What Recommendation Design Taught Me: “Data Preparation Comes Before Algorithm Selection” [Technovate Thinking ⑥ Day3 Part2]

This is part 6 of my 12-part series documenting what I learned in “Technovate Thinking,” a course I took at business school.

The pre-class assignment for the second half of Session 3 was to consider “what kind of recommendations would be most effective for users on an e-commerce site.” A sharp turn from implementing bubble sort in the first half — this was about applying algorithmic thinking to a business challenge.

With bubble sort, there was a clear right answer: “sort numbers in order.” But recommendation design has no single right answer. Defining “what counts as effective” is itself a business judgment — technology alone won’t give you the answer. This shift in framing was the biggest revelation of this session.

Starting with the Definition of “Effective Recommendations”

The first thing I thought about was the definition itself — what does “effective recommendation” actually mean?

A vague “they’re good at suggesting stuff” doesn’t lead anywhere productive. To maximize both user satisfaction and purchase rates, I identified three essential experiences:

• “It fits me”: Precise suggestions based on preferences and purchase history
• “It’s a new discovery”: Encountering products I wouldn’t have searched for but that resonate
• “I can choose without hesitation”: Curation that prevents choice fatigue from too many options

These three are inherently in tension with each other. Push “fits me” too far and new discoveries vanish. Over-curate the options and you can’t leverage personalization accuracy. Recommendation design is the “business judgment” of how to balance these trade-offs — and recognizing that upfront was a major step forward.

From Bubble Sort to Recommendation Design — A Paradigm Shift

In the first half, bubble sort had a clear goal: “sort given numbers in ascending order.” Input and output were defined, and writing the correct algorithm got you to the right answer.

Recommendation design, however, requires you to define “what to recommend” from scratch. Even for the same user, the design changes entirely depending on whether you “surface products similar to past purchases,” “suggest items from categories they’ve never explored,” or “promote products you want to move right now.”

This transition from “problems with a right answer” to “problems where you define the right answer yourself” was a genuine paradigm shift. A pivot from the programming question of “how to structure the steps” to the business question of “what to aim for.” The algorithmic fundamentals I’d built through bubble sort finally began to function as a business tool — and I could feel it happening.

Understanding How Recommendations Work — Key Approaches

Working through the assignment, I learned that there are several established approaches to recommendation.

Collaborative filtering leverages information from users with similar behavior patterns. “People like you also bought…” is its signature approach, and it excels at surfacing products users wouldn’t search for on their own. The catch: it can’t recommend anything for new users or new products with no data — the classic “cold start problem.”

Content-based filtering recommends based on similarity of product attributes (category, description, tags, etc.). Since it only needs product information, it handles new products well and resists cold start issues. However, it tends to surface only similar items, making new discoveries unlikely.

Hybrid approaches combine multiple methods. Accuracy improves, but design and operational complexity increase significantly.

Each method has its strengths and weaknesses. The critical insight: the question isn’t “which method is superior?” but “which method fits your business model and user characteristics?” You choose based on business fit, not technical superiority. This perspective was something Technovate Thinking emphasized repeatedly.

Data Preparation Comes Before Algorithm Selection

After surveying the approaches, the strongest conviction I came away with was: “Before debating which algorithm is best, nothing starts without clean, well-organized data.”

To run any recommendation system, you first need accumulated data — user behavior logs, rating data, product attributes. And that data inevitably contains “dirt.” Missing values from unrated products. Anomalies from bots and unauthorized access. Duplicates from users with multiple accounts. Bias from ratings concentrated on popular items. Run an algorithm on this mess and the recommendations come out full of noise.

In the infrastructure world, there’s a saying: “Get the logging right first — we’ll talk after that.” The recommendation world is exactly the same. No matter how sophisticated the algorithm, if the input data is dirty, the output is garbage. Garbage In, Garbage Out — a truth that holds across every domain.

This insight landed especially hard for someone with my IT infrastructure background. In operations, designing proper logging and building a collection framework always takes priority over tool adoption. When I realized the same principle applies to recommendations, I gained a real conviction: “The principles of the technology world function the same way in a business context.”

Filter Bubbles — When Precision Creates a World Without Discovery

One topic that stuck with me as I researched recommendations was the filter bubble problem.

