CN111753995A - A Locally Interpretable Method Based on Gradient Boosting Trees - Google Patents

Abstract

The invention discloses a local interpretable method based on gradient boosting tree, which uses knowledge distillation to obtain a gradient boosting tree model from a complex model. The weighted average of the contribution of the information gain of the nodes, and the ranking of the importance of the input features can be locally interpretable, so as to explain the complex model. The invention is a general interpretable method, which can extract and explain data sets in various fields, such as natural language processing data sets, image data sets and table data sets. At the same time, the method can be extended to obtain the global interpretation of the model by using the method of sub-module selection using the local interpretation.

Description

A Locally Interpretable Method Based on Gradient Boosting Trees

technical field

The invention relates to the field of artificial intelligence, in particular to a locally interpretable method based on gradient boosting tree, which is applied to extract and explain various artificial intelligence models.

Background technique

As machine learning models are increasingly used in key areas such as self-driving cars, healthcare, financial markets, and legal systems, it becomes critical for humans to understand the predictions made by machine learning algorithms. Many complex models, such as deep neural networks and ensemble learning, are fine-tuned to optimize prediction accuracy, which makes it difficult to interpret predictions. Explainable machine learning addresses this problem in two directions. The first approach attempts to build intrinsically interpretable models based on decision trees (sets or rules), GAMs (Generalized Additive Models), logistic regression, etc., which often face the need to reduce prediction accuracy. Another approach provides a global understanding of the entire model or a local interpretation of individual predictions. Some interpretation methods are model-independent and can be applied to any classifier or regressor, while others are designed for specific models. The form of explanation varies from functional importance to decision sets or rules.

The field of explainable machine learning has recently attracted an increasing number of researchers. With the renaissance of deep learning, understanding complex neural networks has become increasingly difficult. Deep neural networks are still challenging because they typically contain a large number of hidden layers and parameters, as well as correlated activity features residing on the hidden layers. Meanwhile, GBM (Gradient Boosting Machine) is a powerful overall learning algorithm that demonstrates its competitive performance on many tasks such as online advertising. Boosting is a powerful supervised learning method that enhances the predictive performance of a model by iteratively refining and combining multiple weak learners (usually decision trees). Gradient boosting generalizes the boosting method to arbitrary differentiable loss functions, which can be used for regression and classification problems. In practice, GBM works well in many application domains and is supported by many publicly available implementations. Learning competitions like Kaggle are all tree-based gradient boosting methods, especially LightGBM, XGBoost, etc. Probably one of the most popular base learners for GBM is the fixed size CART (Classification and Regression Trees), resulting in GBDT (Gradient Boosted Decision Trees, also known as Gradient Tree Boosting). In the present invention we focus on explaining a single prediction of tree-based GBM in terms of functional importance. For tree-based ensemble methods, although decision trees are relatively easy to understand, the final additive model becomes less transparent after model ensemble.

Recent advances in model-agnostic explanation methods can be used to explain ensemble methods. Model-agnostic explanation methods treat the target model as a black box, thus being able to explain any classifier or regressor. Existing work on model-agnostic methods is often post-hoc analysis of a given black-box model fit to the data. A common approach is to learn another model that approximates the predictions of the original model and is relatively easy to interpret. Earlier work approximated the original prediction globally, while more recent methods, such as LIME and Anchor, were able to obtain locally interpretable models for individual examples. Most model-independent methods perturb the input instance for interpretation according to some perturbation distribution that assigns the most likely important features to the prediction. For complex models, it is often difficult to explain the behavior of the model globally using simple interpretable sets or rules formed by selected important features. Likewise, the interpretation of the entire model may not perfectly explain individual predictions. So in this case it is better to use a native interpretation method with concise and interpretable instructions. To further evaluate the entire model, there is an option to apply the explanations generated by a subset of the input to unknown instances.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a local interpretable method based on gradient boosting tree. By improving the feature importance calculation method of the integrated model, the interpretability of the model can be improved. The original complex model is explained.

