pandas read_json中pd.CategoricalDtype不生效，'category'生效是否为已知问题？

Great question! This isn't a bug in your code—it's a design limitation in pandas' read_json method for older versions (like your 1.5.2 release). Let's break down what's happening and how to handle it:

What's the Issue? #

When you pass pd.CategoricalDtype directly to the dtype parameter in read_json, pandas doesn't recognize it as a valid type specification for JSON parsing. Under the hood, read_json expects simple string identifiers (like 'category', 'int', 'float') for non-nested columns, not pandas-specific type objects. This is different from methods like astype() that do accept pd.CategoricalDtype instances.

As a result, columns mapped to pd.CategoricalDtype fall back to pandas' default JSON parsing behavior, converting string values to the object dtype instead of category.

How to Fix It #

You have a few options depending on your needs:

1. Stick with the String Identifier ('category') #

This is the simplest approach, and it works exactly as you've seen. If you don't need custom categories or ordering, this is perfect:

df = pd.read_json(
    "./data/data.json",
    dtype={
        "facility": 'category',
        "supplier": 'category'
    }
)

2. Convert After Reading (For Custom Categories) #

If you need to define specific categories or ordered behavior, read the column as object first, then convert it using astype() with pd.CategoricalDtype:

df = pd.read_json("./data/data.json")
# Define your custom categorical type
supplier_dtype = pd.CategoricalDtype(categories=["SupplierA", "SupplierB", "SupplierC"], ordered=True)
df["supplier"] = df["supplier"].astype(supplier_dtype)

3. Upgrade to Pandas 2.0+ (If Possible) #

Newer versions of pandas (2.0 and above) have improved support for type handling in read_json. Many type-related edge cases have been addressed, so upgrading might resolve this limitation if you're able to do so.

Verification #

After applying any of these fixes, run df.info()—you should see category listed for both target columns instead of object.

内容的提问来源于stack exchange，提问作者mon