Computer Science/Data Science

[์ถ”์ฒœ์‹œ์Šคํ…œ] Collaborative Denoising AutoEncoders for Top-N Recommender Systems ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ

_cactus 2024. 9. 24. 11:52
๋ฐ˜์‘ํ˜•

Collaborative Denoising Auto-Encoders for Top-N Recommender Systems

1. Introducton

CDAE (Collaborative Denoising AutoEncoder) ๋Š” DAE๋ฅผ Collaborative Filtering์— ์ ์šฉํ•˜์—ฌ top-N ์ถ”์ฒœ์— ํ™œ์šฉํ•œ ๋ชจ๋ธ
๋ชจ๋ธ์€ input์œผ๋กœ corrupted๋œ user-item ์„ ํ˜ธ๋„๋ฅผ ์ฃผ๊ณ  ์ด๊ฒƒ์˜ latent representation์„ ํ•™์Šต
→ ์ด๋Š” corrupted๋˜๊ธฐ ์ „์˜ ์›๋ž˜์˜ input์„ ๋” ์ž˜ ๋ณต์›ํ•ด์คŒ


2. Problem Definition

2.1 Overview of Model-based Recommender

  1. model definition (input๊ณผ output์˜ ๊ด€๊ณ„๋ฅผ ์ˆ˜์‹ํ™”ํ•œ ๊ฒƒ)
  2. objective function (best model์˜ param์„ ์ฐพ๊ธฐ ์œ„ํ•ด ํ•™์Šต๊ณผ์ •์—์„œ ์ตœ์ ํ™”ํ•˜๋Š” ๊ฒƒ)

 

Recommender Models

user u์˜ item i์— ๋Œ€ํ•œ ์„ ํ˜ธ๋„๋ฅผ ๊ตฌํ•˜๋Š” ๋ชจ๋ธ

  • ํ›„๋ณด ๋ชจ๋ธ 4๊ฐ€์ง€
    1. Latent Factor Model (LFM)
      • user์™€ item์˜ latent vector๋“ค์˜ dot product๋กœ ์„ ํ˜ธ๋„๋ฅผ ์ •์˜
    2. Similarity Model (SM)
      • item j์— ๋Œ€ํ•œ user์˜ ์„ ํ˜ธ๋„์™€ item i์™€ j๊ฐ„์˜ ์œ ์‚ฌ๋„์˜ ๊ฐ€์ค‘ํ•ฉ์œผ๋กœ ์ •์˜
    3. Factorized Similarity Model (FSM)
      • SM๋ชจ๋ธ์˜ similarity matrix๋ฅผ 2๊ฐœ์˜ ์ €์ฐจ์› ํ–‰๋ ฌ๋กœ ๋ถ„ํ•ด
        → item๊ฐœ์ˆ˜์— ๋”ฐ๋ผ parameter์˜ ๊ฐœ์ˆ˜๊ฐ€ ๊ธฐํ•˜๊ธ‰์ˆ˜์ ์œผ๋กœ ์ฆ๊ฐ€ํ•˜๋Š” SM๋ชจ๋ธ์˜ ๋‹จ์  ๊ฐœ์„ 
    4. LFSM (LFM+FSM)
      • netflix prize์—์„œ ์šฐ์Šนํ•œ SVD++
๋ฐ˜์‘ํ˜•

Objective Functions for Recommenders

์ถ”์ฒœ๋ชจ๋ธ ํ•™์Šต์‹œ ์‚ฌ์šฉ๋˜๋Š” ๋ชฉ์ ํ•จ์ˆ˜๋Š” 2๊ฐ€์ง€
*์ฃผ์˜ํ•  ์  :
ํ•™์Šต ๋ฐ์ดํ„ฐ ์…‹์œผ๋กœ ๊ด€์ฐฐ๋œ user-item pair๋งŒ ์‚ฌ์šฉ ์‹œ, ๋ฐ์ดํ„ฐ๊ฐ€ implicit feedback์ธ ๊ฒฝ์šฐ y๊ฐ’์ด ์ „๋ถ€ 1์ด๊ธฐ ๋•Œ๋ฌธ์— ์˜ˆ์ธก๊ฐ’๋„ ์ „๋ถ€ 1์ธ ๋ฌธ์ œ ๋ฐœ์ƒ
→ ๋”ฐ๋ผ์„œ ๊ด€์ธก๋˜์ง€ ์•Š์€ ๋ฐ์ดํ„ฐ๋„ ํ•™์Šต๋ฐ์ดํ„ฐ์— ํฌํ•จ์‹œ์ผœ์•ผ ํ•จ ($O'$)

