[python-package] Seeing unexpected behaviour in refit()

[Abir-lal-Roy]

Description

Hi team,
I'm currently working on a classification task that involves incremental learning .To update the model with new data I am using refit() function.Now given an instance X=[x1,x2,x3,...,xn],and tree_index, I am seeing X falling into different leaf node for initial and updated model which is not expected as refit does not change the structure of the main model.Can somebody please help me understand why it is happening ?

Reproducible example

Updating main model

updated_model = initial_model.refit(data = df_train[initial_model.feature_name()],label = df_train["target"],
                                  weight=df_train["sample_weights"],decay_rate=0.95) 

example = df_test[df_test.index==0].values 
l1 = initial_model.predict(example[fixed_model.feature_name()],num_iteration=1,pred_leaf=True) # get the leaf node of 1st tree of main model
l2 = updated_model.predict(example[updated_model.feature_name()],num_iteration=1,pred_leaf=True) # get the leaf node of 1st tree of updated model
l1,l2
(array([177], dtype=int32), array([187], dtype=int32)

I was expecting l1 and l2 to be equal

Environment info

LightGBM version : 4.4.0
python version : 3.8.3

[jameslamb]

Thanks for using LightGBM.

You're right that given two Booster objects, one of which was called by running .refit() on the other, the same sample should fall into the exact same leaf nodes by default. This is because .refit() preserves the tree structure.

I tried tonight with the latest development version of lightgbm (a9df7f1) and observed exactly that behavior:

import lightgbm as lgb
import pandas as pd
from sklearn.datasets import make_regression

X, y = make_regression(n_samples=10_000, n_features=5, random_state=708)
df_train = pd.DataFrame(X)

bst = lgb.train(
    train_set=lgb.Dataset(
        data=df_train,
        label=y
    ),
    params={
        "objective": "regression",
        "num_iterations": 10,
        "seed": 708,
        "deterministic": True,
        "force_row_wise": True,
        "num_threads": 1
    }
)

bst_new = bst.refit(
    data=df_train,
    label=y,
    decay_rate=0.95
)

Chose a sample and saw it fall into the same leaf nodes in both models' first tree:

example_case = df_train.head(1)

orig_leaf_index = bst.predict(example_case, num_iteration=1, pred_leaf=True)
new_leaf_index = bst_new.predict(example_case, num_iteration=1, pred_leaf=True)
print(f"leaf indices: old={orig_leaf_index}, new={new_leaf_index}")
# leaf indices: old=[10], new=[10]

But with different raw predicted values:

bst.predict(example_case, num_iteration=1, raw_score=True)
# array([-12.81107821]

bst_new.predict(example_case, num_iteration=1, raw_score=True)
# array([-17.67358137])

environment info (click me)

• M2 Mac, macOS Sonoma 14.4.1
• Python 3.11.9

conda env export:

