EPPS 6323 Knowledge Mining | Spring 2026
Group Project
Public Trust in Banks and Financial Institutions: A Knowledge Mining Approach
Made in collaboration with the brilliant minds of Soha Arian, Kasra Akbari, and Eman Ajmal.
Project Proposal
Or Click to Show/Hide Proposal
Pitch Deck
Or Click to Show/Hide Pitch Deck
Mid-Semester Progress Report
Or Click to Show/Hide Deck
Shiny Application | Link
Final Report
Or Click to Show/Hide Report
Individual Assignments
Reflections
AI for Science and the Move from Prompts to Research Workflows | April 21st, 2026
People toss around the phrase “AI for Science” like another buzzword, but it’s not that simple. If anything, it’s a bit sprawling. At its core, using AI not just to crunch numbers or fit models, but to also nudge the entire scientific process forward: reading papers, spotting patterns, sketching hypotheses, even drafting reports. Not just prediction, in other words, but participation. Some recent surveys frame it as covering the full research loop, from literature digging to peer review, which honestly, starts to sound less like a tool and more like a junior (and often overconfident) lab partner - although are not all junior lab partners overconfident? Some founders at the top of the industry, like Dario Amodei, have gone a step further, arguing that the real promise isn’t faster autocomplete but something closer to “expert-level” scientific assistance, the kind that actually helps decide what experiment to run next . A bold conviction, sure. But so have been all the convictions that have come true so far with AI.
Now, here’s where things get interesting. Traditional AI use in research is mostly single-shot: you ask a question, get an answer, maybe tweak the prompt and try again. Done. Agentic systems don’t play that game. They stretch things out. They plan, they loop, they call tools, they revise. Instead of one neat response, you get a whole chain of actions unfolding over time. Almost like watching someone think out loud, except they’re also running code, pulling papers, and scribbling notes while they’re at it. Systems like ResearchAgent or those multi-agent “AI Scientist” setups actually iterate: generate an idea, critique it, refine it, test it, repeat . It’s less “here’s your answer” and more “here’s the process.”
And honestly, that shift, from answers to workflows, changes the rhythm of research. Instead of manually juggling literature searches, data analysis, and write-ups, the agent sort of orchestrates it. Planning modules decide what to do next, tool-use components fetch data or run experiments, memory keeps track of what’s already been tried. There’s even evidence that explicit multi-step planning improves performance compared to reactive, one-step approaches (which, yeah, makes sense if you’ve ever tried to wing a complex project without a plan).
That said, and this part gets glossed over way too often, it’s not like humans get to kick back and sip coffee while the AI does everything. If anything, the role expands rather than shrinks. You’re less of a hands-on executor and more of a supervisor or commanding officer. Setting direction, checking whether the reasoning holds up, catching weird mistakes (and there will be weird mistakes). Even the more autonomous systems still lean on human judgment to keep things grounded and, well, scientifically sane. Because at the end of the day, science isn’t just pattern-matching, it’s about causality, evidence, and being able to explain “why” something works. And no matter how slick these agentic workflows get, that part? Is still stubbornly human.
Knowledge Graphs for Research and Knowledge Mining | April 14th, 2026
People tend to think of data as rows and columns first, like spreadsheets stretching out forever, but a knowledge graph doesn’t really behave that way. It’s more like a web: messy, interconnected, almost alive in how things link together. At its core, a knowledge graph represents information as entities (nodes) and relationships (edges), forming what researchers often describe as triples: subject, relationship, object. That structure matters because it encodes meaning directly, not just values sitting in isolation. In contrast, a traditional relational database stores data in tables that need to be stitched together through joins, while a vector store does something entirely different, it maps meaning into numerical space for similarity search. Useful, sure, but kind of blind to explicit relationships. In other words, databases store facts, vector stores store proximity, and knowledge graphs store context.
