recommended the birth of the Non-Aligned Movement in Indonesia to highlight how the Global South navigated this era."

Journey 6: Cross-Geographic Chronology (The Ancient Empires)

The user views "The Fall of the Roman Republic" (Rome, 27 BCE). The user explicitly locks the date slider in the Control Deck but asks the system to explore geography (Passive Mode activated). The globe rotates, maintaining the exact chronological weight, and surfaces "The Western Han Dynasty" (China, 20 BCE). The explanation reads: "We kept the time period identical but shifted the geography to show how another massive empire was managing statecraft at the exact same moment in history."

Journey 7: Deliberate Broadening (Breaking the Tragedy Loop)

The user is viewing "The Transatlantic Slave Trade" (West Africa \-\> Americas). The system recognizes a potential trauma-loop forming if it only suggests further colonial atrocities, a known anti-pattern in engagement algorithms4. The diversity penalty ([Figure omitted from source export]) is temporarily elevated. The system recommends "The Kingdom of Kongo" (Central Africa, 1400s). The explanation reads: "To provide broader context on West and Central African history beyond colonial exploitation, we recommended exploring the complex diplomacy and economy of the Kongo kingdom prior to the transatlantic trade."

Journey 8: Chronological Thematic Chaining (The Ripple of Revolution)

The user views "The French Revolution" (France, 1789). The user sets their learning goal to "Classroom" (triggering primary source density) and pushes the "Consequences" vector. The algorithm moves forward in time, chaining the thematic resonance of the event, and recommends "The Revolutions of 1848" (Europe-wide). The explanation reads: "We moved forward in time to show how the ideological concepts from the French Revolution eventually triggered a subsequent wave of uprisings across the continent."

Journey 9: Independent Parallel Discovery (The Invention of Writing)

The user views "Cuneiform in Sumer" (Mesopotamia, 3400 BCE). Fascinated, the user slides their explicit interest entirely to "Linguistics/Technology". Searching for maximum serendipity in human development, the algorithm recommends "The Cascajal Block / Olmec Writing" (Mesoamerica, 900 BCE). The explanation reads: "You are looking at the invention of writing. We surfaced Olmec script to demonstrate a completely independent, parallel development of written language in a disconnected geography."

Journey 10: Local-to-Global Thematic Linking (Urban Fires)

The user views "The Great Chicago Fire" (USA, 1871\) based on an initial local history query25. The user explicitly clicks on the "Urban Architecture" theme tag within the node. The algorithm searches the global graph for structurally identical catalysts resulting in architectural shifts, recommending "The Great Fire of London" (UK, 1666). The explanation reads: "Because you showed interest in how fires reshape cities, this node explores a historically significant parallel event that led to the first modern building codes."

Reset, Correction, and Feedback Model

To guarantee human autonomy and prevent the algorithmic determinism common in commercial platforms, the system implements a robust correction and reset architecture. A permanent, highly visible "Reset Globe" button serves as the ultimate fail-safe. Activating this control clears the sessionStorage vector entirely, instantly flushing the SBR and MMR algorithms and returning the globe to an unweighted, randomized state. If a user views an explanation and disagrees with the algorithm's assumption (e.g., "I clicked the Black Death by accident; I do not want disease history"), they can utilize a correction mechanism by clicking a "Disregard This" button. The system immediately and actively negates the event's features from the SSV. The feedback model itself represents a departure from commercial norms. Rather than relying on ambiguous "Thumbs Up/Down" inputs that opaquely optimize for engagement4, feedback is structured around relevance, diversity, and educational utility metrics. Users provide feedback such as "Too similar to last item" or "Not academically rigorous enough." This feedback instantly adjusts the [Figure omitted from source export] (relevance), [Figure omitted from source export] (serendipity), and information density parameters for the remainder of the session, shaping the algorithm without ever contributing to a permanent, backend user profile.

