xgboost poisson régression: label doit être non négatif

Question

J'utilise un ordinateur portable Windows 10 avec R et xgboost version 0.6-4. Je reçois une erreur bizarre lors de l'exécution du code suivant.

xgb_params <- list("objective" = "count:poisson", 
       "eval_metric" = "rmse") 
regression <- xgboost(data = training_fold, 
        label = y_training_fold, 
        nrounds = 10, 
        params = xgb_params) 

Error in xgb.iter.update(bst$handle, dtrain, iteration - 1, obj) : 
amalgamation/../src/objective/regression_obj.cc:190: Check failed: 
label_correct PoissonRegression: label must be nonnegative 

Mais quand je regarde le résumé de l'étiquette il est écrit:

Min. 1st Qu. Median Mean 3rd Qu. Max. NA's 
0.1129 0.3387 0.7000 1.0987 1.5265 4.5405 287 

Comment puis-je résoudre ce problème? J'ai essayé d'enlever les NA, mais cela n'a pas aidé.

Merci d'avance!

EDIT

Voici un échantillon du traindata

dput(droplevels(head(train[, c(1,2,4,5,6,8,9,10,11)], 20))) 

structure(list(VacancyId = structure(c(1L, 1L, 1L, 1L, 2L, 2L, 
3L, 4L, 4L, 4L, 4L, 5L, 5L, 5L, 6L, 6L, 6L, 6L, 6L, 6L), .Label = c("55288","56838", "57822", "57902", "57925", "58008"), class = "factor"), 
VacancyBankId = c(2L, 1609L, 1611L, 147L, 17L, 1611L, 2L, 
257L, 1611L, 2L, 147L, 17L, 1611L, 239L, 1609L, 2L, 1609L, 
2L, 2L, 1609L), FunctionId = c(36L, 36L, 36L, 36L, 35L, 35L, 
3L, 4L, 4L, 4L, 4L, 9L, 9L, 9L, 3L, 3L, 3L, 3L, 3L, 3L), 
EducationLevel = c(6L, 6L, 6L, 6L, 6L, 6L, 4L, 6L, 6L, 6L, 
6L, 4L, 4L, 4L, 6L, 6L, 6L, 6L, 6L, 6L), ProvinceId = c(22L, 
22L, 22L, 22L, 24L, 24L, 19L, 16L, 16L, 16L, 16L, 19L, 19L, 
19L, 21L, 21L, 16L, 16L, 22L, 22L), CandidatesCount = c(126L, 
27L, 18L, 12L, 1L, 4L, 2L, 6L, 7L, 7L, 1L, 8L, 15L, 13L, 
7L, 7L, 7L, 7L, 7L, 7L), DurationDays = c(62L, 62L, 62L, 
62L, 18L, 18L, 43L, 61L, 61L, 61L, 61L, 60L, 60L, 60L, 62L, 
62L, 62L, 62L, 62L, 62L), DurationWeeks = c(8.857142857, 
8.857142857, 8.857142857, 8.857142857, 2.571428571, 2.571428571, 
6.142857143, 8.714285714, 8.714285714, 8.714285714, 8.714285714, 
8.571428571, 8.571428571, 8.571428571, 8.857142857, 8.857142857, 
8.857142857, 8.857142857, 8.857142857, 8.857142857), CandidatesPerWeek = c(NA, 
3.048387097, 2.032258065, 1.35483871, 0.388888889, 1.555555556, 
0.325581395, 0.68852459, 0.803278689, 0.803278689, 0.114754098, 
0.933333333, 1.75, 1.516666667, 0.790322581, 0.790322581, 
0.790322581, 0.790322581, 0.790322581, 0.790322581)), .Names = c("VacancyId", "VacancyBankId", "FunctionId", "EducationLevel", "ProvinceId", "CandidatesCount", "DurationDays", "DurationWeeks", "CandidatesPerWeek"), row.names = c(1L, 2L, 3L, 4L, 5L, 6L, 11L, 12L, 13L, 14L, 15L, 16L, 17L, 18L, 26L, 27L, 28L, 29L, 30L, 31L), class = "data.frame") 

Je veux prédire les candidats par semaine avec functionId, Educationlevel, Province et VacancyBankId. Donc y_training_fold sont des candidats par semaine et training_fold sont fonction, éducation, province et vacancybankid.

J'espère que quelqu'un peut m'aider!

Answer 1

Le problème dans votre ensemble de données n'est pas la présence de valeurs négatives dans y_training_fold mais la présence de valeurs non entières.
Voir la simulation suivante avec un vecteur y_training_fold des valeurs non entières:

library(xgboost) 

training_fold <- matrix(rnorm(1000),nrow=100) 
y_training_fold <- matrix(rnorm(100),ncol=1) 

xgb_params <- list("objective" = "count:poisson", 
       "eval_metric" = "rmse") 
regression <- xgboost(data = training_fold, 
        label = y_training_fold, 
        nrounds = 10, 
        params = xgb_params) 

Le message d'erreur est exactement la même que vous avez déclaré:

Error in xgb.iter.update(bst$handle, dtrain, iteration - 1, obj) : 
    [11:46:28] amalgamation/../src/objective/regression_obj.cc:190: 
    Check failed: label_correct PoissonRegression: label must be nonnegative 

Maintenant, essayez avec un vecteur y_training_fold d'entier:

y_training_fold <- matrix(rpois(100,10),ncol=1) 

xgb_params <- list("objective" = "count:poisson", 
       "eval_metric" = "rmse") 
regression <- xgboost(data = training_fold, 
        label = y_training_fold, 
        nrounds = 10, 
        params = xgb_params) 

maintenant xgboost fonctionne bien:

[1]  train-rmse:9.795855 
[2]  train-rmse:9.660112 
[3]  train-rmse:9.492991 
[4]  train-rmse:9.287366 
[5]  train-rmse:9.034582 
[6]  train-rmse:8.724205 
[7]  train-rmse:8.343800 
[8]  train-rmse:7.878869 
[9]  train-rmse:7.312294 
[10] train-rmse:6.632671 

EDIT.