Now, why do relationships matter so much? Because that’s where the “knowledge” part actually kicks in. When entities are connected, you can trace paths, uncover indirect links, and ask layered questions that go beyond surface-level queries. A flat dataset might tell you that a company launched a product and that revenue increased later, but it won’t naturally explain how those events relate unless you manually force that connection. A graph, on the other hand, makes those links explicit. You can follow chains like company -> product -> demand > revenue and start to see patterns emerge. It’s a bit like the difference between reading isolated headlines versus understanding the full story behind them, same pieces, totally different depth.
If I were to sketch a small example from a financial research context, I’d probably start simple with three entities: Company, Financial Metric, and Market Event. Then define relationships such as “influences,” “responds_to,” and “drives_change_in.” For instance, a Market Event (like an interest rate hike) influences a Company, which then drives change in a Financial Metric like revenue or margin. Not perfect, maybe a bit simplified, but it works. This kind of structure allows you to ask questions like, “Which companies were most sensitive to rate changes and saw margin compression afterward?” Try answering that cleanly with a flat dataset. You can do it, but its clunky and full of joins and assumptions.
And that’s really the advantage, if I had to boil it down. Knowledge graphs let you ask “why” and “how” questions, not just “what.” They support multi-step reasoning in a way that spreadsheets just don’t. Research has shown that graph-based representations improve tasks like multi-document question answering because they preserve semantic relationships across data points, making it easier to retrieve and connect relevant information . However, building and maintaining them takes effort, and if the relationships are poorly defined, the whole thing can get messy fast. But when done right, they turn scattered data into something that actually resembles understanding.
Retrieval-Augmented Generation and Research Trustworthiness | April 7th, 2026
Half the battle in working with AI is not in the generation of outputs, but validating whether or not you can actually trust said outputs. That is where Retrieval-Augmented Generation (RAG) becomes essential. In plainest terms, RAG is a setup where an AI doesn’t just pull information from its training memory, but the LLM actually goes out, explores and extracts relevant documents, and then builds a response around that material. It less of a solo act and more like a researcher with a stack of notes spread across their desk: messy, but grounded. For research work in particular, that grounding matters. It hedges the model away from confident guesswork and toward something that resembles evidence-based writing.
The real question is, does this actually make AI more trustworthy? The answer is complicated. When a model ties its answers to retrieved sources, you can at least trace where ideas are coming from, which feels reassuring in the same way footnotes do in a decent paper. The “hallucination” problem that plagues AI is reduced by the anchoring of responses by real documents, but it is not entirely eliminated. RAG adds a layer of accountability, where you are no longer dealing with pure improvisation; there is now a paper trail to how the generated output comes to fruition.
That said, people have started poking holes in RAG lately, and not without reason. One common gripe is that retrieval systems can miss the mark, focusing on irrelevant or shallow information from the documents provided; leading the model down the wrong path, defeating the purpose of RAG. Or as the old data science saying goes: “garbage in, garbage out.” There’s also the issue of fragmentation: the AI might stitch together bits from multiple sources, but the result can feel disjointed or oddly stitched. And sometimes, ironically, the presence of sources gives a false sense of authority. Especially if said sources are outdated, biased, or just not very good.
In the end, RAG is not the silver bullet to mastering AI, instead it is another methodology to treat LLMs as they should be - a tool. Better retrieval methods, more context-aware, can also be used in tandem with RAG to create better results. And since RAG is so dependent on solid sources, selecting quality sources, filtered for credibility, is essential before the model begins to run its output.
The Advantages and Limitations of LLMs | March 31st, 2026
LLMs are, at their best, remarkably efficient pattern synthesizers. They can summarize dense material, translate across domains (technical → plain English, for example), generate structured drafts, and even assist in light data cleaning or coding tasks. In research contexts, that’s huge. Hours of initial groundwork—literature overviews, note consolidation, early-stage framing—can be compressed into minutes. Not perfectly, but close enough to move things forward. They’re especially good at starting things. Blank pages, less so.