Privacy-Threat Assessment

Despite the strict absence of persistent user accounts, session-based recommendation systems (SBRs) face distinct privacy and security threats8 that this architecture must proactively mitigate. The primary threat is browser fingerprinting and state persistence. Third-party trackers could theoretically attempt to read the sessionStorage or utilize canvas fingerprinting to build a shadow profile of the user's historical and political interests over multiple seemingly unconnected sessions3. To mitigate this, the globe must operate in a strict first-party context, utilizing rigorous Content Security Policies (CSP) to block all external script execution. The SSV is genuinely ephemeral, explicitly deleted upon tab closure or session timeout. A secondary threat involves URL parameter leakage. When a user shares a journey via a serialized URL, the encoded SSV becomes public. If the SSV contains hyper-specific combinations of accessibility settings, hardware indicators, or highly localized geographic starting points, it could theoretically deanonymize the user. To counter this, the serialization process acts as a privacy filter, stripping out accessibility settings, user-interface preferences, and dwell-time micro-metrics. Only macro-level algorithmic weights are encoded, ensuring the shared journey remains entirely non-identifying11. Finally, the system must guard against linkage attacks on session data. A network observer analyzing the sequence of API requests to the backend knowledge graph could potentially infer a user's ideological focus (e.g., repeatedly requesting nodes related to anarchist history). To mitigate this, API requests fetch candidate nodes in bulk geographic or chronological blocks (e.g., downloading the entire metadata set for 19th-century Europe) rather than querying single, highly specific node IDs. The client-side application then performs the MMR filtering locally, ensuring the server remains entirely blind to the user's exact micro-trajectory.

Evaluation Plan

Standard recommendation metrics—such as Precision, Recall, and Click-Through Rate—are fundamentally incompatible with an ethical digital humanities tool designed for exploration rather than consumption5. The system's efficacy must be measured using "Beyond-Accuracy" metrics alongside rigorous, qualitative Human-Computer Interaction (HCI) evaluations4. Discovery quality and comprehension will be evaluated via randomized controlled trials where users are tasked with completing historical synthesis assignments. Success is measured by the user's ability to articulate the structural causes and consequences of an event, demonstrating deep comprehension rather than superficial engagement. Geographic diversity ([Figure omitted from source export]) will be monitored continuously. The algorithmic output will be logged in aggregate (strictly stripped of any associated user sessions) to monitor the global distribution of recommended nodes. The system must maintain high intra-batch and inter-batch diversity scores28, mathematically verifying that the Global South, pre-colonial histories, and underrepresented regions occupy at least a proportional share of the algorithmic recommendations. Humanities researchers will evaluate the system's "epistemic trust" through qualitative surveys7. These assessments will determine whether the "Why am I seeing this?" explanations accurately reflect historical realities and whether users genuinely felt in control of their narrative trajectory. Finally, to ensure the system is not inadvertently creating political filter bubbles3, evaluations will apply the MMR [Figure omitted from source export] (Coverage Diversity) metric9 to the session arrays in testing cohorts. If A/B testing reveals users routinely terminating sessions having explored only one narrow ideological category, the underlying diversity penalty ([Figure omitted from source export]) will be algorithmically strengthened for future global deployments, ensuring the system continually challenges over-narrow exploration.

Works cited

1. A Survey on Trustworthy Recommender Systems \- SciSpace, https://scispace.com/pdf/a-survey-on-trustworthy-recommender-systems-9t44xmnc.pdf

2. Enhancing Recommendation Diversity and Novelty with Bi-LSTM and Mean Shift Clustering, https://www.mdpi.com/2079-9292/13/19/3841

3. Filter Bubbles in Recommender Systems: Fact or Fallacy \- A Systematic Review \- arXiv, https://arxiv.org/html/2307.01221

4. Relevance Meets Diversity: A User-Centric Framework for Knowledge Exploration Through Recommendations, https://www.atailab.cn/seminar2024Fall/pdf/2024\_KDD\_Relevance%20Meets%20Diversity%20A%20User-Centric%20Framework%20for%20Knowledge%20Exploration%20Through%20Recommendations.pdf