The concrete technical scheme that realizes the object of the present invention is:

A locally interpretable method based on gradient boosting trees, characterized in that the method includes the following specific steps:

Step 1: Use the training data set to train the parameters of the initial complex model, and extract the input features;

Step 2: Perform knowledge distillation on the trained model to obtain the soft label output of the input features;

Step 3: Use the input features obtained in step 1 and the output soft labels obtained in step 2 to train the gradient boosting tree model to obtain a trained gradient boosting tree model;

Step 4: Extract the feature importance from the trained gradient boosting tree model, sort the feature importance, and select the feature with higher feature importance as the interpretation of the initial complex model.

The training data set described in step 1 is a natural language data set, an image data set and a table data set; the initial model is a long short-term memory network, a convolutional neural network and a multilayer perceptron based on an attention mechanism; the parameter training is performed: The natural language dataset uses an attention-based long-term and short-term memory network; the image dataset uses a convolutional neural network; the tabular dataset uses a multilayer perceptron.

The soft label output of the input feature is obtained by performing knowledge distillation in step 2, and the soft label output formula is:

Among them, Label _soft refers to the soft label output, _zi refers to the final output of the initial model, T is the temperature parameter, i refers to the prediction as the i-th category, and j refers to the total prediction category of the prediction task.

The trained gradient boosting tree model described in step 3 includes M weak discriminators, and each weak discriminator is a decision tree model, wherein M is a parameter of the gradient boosting tree model.

In step 4, the feature importance is extracted from the trained gradient boosting tree model, the feature importance is sorted, and the feature with higher feature importance is selected as the interpretation of the initial complex model, which specifically includes:

The formula for calculating feature importance is:

表示特征P的重要性期望，特征P是由K个数据构成，P_k即为特征的第k个数据；Imp(P_k)即为特征的第k个数据的特征重要性，其中

Imp(P_k)中每个权重γ_mh_m(x)即为训练好的梯度提升树模型中第m个弱判别器对整个模型的贡献程度，

定义为归一化的第m个弱判别器在输入为P_k时的不纯度减少率，不纯度减少率是指弱判别器在预测特征P_k时，节点分割中用到P_k的不纯度减少量占总的不纯度减少量的比值；不纯度的计算是通过特征P_k在决策树模型中经过的划分节点n来计算，即Gain(P_k,n)＝i(n)-p_Li(n_L)-p_Ri(n_R)，其中i(n)表示节点分裂的不纯度，而p_L和p_R分别代表样本分裂后达到n_L和n_R的部分；训练得到的梯度提升树模型中，T_m表示第m个弱判别器，即第m个决策树模型，并用T_m(x)表示输入样本为x时，其中样本x是包含多个特征P，决策树模型T_m在预测时对应的路径；特征P的重要性期望越高表明该特征对于模型决策越重要；将得到的全部特征

按照从大到小排序，以此作为从梯度提升树模型中提取出的解释，同时也作为初始复杂模型的解释。in,

Indicates the importance expectation of the feature P, the feature P is composed of K data, P _k is the kth data of the feature; Imp(P _k ) is the feature importance of the kth data of the feature, where

Each weight γ _m h _m (x) in Imp(P _k ) is the contribution of the mth weak discriminator in the trained gradient boosting tree model to the entire model,

Defined as the impurity reduction rate of the normalized mth weak discriminator when the input is P _k , the impurity reduction rate refers to the impurity of P _k used in node segmentation when the weak discriminator predicts the feature P _k The ratio of the reduction to the total reduction of impurity; the calculation of impurity is calculated by the division node n passed by the feature P _k in the decision tree model, that is, Gain(P _k ,n)=i(n)-p _L i(n _L )-p _R i(n _R ), where i(n) represents the impurity of node splitting, and p _L and p _R represent the part of the sample that reaches n _L and n _R after splitting; the gradient obtained by training In the boosting tree model, T _m represents the mth weak discriminator, that is, the mth decision tree model, and T _m (x) represents when the input sample is x, where the sample x contains multiple features P, and the decision tree model T _m corresponds to the path during prediction; the higher the importance expectation of the feature P, the more important the feature is to the model decision; all the features that will be obtained

Sorted from largest to smallest, as the explanation extracted from the gradient boosted tree model, and also as the explanation of the initial complex model.