name: lgb-dev
channels:
  - conda-forge
dependencies:
  - atk-1.0=2.38.0=hd03087b_2
  - aws-c-auth=0.7.22=hec39e38_2
  - aws-c-cal=0.6.14=h5db4892_1
  - aws-c-common=0.9.19=h99b78c6_0
  - aws-c-compression=0.2.18=h5db4892_6
  - aws-c-event-stream=0.4.2=h5eab607_12
  - aws-c-http=0.8.1=had10953_17
  - aws-c-io=0.14.8=hb5a7b21_5
  - aws-c-mqtt=0.10.4=h78534b8_4
  - aws-c-s3=0.5.9=h1755d02_3
  - aws-c-sdkutils=0.1.16=h5db4892_2
  - aws-checksums=0.1.18=h5db4892_6
  - aws-crt-cpp=0.26.9=h03bff2b_0
  - aws-sdk-cpp=1.11.329=hb37a6d0_3
  - brotli=1.1.0=hb547adb_1
  - brotli-bin=1.1.0=hb547adb_1
  - bzip2=1.0.8=h93a5062_5
  - c-ares=1.28.1=h93a5062_0
  - ca-certificates=2024.7.4=hf0a4a13_0
  - cairo=1.18.0=hd1e100b_0
  - certifi=2024.6.2=pyhd8ed1ab_0
  - cffi=1.16.0=py311h4a08483_0
  - cfgv=3.3.1=pyhd8ed1ab_0
  - click=8.1.7=unix_pyh707e725_0
  - cloudpickle=3.0.0=pyhd8ed1ab_0
  - cmakelint=1.4.3=pyhd8ed1ab_0
  - colorama=0.4.6=pyhd8ed1ab_0
  - coverage=7.5.0=py311hd23d018_0
  - cpplint=1.6.0=pyhd8ed1ab_0
  - cycler=0.12.1=pyhd8ed1ab_0
  - distlib=0.3.8=pyhd8ed1ab_0
  - exceptiongroup=1.2.0=pyhd8ed1ab_2
  - expat=2.6.2=hebf3989_0
  - filelock=3.14.0=pyhd8ed1ab_0
  - font-ttf-dejavu-sans-mono=2.37=hab24e00_0
  - font-ttf-inconsolata=3.000=h77eed37_0
  - font-ttf-source-code-pro=2.038=h77eed37_0
  - font-ttf-ubuntu=0.83=h77eed37_1
  - fontconfig=2.14.2=h82840c6_0
  - fonts-conda-ecosystem=1=0
  - fonts-conda-forge=1=0
  - fonttools=4.53.0=py311hd3f4193_0
  - freetype=2.12.1=hadb7bae_2
  - fribidi=1.0.10=h27ca646_0
  - gdk-pixbuf=2.42.11=h13c029f_0
  - gettext=0.22.5=h8fbad5d_2
  - gettext-tools=0.22.5=h8fbad5d_2
  - gflags=2.2.2=hc88da5d_1004
  - giflib=5.2.2=h93a5062_0
  - glog=0.7.1=heb240a5_0
  - gmp=6.3.0=hebf3989_1
  - graphite2=1.3.13=hebf3989_1003
  - graphviz=9.0.0=h3face73_1
  - gtk2=2.24.33=h7895bb2_4
  - gts=0.7.6=he42f4ea_4
  - harfbuzz=8.4.0=hbe0f7c0_0
  - icu=73.2=hc8870d7_0
  - identify=2.5.36=pyhd8ed1ab_0
  - iniconfig=2.0.0=pyhd8ed1ab_0
  - joblib=1.4.2=pyhd8ed1ab_0
  - kiwisolver=1.4.5=py311he4fd1f5_1
  - krb5=1.21.3=h237132a_0
  - lcms2=2.16=ha0e7c42_0
  - lerc=4.0.0=h9a09cb3_0
  - libabseil=20240116.2=cxx17_hebf3989_0
  - libarrow=16.1.0=h28dd788_6_cpu
  - libarrow-acero=16.1.0=h00cdb27_6_cpu
  - libarrow-dataset=16.1.0=h00cdb27_6_cpu
  - libarrow-substrait=16.1.0=hc68f6b8_6_cpu
  - libasprintf=0.22.5=h8fbad5d_2
  - libasprintf-devel=0.22.5=h8fbad5d_2
  - libblas=3.9.0=22_osxarm64_openblas
  - libbrotlicommon=1.1.0=hb547adb_1
  - libbrotlidec=1.1.0=hb547adb_1
  - libbrotlienc=1.1.0=hb547adb_1
  - libcblas=3.9.0=22_osxarm64_openblas
  - libcrc32c=1.1.2=hbdafb3b_0
  - libcurl=8.8.0=h7b6f9a7_1
  - libcxx=16.0.6=h4653b0c_0
  - libdeflate=1.20=h93a5062_0
  - libedit=3.1.20191231=hc8eb9b7_2
  - libev=4.33=h93a5062_2
  - libevent=2.1.12=h2757513_1
  - libexpat=2.6.2=hebf3989_0
  - libffi=3.4.2=h3422bc3_5
  - libgd=2.3.3=hfdf3952_9
  - libgettextpo=0.22.5=h8fbad5d_2
  - libgettextpo-devel=0.22.5=h8fbad5d_2
  - libgfortran=5.0.0=13_2_0_hd922786_3
  - libgfortran5=13.2.0=hf226fd6_3
  - libglib=2.80.0=hfc324ee_6
  - libgoogle-cloud=2.24.0=hfe08963_0
  - libgoogle-cloud-storage=2.24.0=h3fa5b87_0
  - libgrpc=1.62.2=h9c18a4f_0
  - libiconv=1.17=h0d3ecfb_2
  - libintl=0.22.5=h8fbad5d_2
  - libintl-devel=0.22.5=h8fbad5d_2
  - libjpeg-turbo=3.0.0=hb547adb_1
  - liblapack=3.9.0=22_osxarm64_openblas
  - libnghttp2=1.58.0=ha4dd798_1
  - libopenblas=0.3.27=openmp_h6c19121_0
  - libparquet=16.1.0=hcf52c46_6_cpu
  - libpng=1.6.43=h091b4b1_0
  - libprotobuf=4.25.3=hbfab5d5_0
  - libre2-11=2023.09.01=h7b2c953_2
  - librsvg=2.58.0=hb3d354b_1
  - libsqlite=3.45.3=h091b4b1_0
  - libssh2=1.11.0=h7a5bd25_0
  - libthrift=0.19.0=h026a170_1
  - libtiff=4.6.0=h07db509_3
  - libutf8proc=2.8.0=h1a8c8d9_0
  - libwebp=1.3.2=hf30222e_1
  - libwebp-base=1.3.2=h93a5062_1
  - libxcb=1.15=hf346824_0