5. Relevance meets Diversity: A User-Centric Framework for Knowledge Exploration through Recommendations \- arXiv, https://arxiv.org/html/2408.03772v1

6. Filter Bubbles in Recommender Systems: Fact or Fallacy \- A Systematic Review \- Essex Research Repository, https://repository.essex.ac.uk/36100/1/WIREs\_Revision\_Filter\_Bubble\_R1%20%288%29.pdf

7. An Experimental User Study On The Interplay of Trust and Serendipity In Recommender Systems | Request PDF \- ResearchGate, https://www.researchgate.net/publication/381812984\_Navigating\_Serendipity\_-\_An\_Experimental\_User\_Study\_On\_The\_Interplay\_of\_Trust\_and\_Serendipity\_In\_Recommender\_Systems

8. Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, https://sigir.org/sigir2023/program/proceedings/

9. Sequential Diversification with Provable Guarantees \- DiVA Portal, https://www.diva-portal.org/smash/get/diva2:2051451/FULLTEXT01.pdf

10. Diversification and fairness in top-k ranking algorithms \- Digital Commons @ NJIT, https://digitalcommons.njit.edu/cgi/viewcontent.cgi?article=2752\&context=dissertations

11. Dynamic Privacy-Preserving Recommendations on Academic Graph Data \- MDPI, https://www.mdpi.com/2073-431X/10/9/107

12. (PDF) Diversity and Serendipity Preference-Aware Recommender System \- ResearchGate, https://www.researchgate.net/publication/387217872\_Diversity\_and\_Serendipity\_Preference-Aware\_Recommender\_System

13. Transparency, Privacy, and Fairness in Recommender Systems \- Dominik Kowald, https://dominikkowald.info/documents/others/2024habilitation\_recsys.pdf

14. Why Play at Orientalism? \- Public Books, https://www.publicbooks.org/why-play-at-orientalism/

15. Europa Universalis IV: When Your Teen Wants to Rewrite History \- Screenwise, https://screenwiseapp.com/guides/the-grand-strategy-masterpiece-europa-universalis-4-review

16. Europa Universalis IV and Deep Learning \- Loading Journal, https://loading.journals.publicknowledgeproject.org/index.php/loading/article/download/311/313/2341

17. Processes and Idleness in Europa Universalis 4 \- Game Philosophy Network, https://gamephilosophy.org/wp-content/uploads/confmanuscripts/pcg2017/Blacha%20-%202017%20-%20Processes%20and%20Idleness%20in%20Europa%20Universalis%204.pdf

18. Digital Humanities Archives \- Not Even Past, https://notevenpast.org/tag/digital-humanities/

19. Public Access to Our Public Presence: Sharing Our API | The Art Institute of Chicago, https://www.artic.edu/articles/902/public-access-to-our-public-presence-sharing-our-api

20. Public API | The Art Institute of Chicago, https://www.artic.edu/open-access/public-api

21. Art Institute of Chicago API: Documentation, https://api.artic.edu/

22. a tale of two cities \- The JSTOR Understanding Series, https://labs.jstor.org/understand/work/tale-of-two-cities

23. arXiv daily: Computation and Language (cs.CL) \- Science Cast, https://sciencecast.org/podcasts/arxiv\_daily/computation-and-language

24. EDI Resources \- Cultural Heritage Terminology Network, https://culturalheritageterminology.co.uk/resources/

25. Research \- Chicago History Museum, https://www.chicagohistory.org/education/chicago-metro-history-day/research/

26. Genealogy Resources at Acorn Public Library, https://acornlibrary.org/genealogy/

27. Explore Chicago Collections | The Advancing Genealogist, https://advancinggenealogist.com/2015/10/explore-chicago-collections/

28. Adaptive Quality-Diversity Trade-offs for Large-Scale Batch Recommendation \- arXiv, https://arxiv.org/pdf/2602.02024

29. Exploration Space Theory Formal Foundations for Prerequisite-Aware Location-Based Recommendation \- arXiv, https://arxiv.org/html/2603.06624v1