Is there a way to map pandas categorical variable to integer

[zyxue]

I have toy model that takes three variables, x1, x2, x3, x3 is a categorical variable with two categories, a and b. The model is trained using Python API,

import lightgbm
import pandas as pd

model = lightgbm.LGBMRegressor(
        alpha=0.5,
        objective="quantile",
    )

df = pd.DataFrame(
    [
        [1, 1.2, 'a', 111],
        [2, 2.3, 'b', 222],
    ],
    columns=["x1", "x2", 'x3', "y"],
).assign(x3=lambda df: df.x3.astype('category'))

model.fit(df[["x1", "x2", 'x3']], df["y"])

with open('/tmp/simple-regression-model.lgb', 'wt') as opened:
    opened.write( model.booster_.model_to_string())

The serialized model file is like

tree
version=v3
num_class=1
num_tree_per_iteration=1
label_index=0
max_feature_idx=2
objective=quantile
feature_names=x1 x2 x3
feature_infos=none none -1:0:1
tree_sizes=223

...

[machines: ]
[gpu_platform_id: -1]
[gpu_device_id: -1]
[gpu_use_dp: 0]
[num_gpu: 1]

end of parameters

pandas_categorical:[["a", "b"]]

As seen, at the end of the file it includes the pandas_categorical. Now, I'd like to serve the model in Java, I've made some demo code working,

    try {
      String modelStr = Files.readString(Path.of("/tmp/simple-regression-model.lgb"));
      LGBMBooster booster = LGBMBooster.loadModelFromString(modelStr);
      String[] featureNames = booster.getFeatureNames();

      System.out.println(Arrays.toString(featureNames));

      float[] input = new float[] {1.0f, 1.0f, 1f};  // 1f for x3 is arbitrarily coded.
      double pred = booster.predictForMatSingleRow(input, PredictionType.C_API_PREDICT_NORMAL);

      System.out.println(pred);

    } catch (Exception e) {
    }

But I'm not sure what's the best way to map x3 (e.g. a or b) to its corresponding integral value properly in Java? In Python, it's handled automatically if the input is a pd.DataFrame.

[shuttie]

AFAIK you need to pass the categorical_feature parameter prior to creating the booster. See this example: https://github.com/metarank/ltrlib/blob/02cec0419ccc83a85837d1235d89ebdf385b274c/src/main/scala/io/github/metarank/ltrlib/booster/LightGBMBooster.scala#L57

[kensinxie]

@zyxue Hi, I encountered the same issue. How did you manage to overcome it?

[zyxue]

I switched to catboost, https://catboost.ai/en/docs/concepts/java-package

[kensinxie]

AFAIK you need to pass the categorical_feature parameter prior to creating the booster. See this example: https://github.com/metarank/ltrlib/blob/02cec0419ccc83a85837d1235d89ebdf385b274c/src/main/scala/io/github/metarank/ltrlib/booster/LightGBMBooster.scala#L57

@shuttie Hi, I do set the params categorical_feature when I train the model, so you can see the model shows that :

[boosting: gbdt]
[objective: binary]
...
[categorical_feature: 1,16,17,18,19,20,21,23]
...
pandas_categorical:[[0, 1, 6, 7, 8, 9, 10, 11, 12, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 28, 30, 34, 38, 39, 40, 41, 43, 44, 45, 46, 74, 77, 78, 79, 83, 107, 111, 112, 113, 126, 138, 141, 144, 145, 154, 160, 161], [0, 1, 2, 3, 4, 5], [0, 1, 2, 3, 4, 5, 6, 7], [0, 1, 2, 3, 4], [0, 1, 2, 3, 4, 5, 6, 7, 8], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48], [0, 1, 2, 3, 4], [-1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9]]

When using Python to train or predict a LightGBM model, I can use the following code:

data[cat_cols] = data[cat_cols].astype('category')
model.predict(data)

However, in Java, I can only pass a double list to the model for prediction. And that meke it different for the same model inferenced in py and java

[bfeif]

I switched to catboost, https://catboost.ai/en/docs/concepts/java-package

@zyxue Do you have some example code of this you could share?