Where they struggle, though—and this shows up quickly—is in reasoning under uncertainty and maintaining factual reliability. One failure mode I’ve encountered repeatedly is hallucination. Not the dramatic kind—no wild sci-fi nonsense—but subtle, almost believable inaccuracies. For instance, generating a citation that looks entirely legitimate—correct formatting, plausible author names—but doesn’t actually exist. And if you’re not paying attention, it slips through. That’s the problem. It’s not obviously wrong; it’s quietly wrong. The pattern here is that LLMs optimize for coherence, not truth. If something sounds right within the structure of language, the model will often produce it—even when the underlying fact isn’t grounded. So while they’re strong at assembling information, they’re far less reliable at verifying it.
The parts of research that require judgment, context, and accountability will remain human-driven, and probably become more important, not less. Deciding what questions to ask, interpreting ambiguous results, understanding client or market context, and making final decisions under uncertainty—those aren’t easily outsourced. There’s also a trust component. Clients and stakeholders don’t just want outputs; they want reasoning they can stand behind. So while AI will reshape how research is done, it won’t replace the need for human oversight—it will, if anything, raise the bar for it. And maybe that’s the shift: less time spent producing information, more time spent deciding what actually matters.
Predictive Models as Research Assistants | March 24th, 2026
Close reading is intimate. It’s slow, deliberate, almost like sitting with a single painting and noticing every brushstroke. You catch tone, irony, contradictions, the feel of a text. But it doesn’t scale. At all.
Text mining flips that. It doesn’t “read” in the human sense—it scans, aggregates, counts, clusters. And because of that, it can surface patterns that no individual reader would realistically catch. For instance, you could analyze thousands of financial reports and notice subtle shifts in language—say, increasing use of uncertainty-related words before market downturns.But—and this is where it gets tricky—text mining is blind to meaning in a deeper sense. It might flag that a word appears frequently, but it won’t reliably catch sarcasm, cultural nuance, or context-dependent meaning. It can tell you that something is happening linguistically, not always why. So you end up with this tradeoff: scale versus depth. Breadth versus interpretation. Ideally, you use both—but in practice, people tend to lean too hard one way or the other.
Now, bringing NLP and LLMs into a research workflow—this is where things get interesting, and honestly, a bit uneven depending on how they’re used.
For a project like portfolio analytics or financial research, NLP could be used to extract structured insights from unstructured data—earnings calls, analyst reports, SEC filings. For example, you could build a pipeline that identifies sentiment shifts in quarterly earnings transcripts and correlate that with stock performance or volatility. That’s not hypothetical—that’s very doable.
LLMs, on the other hand, are more like… flexible assistants. They can help summarize dense reports, generate draft analyses, or even translate technical findings into client-facing language. Say you’re building a correlation dashboard—you could use an LLM to explain the results in plain English for an advisor or client. Or, more practically, to help clean and standardize messy portfolio input data. Not perfectly—never perfectly—but faster than doing everything manually.
Still, I’d hesitate to fully rely on them, for they are limited by hallucinations, the inability to decipher meaning (only patterns), and bias. This is why it is essential to make sure the data being fed into your models does not predispose the model to bias, and why verification of findings is essential when working with predictive models.
Prediction vs Explanation and Financial Markets | March 10th, 2026
It’s one of those questions that sounds clean in theory but gets murky fast in practice -prediction versus explanation. Sometimes, honestly, you don’t need to know why something happens to act on it. Take credit risk models, for example. A bank might use a model that predicts the likelihood of default with high accuracy, even if the internal logic is complex or not easily interpretable. From a purely operational standpoint, that can be enough - decisions get made, losses get managed. But shift into a policy or advisory context, say financial planning for clients, and suddenly explanation matters a lot more. If a portfolio underperforms, it’s not enough to say “the model predicted this outcome.” Clients (and regulators) expect a rationale. Why did it happen? Was it market conditions, asset allocation, timing? Prediction gets you speed; explanation gets you trust. And depending on the setting, one without the other can feel incomplete.
Now, thinking about a simple causal structure in a financial context - say, the relationship between portfolio diversification and portfolio returns - you could sketch it out like this: diversification -> returns. Straightforward, at least on the surface. But almost immediately, confounders start creeping in. Market conditions, for one - bull versus bear markets can influence both how diversified a portfolio is and how it performs. Investor risk tolerance is another; more risk-tolerant investors might both diversify differently and experience different return profiles. Even time horizon plays a role. So what looks like a clean causal link quickly becomes tangled. To distinguish causation from mere prediction, you’d need to isolate the effect of diversification itself - through controlled comparisons, maybe matching portfolios with similar risk profiles, or using quasi-experimental designs. Otherwise, you’re left with a predictive association: diversified portfolios tend to perform a certain way, but you can’t confidently say diversification caused that outcome. And that distinction (subtle, but critical) is where a lot of real-world analysis either holds up, or quietly falls apart.
Association rules, data mining, and model evaluation | March 3rd, 2026
A practical pipeline for Global weapons
If I were building a workable machine-learning pipeline from the Global weapons database, I would begin by respecting what the site actually is. It is not just a neat spreadsheet. It is a large, mixed-format encyclopedia that, as of March 2026, advertises 2,847 catalogued systems across 196 countries and 52 categories, with browse paths by system, country, type, and comparisons. Its listing structure already exposes fields such as system, type, origin, status, and year, and recent snippets show records such as KF-21 Boramae and B-21 Raider with type, country, and status labels attached.
Because the schema is heterogeneous, I would not try to model every weapon class at once on day one. That would be asking for a mess. A bomber page such as B-21 Raider foregrounds fields like wingspan, maximum takeoff weight, unit cost, planned fleet, and status; an air-defense page like Iron Dome emphasizes engagement counts and interceptor cost; a fighter page such as HAL Tejas Mk2 stresses radar range, targets tracked, payload, assembly progress, and delivery timeline; and a fighter page like Chengdu J-20 mixes technical numbers with long-form narrative about engines, production, and avionics. That kind of variation is interesting for a researcher, but it is poison for a first-pass model if you throw it all into one flat matrix without discipline.
So my first pipeline would subset to one category, probably fighter aircraft. From there, I would set a concrete target variable such as status - for example, in service, flight testing, or under development - because status is analytically meaningful and already appears in the site’s listing language. Then I would follow the same broad logic the course text recommends for data-mining projects: sample if necessary, build a rich array of features, reduce that feature space, fit the model on training data, validate it on held-out data, and compare alternative methods (uploaded course text, p. 53). In practical terms, that means: collect fighter pages and listing metadata; normalize units and currencies; parse numeric specs where available; create binary indicators from text for concepts like stealth, AESA radar, sensor fusion, data link integration, or manned-unmanned teaming; one-hot encode origin and subtype; and then split train, validation, and test sets with stratification by status. Since the site advertises API access, I would check that route first; if it proves unusable, I would fall back to a respectful scraper and parser.
For modeling, I would keep the sequence boring on purpose at the start: a regularized logistic model as a baseline, then a tree-based model such as random forest or gradient boosting, and then maybe a text-aware model if the prose fields carry signal that the structured specs miss. After that, I would add one genuinely data-mining layer: association rules on binarized design features. That would not replace prediction, but it could surface recurring design packages that a flat accuracy score would never reveal on its own - things like which capability bundles tend to coexist in aircraft at similar maturity levels. That is where data mining, in the richer sense of the term, starts to become more than just classification.
What error analysis adds
The course text is useful here because it pushes beyond a single score. Accuracy, or overall classification rate, is only one summary of fit. A confusion matrix opens that summary back up and shows which cases are true positives, true negatives, false positives, and false negatives; the same chapter also highlights sensitivity, specificity, and the unavoidable trade-off between false positives and false negatives (uploaded course text, pp. 39-40). In a Global weapons model, that matters a lot. A classifier that looks “pretty accurate” overall might still be systematically confusing flight-testing systems with in-service systems, or doing well on countries with dense documentation while failing on countries with sparse or text-heavy entries. Accuracy by itself smooths over those patterns, which is exactly the problem.
