Autoencoder transformation (encode-decode)¶

Considering a dataset with $p$ numerical attributes.

The goal of the autoencoder is to reduce the dimension of $p$ to $k$, such that these $k$ attributes are enough to recompose the original $p$ attributes. However from the $k$ dimensionals the data is returned back to $p$ dimensions. The higher the quality of autoencoder the similiar is the output from the input.

In [1]:

# DAL ToolBox
# version 1.0.777

source("https://raw.githubusercontent.com/cefet-rj-dal/daltoolbox/main/jupyter.R")

#loading DAL
load_library("daltoolbox")

Loading required package: daltoolbox

Registered S3 method overwritten by 'quantmod':
  method            from
  as.zoo.data.frame zoo 


Attaching package: ‘daltoolbox’


The following object is masked from ‘package:base’:

    transform

dataset for example¶

In [2]:

data(sin_data)

sw_size <- 5
ts <- ts_data(sin_data$y, sw_size)

ts_head(ts)

A matrix: 6 × 5 of type dbl
t4	t3	t2	t1	t0
0.0000000	0.2474040	0.4794255	0.6816388	0.8414710
0.2474040	0.4794255	0.6816388	0.8414710	0.9489846
0.4794255	0.6816388	0.8414710	0.9489846	0.9974950
0.6816388	0.8414710	0.9489846	0.9974950	0.9839859
0.8414710	0.9489846	0.9974950	0.9839859	0.9092974
0.9489846	0.9974950	0.9839859	0.9092974	0.7780732

applying data normalization¶

In [3]:

preproc <- ts_norm_gminmax()
preproc <- fit(preproc, ts)
ts <- transform(preproc, ts)

ts_head(ts)

A matrix: 6 × 5 of type dbl
t4	t3	t2	t1	t0
0.5004502	0.6243512	0.7405486	0.8418178	0.9218625
0.6243512	0.7405486	0.8418178	0.9218625	0.9757058
0.7405486	0.8418178	0.9218625	0.9757058	1.0000000
0.8418178	0.9218625	0.9757058	1.0000000	0.9932346
0.9218625	0.9757058	1.0000000	0.9932346	0.9558303
0.9757058	1.0000000	0.9932346	0.9558303	0.8901126

spliting into training and test¶

In [4]:

samp <- ts_sample(ts, test_size = 10)
train <- as.data.frame(samp$train)
test <- as.data.frame(samp$test)

creating autoencoder¶

Reduce from 5 to 3 dimensions

In [5]:

auto <- autoenc_encode_decode(5, 3)

auto <- fit(auto, train)

testing autoencoder¶

presenting the original test set and display encoding

In [6]:

print(head(test))
result <- transform(auto, test)
print(head(result))

         t4        t3        t2        t1        t0
1 0.7258342 0.8294719 0.9126527 0.9702046 0.9985496
2 0.8294719 0.9126527 0.9702046 0.9985496 0.9959251
3 0.9126527 0.9702046 0.9985496 0.9959251 0.9624944
4 0.9702046 0.9985496 0.9959251 0.9624944 0.9003360
5 0.9985496 0.9959251 0.9624944 0.9003360 0.8133146
6 0.9959251 0.9624944 0.9003360 0.8133146 0.7068409
          [,1]      [,2]      [,3]      [,4]      [,5]
[1,] 0.7281464 0.8297843 0.9119417 0.9712389 0.9987636
[2,] 0.8298926 0.9120755 0.9692511 0.9992059 0.9964882
[3,] 0.9135329 0.9704542 0.9987994 0.9969148 0.9618803
[4,] 0.9685430 0.9973767 0.9959966 0.9621131 0.9001518
[5,] 0.9985738 0.9968489 0.9646484 0.8984091 0.8126789
[6,] 0.9928818 0.9628512 0.9030885 0.8135285 0.7084026

In [ ]: