A methods program in online review analytics for graduate programs, research teams, analysts, and data-curious operators. You run each method yourself on a practice dataset, learn to interpret the output correctly, then apply it in two registers: once as a researcher, where it becomes a results paragraph, and once as an analyst or operator, where it becomes something that changes on Monday. The program is a published book, fourteen narrated video lessons, a full set of per-chapter lecture decks, and a practice dataset. Lesson one and the mini-project are free and need no account.
Start lesson one freeOr run the mini-project hereGet the book
This is the first end-to-end analysis in the book, and it comes before any advanced method. In about thirty minutes you load sample review data, summarize it, find what positive and negative reviewers talk about, measure sentiment, and turn the results into recommendations. You do not need to write code. You type plain requests to Gemini and read the output.
Seven steps in all. The first three are below in full, exactly as the book gives them. Open Google Colab with Gemini enabled, download the mini-project file below, upload it to Colab, and type. The prompts name that exact file, so nothing needs adjusting; if you use your own data instead, just swap the filename in Step 1.
What you should see: a table of the first 5 rows, a list of column names, and a single number showing how many reviews are in the file.
What you should see: a table of rating counts, a bar chart showing the distribution visually, and a short paragraph in everyday language describing what the pattern means.
What you should see: a ranked list of 15 words with their counts, showing the most common terms in low-rated reviews.
Steps 4 through 7 continue the same way: the word list for four-star and five-star reviews, a sentiment score computed for every review, sentiment compared against star rating, and three concrete recommendations a manager could act on. That last step is the whole program in miniature, a number you produced yourself turned into something someone can do. The book carries all seven with an answer key, a teaching note on why sentiment and stars do not line up perfectly, and a fix for each step that can misbehave.
Maple City is the book's companion dataset: a fictional mid-sized city whose restaurant reviews carry the whole methods sequence, from text measurement through panel regression, forecasting, spatial diagnostics, and prediction. It is fully synthetic and built with known ground truth, so any result can be checked against the effect it was meant to recover. It is free to copy and reuse.
Download the mini-project fileThe full Maple City practice dataset, the one the book’s worked examples run on, is here too: maple_city_reviews.csv.
Run the method on the practice dataset. Read the output correctly, which is the part most courses skip. Then apply it twice: once as a researcher, where the finding becomes a results paragraph, and once as an analyst or operator, where it becomes a decision. The book is built this way. Chapter 6 does not stop at the regression table; it writes the same finding out for a journal and for hospital leadership, one after the other. Chapter 12 asks you to restate a case finding as a two-sentence executive summary a non-technical manager could act on.
Here is one fixed-effects result from Chapter 6, read both ways. Same model, same numbers, two registers.
"Fixed-effects panel regression on 1,847 hospital-quarter observations reveals that average review sentiment is a significant positive predictor of hospital star ratings (β = 0.342, SE = 0.089, p < 0.001), controlling for review volume, management response rate, and topic composition."
"A meaningful improvement in patient sentiment is associated with roughly a third of a star increase in ratings, controlling for other factors. We identified the specific complaint themes that track with lower ratings: wait times, staff interactions, and cleanliness."
No jargon in the second one, no p-values, concrete themes, and a place to start. Producing both from one result you ran yourself is what the program trains, and it trains it on every method, not only regression.
This is the self-serve level of the discipline, taught honestly as that. It does not hand anyone a production analytics system, and the book is candid about the limits: when the data are thin, a careful descriptive analysis paired with one well-chosen method is the honest move, not the full method stack.
Analyzing Online Reviews with AI: A No-Code Guide from Raw Data to Research Findings, by Jeong-Yeol Park. Its claim is that reviews are multi-layer behavioral data that need an integrated analytical system: text features feed panel regressions, temporal patterns inform causal identification, geographic structure reveals spillovers, and latent mechanisms explain why the patterns exist. Each method chapter names what it adds that the others cannot.
Available as paperback (ISBN 979-8-1890-3688-0) and Kindle (ASIN B0HBLZ5D4J).
Four chapters and one lab, before any model appears. They set the anatomy of a review, the platform biases baked into it, the working environment, and the vocabulary every later chapter assumes.
01Introduction. Why online reviews demand an integrated analytical framework, and how five methods usually taught apart combine into one system.
02Understanding Online Review Data. What each component of a review measures, what it does not, and why review data tells you only what review-writing customers chose to say.
03Getting Started with Google Colab and Gemini. Your AI research environment: account, settings, dataset uploaded, first notebook, first prompt.
04Thinking Like a Data Analyst Before Analyzing Text. The vocabulary the rest of the book assumes: what a dataset is, what a variable is, and the difference between description, prediction, explanation, and causality.
LabMini-Project: Analyze 500 Reviews in 30 Minutes. Your first end-to-end analysis, before any advanced method.
Methods are taught as complements rather than rivals. Several chapters open by telling you plainly whether to read them now or come back later, which is how a methods text should treat a reader's time.
05From Text to Numbers. Every advanced method later in the book depends on this one step: turning review text into valid variables.
06Panel Regression with Integrated Features. Read this chapter if you want to know whether changes in review features are associated with changes in business outcomes over time.
07Causal Inference with Review Data. Read this chapter if you want to claim that one change caused another. Skip it for now if your goal is description, association, or prediction.
08Time-Series Analysis and Forecasting. Read this chapter if your reviews are organized over time and you want to detect trends, seasonality, or early warning signals.
09Spatial Econometric Models and Visualization. Read this chapter if your research question involves geography, neighborhoods, destinations, or spillovers across nearby businesses.
10Structural Equation Modeling for Theory Testing. Interpretation-heavy and code-light: factor loadings, path coefficients, fit indices, indirect effects, and multi-group analysis.
11Machine Learning for Prediction. Core sections first: supervised learning, train and test split, random forest, performance metrics, and interpretation.
Real analytical problems do not arrive labeled with methodological tags. The last two chapters show the methods combining on whole problems, then look at what is changing in the field and what to distrust about it.
12Case Studies. Four integrated cases showing how the methods of Chapters 5 through 11 combine in practice, and why each case moves from one method to the next.
13Where Review Analytics Is Heading. From static to dynamic analysis, text-only to multimodal, isolated reviews to reviews joined with other organizational data, and description to prediction and causal explanation. It also gives the protocol for validating a language-model classifier before trusting it.
For university programs and research groups adopting the book as a methods text: a complete instructor companion of fourteen lecture decks, one per chapter plus the mini-project lab, built to the book's own structure. Every chapter deck carries hands-on exercises and key terms, so a class period ends with students having done something, not only heard it.
The chapter objectives carry the second register too. Chapter 6 asks students to translate regression results into narratives for academic papers, business leaders, and policymakers; Chapter 12 asks them to turn statistical findings into recommendations a practitioner could act on.
Write to us on institutional letterhead at [email protected] and we will send the sample chapters.
The individual path is the book, the practice dataset, and the mini-project as your first sitting. Do that analysis end to end, then work the chapters in the order your question needs: text measurement first, then whichever of panel regression, causal design, forecasting, spatial diagnostics, theory testing, or prediction your data can actually support.
If you will never write a paper, take the second register alone. Analysts and operators can run the same methods, read the same output, and stop at the decision. The mini-project is your entry either way.
Join the list for the practice materials and program updates.
Start with the mini-project on this page, whether the output is headed for a manuscript or for Monday's staff meeting. Institutions asking about adoption and the instructor companion should write on letterhead to [email protected]. Individuals, teams, and analysts can join the list for the practice materials.