CSOAA multiclass classification

Cost Sensitive One Against All (CSOAA) - A multi-class predictive modeling reduction in Vowpal Wabbit

We will use abalone dataset and will try to predict age groups (based on number of abalone shell rings) from physical measurements

library(mltools)
library(rvw)

aburl = 'http://archive.ics.uci.edu/ml/machine-learning-databases/abalone/abalone.data'
abnames = c('sex','length','diameter','height','weight.w','weight.s','weight.v','weight.sh','rings')
abalone = read.table(aburl, header = F , sep = ',', col.names = abnames)
data_full <- abalone

Split number of rings into groups with equal (as possible) number of observations

data_full$group <- bin_data(data_full$rings, bins=3, binType = "quantile")
group_lvls <- levels(data_full$group)
levels(data_full$group) <- c(1, 2, 3)

Prepare variables for CSOAA algorithm

data_full$cost_class_1 <- ifelse(data_full$group == 1, 0.8, 0.1)
data_full$cost_class_2 <- ifelse(data_full$group == 2, 0.8, 0.1)
data_full$cost_class_3 <- ifelse(data_full$group == 3, 0.8, 0.1)
data_full$rings <- factor(data_full$rings)
data_full$tag <- sapply(1:nrow(data_full), function(x) paste0("ex",x))

Prepare indices to split data

ind_train <- sample(1:nrow(data_full), 0.8*nrow(data_full))

Split data into train and test subsets

df_train <- data_full[ind_train,]
df_test <- data_full[-ind_train,]

Set up CSOAA VW model

vwmodel <- vwsetup(dir = "./",
                   option = "csoaa", num_classes = 3)

Train and test

vwtrain(vwmodel, data = df_train,
        namespaces = list(NS1 = list("sex", "rings"), NS2 = list("diameter", "length", "height")),
        targets = c("cost_class_1", "cost_class_2", "cost_class_3"), tag = "tag"
)

#> Converting data.frame to VW format
#> Starting VW training session
#> VW v8.6.1
#> Using data file: ./train.vw
#> Using model file: ./vw_1552857569_mdl.vw
#> Command line parameters: 
#> --csoaa 3 -d ./train.vw -f ./vw_1552857569_mdl.vw --passes 1
#> final_regressor = ./vw_1552857569_mdl.vw
#> Num weight bits = 18
#> learning rate = 0.5
#> initial_t = 0
#> power_t = 0.5
#> using no cache
#> Reading datafile = ./train.vw
#> num sources = 1
#> average  since         example        example  current  current  current
#> loss     last          counter         weight    label  predict features
#> 0.700000 0.700000            1            1.0    known        1        6
#> 0.700000 0.700000            2            2.0    known        2        6
#> 0.350000 0.000000            4            4.0    known        3        6
#> 0.262500 0.175000            8            8.0    known        1        6
#> 0.131250 0.000000           16           16.0    known        3        6
#> 0.087500 0.043750           32           32.0    known        1        6
#> 0.054687 0.021875           64           64.0    known        3        6
#> 0.027344 0.000000          128          128.0    known        1        6
#> 0.013672 0.000000          256          256.0    known        2        6
#> 0.006836 0.000000          512          512.0    known        1        6
#> 0.003418 0.000000         1024         1024.0    known        3        6
#> 0.001709 0.000000         2048         2048.0    known        1        6
#> 
#> finished run
#> number of examples = 3341
#> weighted example sum = 3341.000000
#> weighted label sum = 0.000000
#> average loss = 0.001048
#> total feature number = 20044

vwpreds <- predict(vwmodel, data = df_test, full_probs = T, quiet = T)

head(vwpreds)

#>   V1   V2
#> 1  3  ex2
#> 2  2  ex8
#> 3  2 ex21
#> 4  2 ex23
#> 5  2 ex26
#> 6  1 ex35