In summary, error analysis showed me where the model was making the wrong kind of mistakes, not just how many mistakes it made. The accuracy score compressed everything into one headline number, but the confusion matrix and subgroup review exposed whether the model was confusing adjacent status categories, relying too heavily on one country or one documentation style, or failing whenever structured specs were missing. That mattered because some errors were understandable ambiguity, while others were signs of shallow pattern-matching. The course text defines overfitting as fitting not only the signal in the training data, but also the noise, which is why a model can look excellent on familiar data and then stumble on new cases (uploaded course text, pp. 32-33). I ended up thinking about overfitting as the difference between knowledge and residue: knowledge is the structure that travels to new data, while noise is the leftover clutter the model memorizes because it happened to be there. Amodei makes a similar point from another angle when he argues that prediction alone does not necessarily yield insight; a model can capture patterns that are hard to understand and still fail to provide real explanatory value. That is exactly why error analysis matters. It tests whether the model learned something durable, or just something convenient.
Taken together, the sources point to a pretty coherent answer. Association rules are a noncausal, pattern-finding method that works especially well on categorical data and is often judged by support, confidence, and lift. Unsupervised learning is one methodological branch inside the larger practice of data mining, not the whole thing. And for a source like Global weapons, the smartest path is not to toss the entire encyclopedia into one model and hope for magic; it is to define a narrow prediction target, engineer comparable features, validate carefully, and then use error analysis to separate transferable structure from sample-specific noise. That last distinction is really the center of the whole exercise. The model is only useful if it learns signal that survives contact with new data.
Knowledge, Information, and AI | February 24th, 2026
It’s tempting to say humans “find” knowledge the way you’d find your keys—look hard enough, eventually it turns up—but that’s not really how it works. It’s messier than that. People stumble into knowledge, argue their way toward it, inherit it, misinterpret it, revise it… sometimes all in the same afternoon. We piece it together through experience, observation, trial-and-error, and, honestly, a fair amount of guesswork dressed up as confidence. Think about how often understanding comes after confusion, not before. A student rereads something three times, suddenly it clicks—was the knowledge always there? Maybe. But it had to be processed, wrestled with a bit. That’s the human part.
And that’s where the difference between information and knowledge starts to show. Information is just… raw material. Facts, data points, numbers on a page, headlines scrolling by faster than anyone can really digest them. Knowledge, on the other hand, is what happens when that material gets interpreted, connected, and—this is key—understood in context. It’s the difference between knowing that a stock dropped 5% and understanding why it dropped, what that implies, and whether it even matters. One is static. The other is alive, evolving. You could say information fills the room; knowledge rearranges the furniture so you can actually move around.
Now, when you look at AI systems—especially modern ones—you start to notice a pattern in where things go sideways. They rarely fail because they lack information. Quite the opposite. The failures tend to cluster around interpretation. For instance: hallucinations (confidently generating false or nonexistent facts), context blindness (missing nuance, sarcasm, or cultural meaning), overgeneralization (applying patterns too broadly), and brittleness when faced with slightly unfamiliar inputs. There’s also the issue of opacity—models producing outputs without clear reasoning pathways—and bias inheritance, where existing data imbalances quietly shape outcomes. None of these are random glitches; they’re recurring themes.
If you step back a bit, a pattern emerges—actually, a few. First, AI tends to confuse correlation with understanding. It’s excellent at spotting patterns, but not always at grasping meaning. Second, it operates without lived experience, which sounds obvious, but has real consequences—it can’t “ground” information the way humans do. Third, it often projects certainty where uncertainty would be more appropriate, which… people do too, to be fair, but AI does it at scale. So in a strange way, AI is overloaded with information but still struggles to produce what we’d comfortably call knowledge. And maybe that’s the core tension: knowing that something is true isn’t the same as knowing why it matters.