The invention is a general interpretable method, which can extract and explain data sets in various fields, such as natural language processing data sets, image data sets and table data sets. At the same time, the method can be extended to obtain the global interpretation of the model by using the method of sub-module selection using the local interpretation.

Description of drawings

1 is a specific flow chart of Embodiment 1 of the present invention;

2 is a frame diagram of an initial image processing model in Embodiment 2 of the present invention;

3 is a frame diagram of an initial model of natural language processing according to Embodiment 1 of the present invention;

Fig. 4 is the initial model frame diagram of table task according to Embodiment 3 of the present invention;

Figure 5 is a flow chart of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the invention.

The present invention provides a local interpretation algorithm based on a gradient boosting tree model. By calculating the relative importance of nodes passed through in the process of sample prediction, sorting is performed to obtain the importance ranking of input features to obtain local interpretability. The invention is a general interpretable method, which can extract and explain data sets in various fields, such as natural language processing data sets, image data sets and table data sets.

The process of the present invention is as shown in Figure 5, including initial complex model training, extraction of input features and output soft labels, gradient boosting tree model training, extraction of feature importance, and ranking of feature importance to generate explanations.

First, the original sample data is divided into training set and test set; secondly, the original model is trained on the training set, and the soft label output of the original model is extracted by the method of knowledge distillation; then the input and soft label of the training set are used. The output is used to train the gradient boosting tree model; then, on the test set, the feature importance of the sample is calculated using the feature calculation method of the present invention for a single sample, and the interpretation of the sample is sorted.

The present invention proposes a formula for calculating feature importance, as follows:

表示特征P的重要性期望，特征P是由K个数据构成，P_k即为特征的第k个数据；Imp(P_k)即为特征的第k个数据的特征重要性，其中

Imp(P_k)中每个权重γ_mh_m(x)即为训练好的梯度提升树模型中第m个弱判别器对整个模型的贡献程度，

定义为归一化的第m个弱判别器在输入为P_k时的不纯度减少率，不纯度减少率是指弱判别器在预测X时，节点分割中用到P_k的不纯度减少量占总的不纯度减少量的比值；不纯度的计算是通过特征P_k在决策树模型中经过的划分节点n来计算，即Gain(P_k,n)＝i(n)-p_Li(n_L)-p_Ri(n_R)，其中i(n)表示节点分裂的不纯度，而p_L和p_R分别代表样本分裂后达到n_L和n_R的部分；训练得到的梯度提升树模型中，T_m表示第m个弱判别器，即第m个决策树模型，并用T_m(x)表示输入样本为x时(样本x是包含多个特征P)，决策树模型T_m在预测时对应的路径；特征P的重要性期望越高表明该特征对于模型决策越重要；将得到的全部特征

按照从大到小排序，以此作为从梯度提升树模型中提取出的解释，同时也作为初始复杂模型的解释。in,

Indicates the importance expectation of the feature P, the feature P is composed of K data, P _k is the kth data of the feature; Imp(P _k ) is the feature importance of the kth data of the feature, where

Each weight γ _m h _m (x) in Imp(P _k ) is the contribution of the mth weak discriminator in the trained gradient boosting tree model to the entire model,

Defined as the impurity reduction rate of the normalized mth weak discriminator when the input is P _k , the impurity reduction rate refers to the impurity reduction of P _k used in node segmentation when the weak discriminator predicts X The ratio of the total impurity reduction; the calculation of impurity is calculated by the division node n passed by the feature P _k in the decision tree model, that is, Gain(P _k ,n) ₌ i(n)-p Li( n _L )-p _R i(n _R ), where i(n) represents the impurity of node splitting, and p _L and p _R represent the parts that reach n _L and n _R after sample splitting; the gradient boosting tree obtained by training In the model, T _m represents the m th weak discriminator, that is, the m th decision tree model, and T _m (x) represents when the input sample is x (sample x contains multiple features P), the decision tree model T _m is in The corresponding path during prediction; the higher the importance expectation of the feature P, the more important the feature is to the model decision; all the features that will be obtained

Sorted from largest to smallest as the explanation extracted from the gradient boosted tree model, as well as the explanation of the initial complex model.