When a system over-tailors to a user’s preferences, they end up seeing only similar products and lose the chance for new discovery. The more you improve recommendation accuracy, the more the user gets trapped inside the “bubble” of their own tastes.

This isn’t just an e-commerce problem. News apps that only show articles aligned with your views, social media timelines filled with only like-minded opinions — filter bubbles accelerate information bias and can even deepen societal division.

So what’s the fix? One approach is “serendipity design” — deliberately mixing in a certain percentage of recommendations outside the user’s usual preferences. When a music streaming service occasionally drops in a song from a genre you never listen to, that’s this approach in action. It keeps users engaged by preventing monotony while also mitigating the filter bubble.

But mix in too much and users leave because “irrelevant stuff keeps showing up.” This calibration is exactly where technology ends and business judgment begins. Where you set the balance between “precision” and “diversity” is a statement about what the service values most.

Among the three elements of “effective recommendations” I defined earlier, the inherent trade-off between “it fits me” and “it’s a new discovery” is rooted precisely in this filter bubble problem. Pursue precision and you reinforce the bubble; emphasize diversity and precision drops. Which way to lean is a business decision — not something engineers alone can determine.

Organizing with the Technovate Framework: Desired State → Technology → Validation

I used the framework that recurs throughout this course — “Envision the desired state → Select the technology → Validate and improve” — to organize my thinking on recommendation design.

Start with the “desired state.” What customer experience do you want to create? What data is needed for that? Establishing a log collection infrastructure and data quality management is the starting point.

Then “technology selection.” Form hypotheses based on user behavior and choose the right approach from multiple options. Factor in the balance of cost, speed, and accuracy.

Finally, “validation and improvement.” Quantitatively evaluate effectiveness through A/B testing, and adjust the algorithm while monitoring KPIs like click-through rate, conversion rate, and repeat purchase rate. This cycle isn’t a one-time event — it’s meant to run continuously. User preferences shift. Product lineups change by season. A recommendation system you build and forget will quickly lose its edge.

Organizing my thinking through this framework made one thing crystal clear: recommendation design is “business design,” not “technology selection.” “Which algorithm to use” is a means — the starting point is the business question, “What customer experience do we want to deliver?” This sequencing is exactly what Technovate Thinking emphasizes over and over.

Recommendations = Business Strategy — A Management Problem, Not a Technology One

The most memorable takeaway from this assignment: recommendations aren’t a technology problem — they’re business strategy itself.

What recommendations you serve speaks to the very core of “what value does this service provide?” Do you pursue personalization accuracy to maximize purchase rates? Emphasize serendipity to build brand distinctiveness? Or leverage the trustworthiness of popularity rankings to attract new users? Your recommendation strategy is a mirror of your business strategy.

With an IT infrastructure background, it’s easy to frame recommendation engines as a technology selection problem — “do we adopt one or not?” and “which tool do we use?” But the real question is business design: “Through what we recommend, what customer experience are we crafting?”

In the previous post about Transition Diagrams, I learned that “screen transition design isn’t a UI problem — it’s a business strategy problem.” This time, I learned that “recommendation design isn’t an algorithm problem — it’s a business strategy problem.” Behind every technology lies a business decision — that’s the consistent message of Technovate Thinking.

Lesson: “Researching” and “Designing” Are Different Skills

What this assignment drove home: “researching” and “designing” are completely different skills. Surveying algorithm types and case studies is input work. Selecting the optimal approach for a specific service and designing KPIs for validation is output work. Only when you can do the latter does knowledge become actionable for business.

Experiencing a pure algorithm through bubble sort in the first half, then applying it to a business challenge in the second — Session 3’s structure was brilliantly designed. Bubble sort taught me “how to solve the same problem in fewer steps.” Recommendation design gave me practice in “translating a business problem into algorithmic language.” Having these two back-to-back was what made algorithmic thinking finally click as a business tool — this was the session where I first truly felt that connection.

Next Up: The Report Submission Session — A Comprehensive Exercise Using Everything Learned So Far

Session 4 is the report submission round. A comprehensive exercise that mobilizes everything covered so far — programming fundamentals, Transition Diagrams, algorithmic thinking, and recommendation design. I’ll write about that next.

Books Referenced in This Post