Reflection on Assignment 2 | February 17th, 2026
One thing that jumps out immediately across the three responses is that they are all, broadly speaking, answering the same question, but they are doing very different kinds of work. The Grok response leans heavily into critical theory and normative critique, almost treating the review as a conceptual intervention as much as a systematic synthesis. Its strongest contribution is arguably its framing of surveillance as a mechanism of “racial surveillance capitalism,” and its attention to feedback loops, intersectionality, and structural oppression gives the review real interpretive depth . The Copilot draft, by contrast, is the most conventionally “systematic” in form - more exhaustive in scope, stronger on methodology, and much more explicit about review protocols, inclusion criteria, and empirical gaps. It feels closer to something one might actually submit as a structured literature review. The Chat response is more compressed and economical, but interestingly, it sharpens the core mechanisms of oppression into a cleaner analytic framework, which gives it conceptual efficiency even if it sacrifices some breadth .
In terms of validity, they also diverge. Copilot appears strongest on methodological validity because it most clearly approximates PRISMA-like review logic, incorporates broader source coverage, and engages with causal inference through an operationalized hypothesis involving difference-in-differences design . That matters, because systematic reviews live or die partly on transparent method, not just persuasive writing. Grok, however, arguably offers stronger construct validity around the oppression question itself; it captures dimensions of harm that a narrower empirical framing might understate. Chat sits somewhere in between, balancing analytic clarity with reasonable methodological grounding. What’s interesting—maybe a little ironic—is that the “best” output depends on the evaluation criterion. If judged on systematic rigor, Copilot may lead. If judged on theoretical sophistication, Grok has an edge. If judged on concise synthesis, Chat performs surprisingly well.
There are also patterns in what all three emphasize, and those overlaps tell us something important. All three converge on predictive policing, facial recognition, and algorithmic feedback loops as dominant themes. All three identify underrepresentation, bias amplification, and opacity as recurring mechanisms of harm. And all three point to similar research gaps: too little longitudinal evidence, insufficient Global South analysis, weak intersectional modeling, and limited participatory approaches. That convergence increases confidence that these themes are not artifacts of a single model’s framing but rather robust signals in the literature itself. At the same time, there are differences in outputs worth noting. Grok tends to foreground moral-political critique, Copilot leans institutional and policy-oriented, while Chat is more distilled and hypothesis-driven. In a way, they reflect different epistemic styles: interpretive, procedural, and synthetic.
If I had to reflect on what this comparison reveals more broadly, it’s that output quality is not the same as output validity. A response can sound comprehensive and still omit causal rigor; it can be methodologically careful yet flatten theory. That tension matters. In some sense, comparing these models resembles error analysis in machine learning—accuracy alone does not reveal what a system is actually doing well or poorly. Here, the comparison surfaced biases in emphasis, differences in evidentiary structure, and tradeoffs between breadth and depth that a single response would have obscured. And maybe that’s the bigger lesson: triangulating outputs may produce more reliable knowledge than treating any one generated review as authoritative. One model gives you a map. Three, maybe, give you terrain.
Artifical General Intelligence & O’Neil’s Weapons of Math Destruction | Febuary 10th, 2026
AGI (Artificial General Intelligence) is often discussed as though it were an inevitable milestone, but in practice, it remains more of an evolving concept than a concrete achievement. At its core, AGI refers to a form of machine intelligence capable of performing a wide range of cognitive tasks across domains, rather than being confined to narrow, specialized functions. In other words, it is intended to approximate the adaptability and reasoning capacity of human intelligence - though whether it can truly replicate that depth is still an open question. What makes this especially relevant, when considered alongside Cathy O’Neil’s Weapons of Math Destruction, is the recognition that even today’s narrower algorithms already exert significant influence while operating with limited transparency . Extending that dynamic to more generalized systems introduces not just technical challenges, but broader concerns around interpretability and accountability.
In the context of scientific research, AGI holds the potential to fundamentally reshape how knowledge is generated. It could, for instance, accelerate hypothesis formation, automate complex simulations, and uncover patterns within datasets that would otherwise remain inaccessible. That said, this potential comes with a degree of caution. As O’Neil highlights, reliance on opaque computational systems can lead to outcomes that are difficult to interrogate or validate . If AGI-driven tools are adopted without sufficient scrutiny, there is a risk that researchers may begin to accept results without fully understanding the processes behind them. This introduces a subtle but important tension: while AGI may enhance the efficiency and scale of scientific inquiry, it also necessitates a renewed emphasis on critical evaluation to ensure that scientific rigor is not inadvertently compromised.