Example 1

The main parts and implementation strategies of the present invention are described below:

Figure 1 shows the specific process of Embodiment 1, including training of an initial complex model, extracting input features and output soft labels, training a gradient boosting tree model, and extracting feature importance from the gradient boosting tree model and sorting to obtain an explanation. Step 1: Use the training data set to train the parameters of the initial complex model and extract the input features

First, use the natural language processing data set SST2 to construct the training data set, and then design and construct the initial complex model corresponding to the natural language processing task. Its structure is shown in Figure 3, including the word embedding layer, the long short-term memory network layer, and the attention layer , random deactivation layer, fully connected network layer. Use this initial complex model for training on the training dataset. After training the model, the output of the word embedding layer is extracted as the input feature.

Step 2: Perform knowledge distillation on the trained model to obtain the soft label output of the input features

Knowledge distillation is a technique often used in model compression and transfer learning to transfer knowledge from a complex network to another simpler model. For complex models, it is actually very difficult to interpret them directly. In order to be able to explain, knowledge distillation is a very useful technique to distill a model with low interpretability onto a model with higher interpretability. The former explanation can be obtained by explaining it. Therefore, the present invention uses the method of knowledge distillation to extract the soft label corresponding to the input feature from the final output layer of the initial model. The formula of the soft label is:

where _zi is the logits output of the model, T is the temperature of knowledge distillation, and i corresponds to the number of types of prediction tasks. For natural language processing tasks, T is set to 2. The output soft label Label _soft corresponding to the input feature is obtained by calculation. Step 3: Gradient boosted tree training using input features and output soft labels

The input features and the soft labels obtained in step 2 are constructed into a dataset, and the gradient boosting tree model is used for training. For different tasks, the parameters of the constructed gradient boosting tree need to be adjusted accordingly to ensure the trained gradient boosting tree. The accuracy is high enough, and obtaining a high-precision gradient boosting tree model is a prerequisite for extracting a better interpretation of the model. For natural language processing tasks, set the parameter M of the gradient boosting tree to 100, that is, the gradient boosting tree model contains 100 weak discriminators, which is 100 decision tree models.

Step 4: Extract the feature importance from the trained gradient boosting tree model, sort the feature importance, and select the feature with higher feature importance as the interpretation of the initial complex model

按照从大到小排序，以此作为从梯度提升树模型中提取出的解释，同时也作为初始复杂模型的解释，即为局部解释。在自然语言处理任务中，输入的样本为句子，句子中的每个词就是样本的特征，通过上述计算可以得到每个词的特征重要性，以此进行排序可以得到词的重要性排序，重要的词作为该样本的解释。Using the calculation method proposed by the present invention, select any sample from the test set, use the gradient boosting tree model trained in step 3 to predict the sample, and at the same time predict the sample, calculate the feature importance for each feature of the sample. The higher the importance expectation of P indicates that the feature is more important for the model decision; all the features that will be obtained

Sorted from large to small, as the explanation extracted from the gradient boosting tree model, and also as the explanation of the initial complex model, that is, the local explanation. In the natural language processing task, the input sample is a sentence, and each word in the sentence is the feature of the sample. Through the above calculation, the feature importance of each word can be obtained, and the order of the importance of the words can be obtained by sorting. word as an explanation for the sample.

Example 2

The main parts and implementation strategies of the present invention are described below:

Step 1: Use the training data set to train the parameters of the initial complex model and extract the input features

First, use the image processing data set MNIST to build a training data set, and then design and construct the initial complex model corresponding to the image processing task. Its structure is shown in Figure 2, including convolution layer, activation layer, pooling layer, and random deactivation layer , Fully connected network layer. Use this initial complex model for training on the training dataset. Directly use image 2D pixel data as input features.

Step 2: Perform knowledge distillation on the trained model to obtain the soft label output of the input features

For image processing tasks, T is set to 1. The output soft label Label _soft corresponding to the input feature is obtained by calculation.

Subsequent steps are the same as those in Embodiment 1.

