[python-package] plot_importance() fails when model has 0 splits

[Dixhom]

Description

lgb.plot_importance() throws an error: ValueError: not enough values to unpack (expected 2, got 0)

Reproducible example

import numpy as np
import lightgbm as lgb
from sklearn.model_selection import train_test_split

X = np.array([[  4. , -10.1,  10.4],
        [ 41. , -12.9,  11.3],
        [ 19. , -12.3,  11. ],
        [ 11. ,  -7.3,  11.7],
        [ 36. , -15.3,  11.6],
        [ 15. , -11.6,  10.8],
        [ 35. , -10.6,  10.3],
        [ 38. ,  -8.3,  10.4],
        [ 27. ,  -8.2,   8.8],
        [ 37. ,  -9.2,   9.6]])
y = np.array([4.8, 4.8, 4.7, 5.5, 4. , 5.1, 5.2, 5.4, 5.4, 5.4])

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.5)

lgb_train = lgb.Dataset(X_train, y_train)
lgb_val = lgb.Dataset(X_test, y_test, reference=lgb_train)

params = {
    'boosting_type': 'gbdt',
    'objective': 'regression',
    'metric': 'rmse',
    'random_state': 0,
    'verbose' : -1,
}

model = lgb.train(params,
                lgb_train,
                valid_sets=[lgb_train, lgb_val],
                num_boost_round = 100000,
                callbacks=[lgb.early_stopping(stopping_rounds=1000, 
                                verbose=True)])

lgb.plot_importance(model, figsize=(8,4), max_num_features=5, importance_type='gain') # <-- this throws an error

Environment info

LightGBM version: 4.6.0

Command(s) you used to install LightGBM:

pip install lightgbm

• Windows 11
• VSCode 1.98.1
• Python 3.12.4

Additional Comments

When the data is from from sklearn.datasets import load_diabetes, it does not throw an error. A graph for feature importance is shown.

[Dixhom]

I identified the cause. When the length of data is 49 or less, the error happens. Otherwise, the plot shows up. But I don't know why it's like this.
(update 2025-03-23: This is for diabetes data)

[jameslamb]

Thanks for using LightGBM, and for taking the time to prepare a report with a reproducible example!

Some suggestions for the next time you submit a report to an open source project:

1. Don't forget the "minimal" part of "minimal, reproducible example". Try removing things until you're left with the smallest amount of code necessary to reproduce the problem. For example, doing a train-test split and training for between 1,000 and 100,000 iterations on a dataset with only 10 rows and 3 columns was way too much.
2. Post all of the logs, error messages, and stack traces, not just the one line you think is relevant. Notice in the output I shared below, the stack trace points to an exact code path in lightgbm that helps us to identify where the error is coming from.

I've found these docs helpful, you might too: https://stackoverflow.com/help/minimal-reproducible-example

I'll put up some findings about the report in the next comment.

[jameslamb]

When I ran your example, I noticed that the resulting model has only 1 tree, and 0 splits.

model.trees_to_dataframe()
#    tree_index  node_depth node_index left_child right_child parent_index  ... decision_type missing_direction missing_type value weight count
# 0           0           1       0-L0       None        None         None  ...          None              None         None  5.03   None  None
# 
# [1 rows x 15 columns]

I suspect that could be the problem. I'm able to reproduce this issue on the latest development version, on an M2 Mac, with Python 3.12 and matplotlib==3.10.1.

import numpy as np
import lightgbm as lgb

X = np.array([[  4. , -10.1,  10.4],
        [ 41. , -12.9,  11.3],
        [ 19. , -12.3,  11. ],
        [ 11. ,  -7.3,  11.7],
        [ 36. , -15.3,  11.6],
        [ 15. , -11.6,  10.8],
        [ 35. , -10.6,  10.3],
        [ 38. ,  -8.3,  10.4],
        [ 27. ,  -8.2,   8.8],
        [ 37. ,  -9.2,   9.6]])
y = np.array([4.8, 4.8, 4.7, 5.5, 4., 5.1, 5.2, 5.4, 5.4, 5.4])

model = lgb.train(
    params={
        "objective": "regression",
    },
    train_set=lgb.Dataset(X, label=y),
    num_boost_round=1,
)

lgb.plot_importance(model, figsize=(8,4), max_num_features=5, importance_type="gain")

The logs from lgb.train() show a warning explaining why the model has 0 splits.

[LightGBM] [Warning] There are no meaningful features which satisfy the provided configuration. Decreasing Dataset parameters min_data_in_bin or min_data_in_leaf and re-constructing Dataset might resolve this warning.
[LightGBM] [Info] Total Bins 0
[LightGBM] [Info] Number of data points in the train set: 10, number of used features: 0
[LightGBM] [Info] Start training from score 5.030000
[LightGBM] [Warning] Stopped training because there are no more leaves that meet the split requirements

And the traceback from lgb.plot_importance() shows where in lightgbm the error is coming from:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/jlamb/miniforge3/envs/lgb-dev/lib/python3.11/site-packages/lightgbm/plotting.py", line 132, in plot_importance
    labels, values = zip(*tuples)
    ^^^^^^^^^^^^^^
ValueError: not enough values to unpack (expected 2, got 0)

Looks like somewhere around here, lightgbm is not correctly handling empty models:

https://github.com/microsoft/LightGBM/blob/6437645c4a0c17046be59e4f57d09952e2e0185f/python-package/lightgbm/plotting.py#L121-L132

It looks like this is a bug in lightgbm. I've labeled it that way and changed the title to reflect the problem.

When the length of data is 49 or less, the error happens. Otherwise, the plot shows up. But I don't know why it's like this.

I'm not certain what this refers to, but it generally sounds right to me that if the root cause is "lgb.plot_importance() cannot handle empty models", then adding more data could "resolve" this error by producing a non-empty model. You can run model.dump_model() or model.trees_to_dataframe() to check whether a given model has any splits.

[Dixhom]

@jameslamb Thank you for your support.

When the length of data is 49 or less, the error happens. Otherwise, the plot shows up. But I don't know why it's like this.

This refers to the diabetes dataset. Sorry, I should have explained.

The entire dataset is here (https://limewire.com/d/wfdd4#QUSxLMqCKE).

I ranmodel.dump_model() and got following.

 'tree_info': [{'tree_index': 0,
   'num_leaves': 1,
   'num_cat': 0,
   'shrinkage': 1,
   'tree_structure': {'leaf_value': 5.813157884698165, 'leaf_count': 38}}],

I also ran ` and got below.

tree_index | node_depth | node_index | left_child | right_child | parent_index | split_feature | split_gain | threshold | decision_type | missing_direction | missing_type | value | weight | count
-- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | --
0 | 1 | 0-L0 | None | None | None | None | None | None | None | None | None | 5.852632 | None | None

So it doesn't seem like there is any split in the tree.

[juanf-0gravity]

I can take a look at this. The fix should be straightforward, just need to handle the empty tuples case in plot_importance after the ignore_zero filter.

[jameslamb]

@juanf-0gravity sure, thanks! We'd welcome a contribution.

Please follow the project's existing style and please do ensure this case is covered by the unit tests. Open a draft PR when you have something you think is working locally, and you can @ me on that PR to ask questions about the contribution process.