Choosey Recommendation Methodology
How each result is produced, what data is used, and what the tools do not decide.
1. Shared principles
Choosey reduces the number of options involved in repetitive daily choices. It does not attempt to judge a visitor's ability or replace personal responsibility. Each tool uses conditions the visitor selects directly. When random selection is part of the process, identical inputs can produce different results so that visitors can discover more than one suitable candidate.
- Remove candidates that do not match the selected conditions.
- When several candidates remain, select one or more from that eligible pool.
- Do not automate medical, legal, financial, or other consequential judgments.
- Allow visitors to retry or adapt a result that does not fit the current situation.
2. Data and result generation by tool
| Tool | Data standard | How results are generated |
|---|---|---|
| Meal picker | 62 menus classified by dinner, late-night food, dessert, and preference tags | Select one candidate matching both meal type and preference |
| Weather outfits | 90 outfit records divided by gender selection and five temperature bands | Use Open-Meteo temperature and display three matching outfits |
| Activity picker | 100 missions classified by five energy levels and indoor or outdoor location | Select an exact match, then broaden to the same energy level if necessary |
| Decision-speed test | Ten everyday preference questions with per-question time limits | Convert response time—not answer content—into 0–20 points per question |
| Animal-face test | Nine appearance labels in a browser image model | Display the three labels with the highest relative classification scores |
3. Weather and image processing
With location permission, the outfit tool sends latitude and longitude to Open-Meteo to retrieve current, daily-low, and daily-high temperatures. A fallback keeps the flow usable when permission or retrieval fails. Current classification centers on temperature; humidity, wind, and rain are not included in the automatic category calculation and should be applied manually by the visitor.
A photo selected for the animal-face test becomes input to a TensorFlow.js model inside the browser. Choosey does not implement a server upload endpoint for storing that photo. The output is entertainment-oriented visual classification rather than identification or personality analysis, and it can change with lighting, angle, filters, and training-data composition.
4. Random values versus measured values
Randomness is used to choose eligible meal and activity candidates. The 60–99% success rate on an activity result card is also a playful display, not a statistic calculated from user history. In contrast, the decision-speed score converts locally measured response times through fixed thresholds. Each guide identifies these differences so that a decorative number is not mistaken for analysis based on collected behavior.
5. Review and failure handling
- Validate required data fields and referenced images.
- Check that Korean and English pages provide the same function without mixed-language results.
- Provide a way back when no candidate matches or an external request fails.
- Review error and correction reports submitted through the Firestore-backed QnA form.
6. What visitors must verify
A recommendation is a starting point for organizing current conditions. The tools cannot fully know allergies, health, travel safety, dress codes, actual weather, or budget. Check whether a result fits the real situation before acting, and prioritize qualified advice for consequential decisions.
Technical references: TensorFlow.js, Teachable Machine, and Open-Meteo.