Example 3

The main parts and implementation strategies of the present invention are described below:

Step 1: Use the training data set to train the parameters of the initial complex model and extract the input features

First, a training dataset is constructed using the table processing dataset adult, and then an initial complex model corresponding to the image processing task is designed and constructed. Its structure is shown in Figure 4, including a fully connected network layer. Use this initial complex model for training on the training dataset. Use tabular data directly as input features.

Step 2: Perform knowledge distillation on the trained model to obtain the soft label output of the input features

For form processing tasks, T is set to 1. The output soft label Label _soft corresponding to the input feature is obtained by calculation.

Subsequent steps are the same as those in Embodiment 1.

Specific embodiments of the present invention have been described above with reference to the accompanying drawings. However, those skilled in the art can understand that several improvements and equivalent substitutions can be made without departing from the spirit and principle of the present invention. The improved and equivalently replaced technologies and solutions in the claims of the present invention all fall into the protection scope of the present invention.

Claims (5)

1. A local interpretable method based on a gradient lifting tree is characterized by comprising the following specific steps:

step 1: performing parameter training on the initial complex model by using a training data set, and extracting input features;

step 2: knowledge distillation is carried out on the trained model to obtain soft label output of input characteristics;

and step 3: training a gradient lifting tree model by using the input features obtained in the step 1 and the output soft labels obtained in the step 2 to obtain a trained gradient lifting tree model;

and 4, step 4: extracting feature importance from the trained gradient lifting tree model, sequencing the feature importance, and selecting the features with higher feature importance as the explanation of the initial complex model.

2. The gradient spanning tree-based locally interpretable method of claim 1, wherein the training dataset of step 1 is a natural language dataset, an image dataset, and a table dataset; the initial model is a long-short term memory network, a convolutional neural network and a multilayer perceptron based on an attention mechanism; the parameter training is carried out: the natural language data set uses a long-short term memory network based on an attention mechanism; the image dataset using a convolutional neural network; the tabular data set uses a multi-layer perceptron.

3. The gradient-boosting tree-based locally interpretable method according to claim 1, wherein the knowledge distillation performed in step 2 obtains a soft label output of the input features, and the soft label output formula is as follows:

wherein, Label_softIs referred to as soft tag output, z_iThe method refers to the final output of an initial model, T is a temperature parameter, i refers to the prediction of the ith class, and j refers to the prediction class of the total prediction tasks.

4. The local interpretable method of claim 1, wherein the trained gradient-boosting tree model obtained in step 3 comprises M weak classifiers, each weak classifier being a decision tree model, wherein M is a parameter of the gradient-boosting tree model.

5. The local interpretable method of claim 1, wherein the step 4 of extracting feature importance from the trained gradient lifting tree model, sorting the feature importance, and selecting features with higher feature importance as an interpretation of the initial complex model specifically comprises:

the calculation formula of the feature importance is as follows:

wherein,

expressing the importance expectation of a feature P, which is composed of K data, P_kThe k data of the characteristic is obtained; imp (P)_k) I.e., the feature importance of the kth data of the feature, wherein

Imp(P_k) Each weight gamma in_mh_m(x) Namely the contribution degree of the m-th weak discriminator in the trained gradient lifting tree model to the whole model,

the m weak discriminator defined as normalization has an input of P_kThe impurity reduction rate of the time, which is the prediction characteristic P of the weak discriminator_kIn time, P is used in node segmentation_kThe impurity reduction of (a) is a ratio of the total impurity reduction of (b); the purity of the reaction is calculated by the characteristic P_kCalculated by dividing node n through the decision tree model, i.e. Gain (P)_k，n)＝i(n)-p_Li(n_L)-p_Ri(n_R) Wherein i (n) represents the purity of the node split, and p_LAnd p_RRespectively represent that the sample reaches n after splitting_LAnd n_RA moiety of (a); in the trained gradient lifting tree model, T_mRepresenting the m-th weak arbiter, i.e. the m-th decision tree model, and using T_m(x) When the input sample is x, wherein the sample x contains multiple features P, the decision tree model T_mA corresponding path at the time of prediction; the higher the importance expectation of the feature P indicates that the more important the feature is for model decision making; all features to be obtained

And sequencing the models from large to small to serve as the explanation extracted from the gradient lifting tree model and serve as the explanation of the initial complex model.

Images