Utilisation de vos données la solution au problème est:

dts <- structure(list(VacancyId = structure(c(1L, 1L, 1L, 1L, 2L, 2L, 
3L, 4L, 4L, 4L, 4L, 5L, 5L, 5L, 6L, 6L, 6L, 6L, 6L, 6L), .Label = c("55288","56838", "57822", "57902", "57925", "58008"), class = "factor"), 
VacancyBankId = c(2L, 1609L, 1611L, 147L, 17L, 1611L, 2L, 
257L, 1611L, 2L, 147L, 17L, 1611L, 239L, 1609L, 2L, 1609L, 
2L, 2L, 1609L), FunctionId = c(36L, 36L, 36L, 36L, 35L, 35L, 
3L, 4L, 4L, 4L, 4L, 9L, 9L, 9L, 3L, 3L, 3L, 3L, 3L, 3L), 
EducationLevel = c(6L, 6L, 6L, 6L, 6L, 6L, 4L, 6L, 6L, 6L, 
6L, 4L, 4L, 4L, 6L, 6L, 6L, 6L, 6L, 6L), ProvinceId = c(22L, 
22L, 22L, 22L, 24L, 24L, 19L, 16L, 16L, 16L, 16L, 19L, 19L, 
19L, 21L, 21L, 16L, 16L, 22L, 22L), CandidatesCount = c(126L, 
27L, 18L, 12L, 1L, 4L, 2L, 6L, 7L, 7L, 1L, 8L, 15L, 13L, 
7L, 7L, 7L, 7L, 7L, 7L), DurationDays = c(62L, 62L, 62L, 
62L, 18L, 18L, 43L, 61L, 61L, 61L, 61L, 60L, 60L, 60L, 62L, 
62L, 62L, 62L, 62L, 62L), DurationWeeks = c(8.857142857, 
8.857142857, 8.857142857, 8.857142857, 2.571428571, 2.571428571, 
6.142857143, 8.714285714, 8.714285714, 8.714285714, 8.714285714, 
8.571428571, 8.571428571, 8.571428571, 8.857142857, 8.857142857, 
8.857142857, 8.857142857, 8.857142857, 8.857142857), CandidatesPerWeek = c(NA, 
3.048387097, 2.032258065, 1.35483871, 0.388888889, 1.555555556, 
0.325581395, 0.68852459, 0.803278689, 0.803278689, 0.114754098, 
0.933333333, 1.75, 1.516666667, 0.790322581, 0.790322581, 
0.790322581, 0.790322581, 0.790322581, 0.790322581)), 
.Names = c("VacancyId", "VacancyBankId", "FunctionId", "EducationLevel", 
"ProvinceId", "CandidatesCount", "DurationDays", "DurationWeeks", 
"CandidatesPerWeek"), row.names = c(1L, 2L, 3L, 4L, 5L, 6L, 11L, 12L, 
13L, 14L, 15L, 16L, 17L, 18L, 26L, 27L, 28L, 29L, 30L, 31L), 
class = "data.frame") 

# Delete missing values 
dts <- na.omit(dts) 

# Build X matrix of potential predictors 
# Important: do not use the first column (ID) and the last (response variable) 
training_fold <- as.matrix(dts[,-c(1,9)]) 
# Round to the nearest integer the response variable 
y_training_fold <- as.matrix(dts[,9]) 
y_training_fold <- round(y_training_fold) 

xgb_params <- list("objective" = "count:poisson", 
       "eval_metric" = "rmse") 
(regression <- xgboost(data = training_fold, 
        label = y_training_fold, 
        nrounds = 10, 
        params = xgb_params)) 
# Output 
##### xgb.Booster 
# raw: 4.6 Kb 
# call: 
# xgb.train(params = params, data = dtrain, nrounds = nrounds, 
#  watchlist = watchlist, verbose = verbose, print_every_n = print_every_n, 
#  early_stopping_rounds = early_stopping_rounds, maximize = maximize, 
#  save_period = save_period, save_name = save_name, xgb_model = xgb_model, 
#  callbacks = callbacks) 
# params (as set within xgb.train): 
# objective = "count:poisson", eval_metric = "rmse", silent = "1" 
# xgb.attributes: 
# niter 
# callbacks: 
# cb.print.evaluation(period = print_every_n) 
# cb.evaluation.log() 
# cb.save.model(save_period = save_period, save_name = save_name) 
# niter: 10 
# evaluation_log: 
#  iter train_rmse 
#  1 0.914084 
#  2 0.829741 
# ---     
#  9 0.332951 
#  10 0.291877