astype函数_xgboost 新增的这个函数不错

这段时间踩坑太多，准备写点文章来转转运，今天先分享几行比较有用的 code 起个头

一、持久化

不废话，直接上code

# xgb原生二进制文件，适用于跨平台使用model.save_model('xgbclf.model')# sklearn 框架生成的二进制文件，与 xgb 原生的有些不一样joblib.dump(model,'xgbclf.model')

注：本文采用的是sklearn 框架，采用原生框架的稍微改改code就可以运行啦

二、可解释性

可读的 xgb 树

model.get_booster().dump_model('xgbclf.text')

可分析的 xgb 树

注: 需要 0.82 及之后的版本

xgbtree = model.get_booster().trees_to_dataframe()

单样本解析

模型的特征重要度，单个样本的预测概率和特征重要度都可以从下方函数中得到

def _xgb_tree_leaf_parse(xgbtree,nodeid_leaf):    '''给定叶子节点，查找 xgbtree 树的路径    '''    leaf_ind=list(nodeid_leaf)    result=xgbtree.loc[(xgbtree.ID.isin(leaf_ind)),:]    result['Tag']='Leaf'    node_id=list(result.ID)    while len(node_id)>0:        tmp1=xgbtree.loc[(xgbtree.Yes.isin(node_id)),:]        tmp2=xgbtree.loc[(xgbtree.No.isin(node_id)),:]        tmp1['Tag']='Yes'        tmp2['Tag']='No'        node_id=list(tmp1.ID)+list(tmp2.ID)        result=pd.concat([result,tmp1,tmp2],axis=0)    return resultdef xgb_parse(model,feature=None):    '''给定模型和单个样本，返回该样本的xgbtree树路径以及该样本的特征重要度    '''    feature_names=model.get_booster().feature_names    #missing_value=model.get_params()['missing']    f0=pd.DataFrame({'GainTotal':model.feature_importances_,'Feature':feature_names})    f0=f0[['Feature','GainTotal']]    xgbtree=model.get_booster().trees_to_dataframe()    if feature is None:        return xgbtree,f0.sort_values(by='GainTotal',ascending=False).reset_index(drop=True)    ind=model.get_booster().predict(xgb.DMatrix(feature),validate_features=False,pred_leaf=True)[0]    ind=pd.Series(np.arange(model.n_estimators)).astype(np.str)+'-'+pd.Series(ind).astype(np.str)    result=_xgb_tree_leaf_parse(xgbtree,ind)    result=result.sort_values(by=['Tree','Node'])    loc=int(np.where(result.columns=='Feature')[0][0])+1    result.insert(loc,'FeatureValue',result.Feature.replace(feature.to_dict()))    #result.loc[(result.FeatureValue==missing_value)|(result.FeatureValue.isnull()),'Tag']='Missing'    result=result[['Tree','Node','ID','Feature','FeatureValue','Split','Tag','Yes','No','Missing','Gain','Cover']]    f_=result.groupby('Feature')['Gain'].mean().drop('Leaf',axis=0)    f_=f_/f_.sum()    f_=pd.DataFrame(f_).reset_index()    f=pd.merge(f0[['Feature','GainTotal']],f_[['Feature','Gain']],on='Feature',how='left').fillna(0)    f['diff']=np.round((f['Gain']-f['GainTotal'])/f['GainTotal'],2)    f=f[['Feature','Gain','GainTotal','diff']].sort_values(by='Gain',ascending=False).reset_index(drop=True)    return result,f

ps：哈哈，没想到我也有流量变现的资格啦，就是那个文中广告真的很影响用户体验。。。