Skip to content
JournalTORIS / FIELD NOTES

← Blog Project notes

What Recommendations Needed First Was Explainable Rules, Not a Model

A look back at LOVETRIP's course recommendations and YM Guide's policy recommendations: what I learned about early-product judgment by separating candidate filtering, scoring, and result explanation.

2 min read Read the Korean original

Joohwan Yu · Founder, Full Stack Developer

Toris · Published

Contents

I worked on recommendation features while building LOVETRIP and YM Guide. Travel courses and youth policies are entirely different subjects, but they shared one thing early on: there wasn’t enough user behavior data. Under this condition, rather than starting right away with learning-based recommendation, we chose to tune rules and scores.

A list that merely matched the conditions was not enough

YM Guide’s earlier records preserve the experience of changing from simple filtering to score-based sorting. Even when we showed every policy that matched the conditions, users still had to decide again where to start looking. In LOVETRIP too, the mere existence of place candidates did not produce a good route.

So we first separated the conditions for excluding candidates from the priorities among candidates. A budget constraint that must be respected and a bonus based on preference play different roles. If you express a condition that must be respected only as a low score, an inappropriate candidate can survive because its other scores are high.

Why we chose rules, and the price

A rule-based approach makes it easy to trace why a result came out. When you find an awkward course in use, you can discuss which condition or weight to change. This was useful at the stage where we were collecting user reactions and revising the product itself.

On the other hand, if you keep adding rules, exceptions pile up. A weight fixed for one case can worsen another, and there is a risk that the developer’s preferences get applied to all users. Being rule-based doesn’t make a result objective. The simplicity of a choice and the sufficiency of verification are separate matters.

If I were doing it now, I’d keep comparison examples

When changing recommendation logic, I would compare before and after results with the same bundle of inputs. That means fixing distinct cases in place, such as a trip with a small budget, a trip with a narrowed travel range, and a policy search with almost no conditions. Even if it’s hard to build a full set of correct answers, it helps to catch results that have clearly gotten worse.

After this experience, I stopped setting a sophisticated algorithm as the first goal of recommendation. What came first was whether users can understand the result and whether I can explain the reason for a change. Leaving the question of moving to a more complex model as a decision for after data and verification criteria exist was the right fit for an early product.

Related posts: YM Guide’s recommendation design, LOVETRIP

Advertisement