def model_train(params):
ins_pred=[]
oos_pred=[]
ins_tgt=[]
oos_tgt=[]
day_errors=[]
var_importance={}
test_scores=[]
for i in range(39):
ins_msk=store['train_y'].iloc[:,i].fillna(0)>0
oos_msk=store['val_y'].iloc[:,i].fillna(0)>0
if params!=[]:
idv=['%s_visitors_%d'%(x,y) for x,y in product(['min','max','avg','med','cnt'],[1,2,3,7,14,21,35,56,91,147])]
idv+=[x for x in store['train_x'].columns if 'wd%d'%(6-i%7) in x]
idv+=['air_reserve_%d'%i]
train_lgb=lgb.Dataset(store['train_x'].loc[ins_msk,idv],store['train_y'].loc[ins_msk,:].iloc[:,i])
model=lgb.train(params, train_lgb)
ins_pred.append(pd.Series(model.predict(store['train_x'].loc[ins_msk,idv])))
oos_pred.append(pd.Series(model.predict(store['val_x'].loc[oos_msk,idv])))
else:
idv=['med_visitors_56','avg_visitors_wd%d_13'%(6-i%7)]
model=LinearRegression()
ins_vars=store['train_x'].loc[ins_msk,idv]
ins_vars['med_visitors_56']=ins_vars['med_visitors_56'].replace(-99999,2.86)
ins_vars['avg_visitors_wd%d_13'%(6-i%7)]=np.where(ins_vars['avg_visitors_wd%d_13'%(6-i%7)]==-99999,ins_vars['med_visitors_56'],ins_vars['avg_visitors_wd%d_13'%(6-i%7)])
oos_vars=store['val_x'].loc[oos_msk,idv]
oos_vars['med_visitors_56']=oos_vars['med_visitors_56'].replace(-99999,2.86)
oos_vars['avg_visitors_wd%d_13'%(6-i%7)]=np.where(oos_vars['avg_visitors_wd%d_13'%(6-i%7)]==-99999,oos_vars['med_visitors_56'],oos_vars['avg_visitors_wd%d_13'%(6-i%7)])
model.fit(ins_vars,store['train_y'].loc[ins_msk,:].iloc[:,i])
ins_pred.append(pd.Series(model.predict(ins_vars)))
oos_pred.append(pd.Series(model.predict(oos_vars)))
print 'COEF: ',np.array([model.intercept_]),model.coef_
ins_tgt.append(pd.Series(store['train_y'].loc[ins_msk,:].iloc[:,i].values))
oos_tgt.append(pd.Series(store['val_y'].loc[oos_msk,:].iloc[:,i].values))
ins_error=np.sqrt(mean_squared_error(ins_tgt[-1],ins_pred[-1]))
oos_error=np.sqrt(mean_squared_error(oos_tgt[-1],oos_pred[-1]))
print 'RMSLE: ',np.array([ins_error,oos_error])
day_errors.append([ins_error,oos_error])
var_importance[i]=pd.Series(model.feature_importance(),index=idv)
test_score=pd.Series(np.expm1(model.predict(store['test_x'].loc[:,idv])),index=store['test_x'].index.tolist()).to_frame('visitors')
test_score['visit_date']=date(2017,4,23)+timedelta(days=i)
test_scores.append(test_score)
ins_error=np.sqrt(mean_squared_error(pd.concat(ins_tgt),pd.concat(ins_pred)))
oos_error=np.sqrt(mean_squared_error(pd.concat(oos_tgt),pd.concat(oos_pred)))
day_errors.append([ins_error,oos_error])
day_errors=pd.DataFrame(day_errors,columns=['train','test'])
var_importance=pd.DataFrame(var_importance).unstack()
var_importance=var_importance[var_importance>0]
test_scores=pd.concat(test_scores)
return day_errors,var_importance,test_scores
categorical_feature=categorical_feats)
num_round = 10000
regressor = lgb.train(param,
trn_data,
num_round,
valid_sets=[trn_data, val_data],
verbose_eval=100,
early_stopping_rounds=200)
validation_prediction[validation_idx] = regressor.predict(
valid_df_x, num_iteration=regressor.best_iteration)
fold_importance_df = pd.DataFrame()
fold_importance_df["feature"] = features
fold_importance_df["importance"] = regressor.feature_importance()
fold_importance_df["fold_n"] = fold_n + 1
feature_importance_df = pd.concat(
[feature_importance_df, fold_importance_df], axis=0)
predictions += regressor.predict(
test_df[features],
num_iteration=regressor.best_iteration) / folds.n_splits
print("CV score: {:<8.5f}".format(
mean_squared_error(validation_prediction, target)**0.5))
stop = time.time()
print("Time taken by the algorithm:", round(stop - start, 2), "seconds")
#Importance of each feature (Plotting)
cols = (feature_importance_df[[