Breiman Vs. Galit | February 3rd, 2026
Breiman’s “Statistical Modeling: The Two Cultures” and Shmueli’s “To Explain or to Predict?” approach statistical modeling from closely related, yet meaningfully distinct, perspectives. Breiman frames the divide as a fundamental cultural split between data modeling and algorithmic modeling, advocating strongly for the latter as a more pragmatic approach. His argument is direct, emphasizing predictive accuracy and real-world performance over traditional inferential assumptions. In contrast, Shmueli offers a more structured and reconciliatory framework, distinguishing between explanatory and predictive modeling as separate, but equally valid, objectives within statistical practice.
While Breiman’s work challenges the statistical community to reconsider its priorities, Shmueli refines that challenge by formalizing the distinction and clarifying when each approach is appropriate. Breiman’s tone suggests a shift in paradigm, whereas Shmueli’s analysis provides guidance for coexistence and methodological alignment. Taken together, the two articles complement one another: Breiman identifies the problem, and Shmueli organizes the solution space. The combination ultimately reinforces the importance of aligning modeling techniques with the intended goal, whether explanation or prediction, rather than treating statistical methods as interchangeable.
Class Reading List
Note: Reflections are in response to readings listed in order read.
Bowen, J. A., & Watson, C. E. (2024). Teaching with AI: a practical guide to a new era of human learning. Johns Hopkins University Press.
Wickham, Hadley and Garrett Grolemund and. 2023. R for Data Science. Latest online edition (https://r4ds.hadley.nz)
Breiman, Leo. 2001. “Statistical modeling: The two cultures (with comments and a rejoinder by the author)”. Statistical Science, 16(3), pp.199-231.
Wickham, Hadley, Mine Çetinkaya-Rundel, and Garrett Grolemund. 2023. R for data science. O’Reilly Media, Inc. (Online: https://r4ds.hadley.nz) (WRG) (Chapters 1-8)
Silge, Julia & Robinson, David. 2019. Text mining with R: a tidy approach. Latest online edition (https://www.tidytextmining.com/
Attewell, Paul, and David Monaghan. 2015. Data mining for the social sciences: An introduction. Univ of California Press. (AM) Chapters 1 - 4
Young, Cristobal and Katherine Holsteen. 2017. “Model Uncertainty and Robustness: A Computational Framework for Multimodel Analysis” Sociological Methods & Research. Vol. 46(1) 3-40
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics, 4171–4186. (BERT model, a cornerstone for modern contextual language representations)
Chang, Yupeng, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen et al. 2024. “A survey on evaluation of large language models.” ACM transactions on intelligent systems and technology 15, no. 31-45.
Min, Bonan, Hayley Ross, Elior Sulem, Amir Pouran Ben Veyseh, Thien Huu Nguyen, Oscar Sainz, Eneko Agirre, Ilana Heintz, and Dan Roth. 2023. “Recent advances in natural language processing via large pre-trained language models: A survey.” ACM Computing Surveys 56, no. 2: 1-40.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30, 5998–6008. (must read for all language model studies and transformer model)
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics, 4171–4186. (BERT model, a cornerstone for modern contextual language representations)
Linegar, Mitchell, Rafal Kocielnik, and R. Michael Alvarez. 2023. “Large language models and political science.” Frontiers in Political Science 5.
Thapa, Surendrabikram, Shuvam Shiwakoti, Siddhant Bikram Shah, Surabhi Adhikari, Hariram Veeramani, Mehwish Nasim, and Usman Naseem. 2025. “Large language models (LLM) in computational social science: prospects, current state, and challenges.” Social Network Analysis and Mining 15, no. 1: 1-30.
Koroteev, Mikhail V. “BERT: a review of applications in natural language processing and understanding.” arXiv preprint arXiv:2103.11943 (2021).
Molnar, Christoph. 2019. Interpretable Machine Learning: A Guide for Making Black Box Models Explainable. (https://christophm.github.io/interpretable-ml-book/)