  - libxml2=2.12.6=h0d0cfa8_2
  - libzlib=1.2.13=h53f4e23_5
  - llvm-openmp=18.1.8=hde57baf_0
  - lz4-c=1.9.4=hb7217d7_0
  - matplotlib=3.8.4=py311ha1ab1f8_2
  - matplotlib-base=3.8.4=py311h000fb6e_2
  - munkres=1.1.4=pyh9f0ad1d_0
  - mypy=1.10.0=py311hd23d018_0
  - mypy_extensions=1.0.0=pyha770c72_0
  - ncurses=6.4.20240210=h078ce10_0
  - nodeenv=1.8.0=pyhd8ed1ab_0
  - numpy=2.0.0=py311h4268184_0
  - openjpeg=2.5.2=h9f1df11_0
  - openssl=3.3.1=hfb2fe0b_1
  - orc=2.0.1=h47ade37_1
  - packaging=24.0=pyhd8ed1ab_0
  - pandas=2.2.2=py311h4b4568b_1
  - pandoc=3.1.13=hce30654_0
  - pango=1.52.2=hb067d4f_0
  - pcre2=10.43=h26f9a81_0
  - pillow=10.3.0=py311h0b5d0a1_0
  - pip=24.0=pyhd8ed1ab_0
  - pixman=0.43.4=hebf3989_0
  - platformdirs=4.2.1=pyhd8ed1ab_0
  - pluggy=1.5.0=pyhd8ed1ab_0
  - pre-commit=3.7.0=pyha770c72_0
  - psutil=5.9.8=py311h05b510d_0
  - pthread-stubs=0.4=h27ca646_1001
  - pyarrow=16.1.0=py311h35c05fe_2
  - pyarrow-core=16.1.0=py311hb5ba6a5_2_cpu
  - pycparser=2.22=pyhd8ed1ab_0
  - pyparsing=3.1.2=pyhd8ed1ab_0
  - pytest=8.2.2=pyhd8ed1ab_0
  - pytest-cov=5.0.0=pyhd8ed1ab_0
  - python=3.11.9=h932a869_0_cpython
  - python-dateutil=2.9.0=pyhd8ed1ab_0
  - python-graphviz=0.20.3=pyh717bed2_0
  - python-tzdata=2024.1=pyhd8ed1ab_0
  - python_abi=3.11=4_cp311
  - pytz=2024.1=pyhd8ed1ab_0
  - pyyaml=6.0.1=py311heffc1b2_1
  - re2=2023.09.01=h4cba328_2
  - readline=8.2=h92ec313_1
  - ruff=0.4.7=py311hd374d79_0
  - scikit-learn=1.5.1=py311hbfb48bc_0
  - scipy=1.14.0=py311hceeca8c_0
  - setuptools=69.5.1=pyhd8ed1ab_0
  - shellcheck=0.10.0=hecfb573_0
  - six=1.16.0=pyh6c4a22f_0
  - snappy=1.2.1=hd02b534_0
  - threadpoolctl=3.5.0=pyhc1e730c_0
  - tk=8.6.13=h5083fa2_1
  - toml=0.10.2=pyhd8ed1ab_0
  - tomli=2.0.1=pyhd8ed1ab_0
  - tornado=6.4.1=py311hd3f4193_0
  - typing_extensions=4.11.0=pyha770c72_0
  - tzdata=2024a=h0c530f3_0
  - ukkonen=1.0.1=py311he4fd1f5_4
  - virtualenv=20.26.0=pyhd8ed1ab_0
  - wheel=0.43.0=pyhd8ed1ab_1
  - xorg-libxau=1.0.11=hb547adb_0
  - xorg-libxdmcp=1.1.3=h27ca646_0
  - xz=5.2.6=h57fd34a_0
  - yaml=0.2.5=h3422bc3_2
  - zlib=1.2.13=h53f4e23_5
  - zstd=1.5.6=hb46c0d2_0
  - pip:
      - altgraph==0.17.4
      - annotated-types==0.7.0
      - attrs==24.2.0
      - backports-tarfile==1.2.0
      - build==1.2.1
      - charset-normalizer==3.3.2
      - check-wheel-contents==0.6.0
      - contourpy==1.3.0.dev1
      - delocate==0.11.0
      - docutils==0.21.2
      - idna==3.8
      - importlib-metadata==8.4.0
      - jaraco-classes==3.4.0
      - jaraco-context==6.0.1
      - jaraco-functools==4.0.2
      - keyring==25.3.0
      - lightgbm==4.5.0.99
      - macholib==1.16.3
      - markdown-it-py==3.0.0
      - mdurl==0.1.2
      - more-itertools==10.4.0
      - nh3==0.2.18
      - pkginfo==1.10.0
      - pydantic==2.8.2
      - pydantic-core==2.20.1
      - pydistcheck==0.7.1
      - pygments==2.18.0
      - pyproject-hooks==1.0.0
      - readme-renderer==44.0
      - requests==2.32.3
      - requests-toolbelt==1.0.0
      - rfc3986==2.0.0
      - rich==13.8.0
      - twine==5.1.1
      - urllib3==2.2.2
      - wheel-filename==1.4.1
      - zipp==3.20.1
prefix: /Users/jlamb/miniforge3/envs/lgb-dev

So to investigate this further, we're going to need your help.

Can you share a minimal, reproducible example? That'd be code that any of us can copy, paste, and run which reproduces the behavior you're seeing. You could use my example above as a starting point.

Your report leaves out a lot of important details, including the content of the training data. You may want to review "How to create a minimal, reproducible example" in the Stack Overflow docs (link) for some tips on how to report software questions like this.

If you're unable to reproduce this behavior, then could you provide the content of both models in text files? You could called .save_model() to produce those files.