PART 01
原始样本Raw sample
每个水果还有纹理、产地、照片背景等未被使用的信息。Each fruit also has texture, origin, and image background that remain unused here.
01 · 同一盒水果01 · The same box of fruit
如果只看编号，不看颜色和重量，模型还看得出哪些水果相似吗？If a model sees only IDs—not color or weight—can it still tell which fruit is similar?
02特征选择台Feature Selector
勾选重映射型Toggle and remap
已选特征Selected features
2
个属性attributes
X AXIS
—
第一坐标first coordinate
Y AXIS
—
第二坐标second coordinate
03从样本到特征的 4 步Four Steps from Sample to Feature
信息筛选Information filterPART 02
选择属性Select attributes
任务决定颜色、重量或其他属性是否值得保留。The task determines whether color, weight, or another attribute is useful.
PART 03
编码成数Encode as numbers
颜色类别也要编码，才能进入可计算的坐标位置。Even color categories must be encoded before they become computable coordinates.
PART 04
交给模型Pass to the model
模型只能比较输入特征，不能自动使用你没有提供的信息。A model compares the features it receives; it cannot use information you withheld.
04最小机制：样本经过特征函数Minimal Mechanism: A Sample Passes Through a Feature Function
x = f(sample)x = [f1(s), f2(s)]
上方勾选项决定 f1 与 f2。换特征会换坐标，不会改变原水果。The toggles choose f1 and f2. New features
change coordinates, not the fruit itself.
勾选属性Toggle an attribute可用信息改变Usable information changes
重新编码Re-encode每条样本得到新坐标Each sample gets new coordinates
散点移动Points move相似关系随表示改变Similarity follows the representation
05边界实验：把编号当成大小Boundary Test: Treat an ID as Magnitude
主动制造失败Create a failure只保留编号Keep only the ID
点击后散点按编号排成一条线。视觉上有远近，但编号 90 并不一定比编号 10 “更多”。The points line up by ID. They look ordered, but item 90 is not necessarily “more” than item 10.
尚未执行。数字长得像数量，也可能只是名字。Not run yet. A number may look quantitative while serving only as
a name.
✗ 原始数据里的每一列都是好特征。Every raw column is a useful feature.
→ ✓ 特征必须对当前比较或预测有可解释用途。A feature needs a defensible use for the task.
✗ 数字越大，含义一定越多。A larger number always means more.
→ ✓ 编号可能只是离散标识，没有大小意义。An ID may be a categorical label with no
magnitude.
✗ 模型能自动补回没给的属性。The model can recover any omitted attribute.
→ ✓ 输入表示划定了模型可用信息的边界。The input representation bounds usable
information.
06 · 一句话带走06 · One line to keep