  1. point-wise
    • item ๊ฐ๊ฐ์„ ๋…๋ฆฝ์ ์ธ ๊ฐœ์ฒด๋กœ ๋ณด๊ณ  ๊ฐ๊ฐ์— ๋Œ€ํ•˜์—ฌ ์„ ํ˜ธ๋„ ๊ณ„์‚ฐํ•˜์—ฌ ์˜ˆ์ธก๊ฐ’์„ ๋ฑ‰์Œ
  2. pair-wise
    • item pair (positive item, negative item)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์„ ํ˜ธ๋„ ๊ณ„์‚ฐ
    • item๊ฐ„์— ์ƒ๋Œ€์  rank๊ฐ€ ๋ฐ˜์˜๋จ

⇒ ์ผ๋ฐ˜์ ์œผ๋กœ top-N recommender๋ฅผ ์ตœ์ ํ™”ํ•˜๋Š”๋ฐ๋Š” pair-wise ํ•จ์ˆ˜๊ฐ€ ๋” ์ ํ•ฉ
 
 
loss function

  • square loss
  • log loss
  • hinge loss
  • cross entropy loss

 

2.2 Denoising Auto-Encoders

denoising auto-encoder(DAE)๋Š” classical auto-encoder์˜ ํ™•์žฅํŒ

neural network structure

  • 1๊ฐœ์˜ hidden-layer
  • input vector $x$
  • hidden representation $z$
    • mapping function $h$$W$ : weight matrix, $b$ : offset vector
    • $z=h(x)=\sigma(W^Tx+b)$
    • reconstructed vector $\hat{x}$
    • $\hat{x} = \sigma(W'z+b'),\ \ W' =W$ (tied weights)

 

corrupted input vector

DAE๋Š” input ํ•™์Šต ๋ฐ์ดํ„ฐ $x$์— ๋ณ€ํ˜•์„ ์ค€ nosiy input $\tilde{x}$ ๋ฅผ ์‚ฌ์šฉ
→ robustํ•œ ๋ชจ๋ธ
corruption method

  1. additive Gaussian noise
  2. multiplicative mask-out/drop-out noise
  3. : input์˜ ์ผ๋ถ€๋ฅผ q์˜ ํ™•๋ฅ ๋กœ 0์œผ๋กœ ๋žœ๋คํ•˜๊ฒŒ ๋ฎ์–ด์“ฐ๋Š” ๋ฐฉ์‹

3. Proposed Methodology

3.1 Collaborative Denoising Auto-Encoder

CDAE๋Š” DAE๋ž‘ ๋น„์Šทํ•˜๊ฒŒ 1๊ฐœ์˜ hidden layer๋กœ ๊ตฌ์„ฑ๋œ nerural net์ธ๋ฐ์š”, ์ฐจ๋ณ„์ ์€
DAE์™€์˜ ์ฐจ๋ณ„์ 

  • user์— ๋Œ€ํ•œ latent vector encoding

 

๊ตฌ์กฐ

Input Layer

  • $I+1$๊ฐœ์˜ ๋…ธ๋“œ
    • $I$๊ฐœ์˜ item input node
    • $1$๊ฐœ์˜ user input node
      • user-specific node
  • item input node
    • $y_{ui}$ : user $u$ ์˜ item $i$ ์— ๋Œ€ํ•œ ์„ ํ˜ธ๋„ (0/1)

Hidden Layer

  • $K$๊ฐœ์˜ ๋…ธ๋“œ
  • input vector ํฌ๊ธฐ๋ณด๋‹ค ์ž‘์Œ
  • bias node

Output Layer

  • $I$๊ฐœ์˜ ๋…ธ๋“œ
  • reconstructed ๋œ $y_{u}$

 

Weight

  • $W$ : item input node์™€ hidden layer node ์‚ฌ์ด์˜ weight matrix
  • $V_{u}$ : user input node์˜ weight vector
  • $W'$ : hidden layer์™€ output layer ์‚ฌ์ด์˜ weight matrix
  • $b$ : hidden layer์˜ bias node์˜ weight vector

 

 

input vector

  • $y_{u} = {{y_{u1}, y_{u2}, ... , y_{uI}}}$
    • sparse binary vector
    • $|O_{u}|$๊ฐœ์˜ $y_{ui}$ = 1, ๋‚˜๋จธ์ง€๋Š” 0
  • : user $u$์˜ item ์ง‘ํ•ฉ $I$์•ˆ์˜ ๋ชจ๋“  item์— ๋Œ€ํ•œ feedback์ด ๋‹ด๊ธด vector

corrupted input vector

  • $\tilde{y_{u}} = {{\tilde{y_{u1}}, \tilde{y_{u2}}, ... , \tilde{y_{uI}}}}$
  • q์˜ ํ™•๋ฅ ๋กœ 0์ด ์•„๋‹Œ $y$๋ฅผ drop out

 

 

CDAE parameter ํ•™์Šต๊ณผ์ •

  • parameterํ•™์Šต์— SGD๊ธฐ๋ฒ• ์ ์šฉ
    • ์ „์ฒด user์— ๋Œ€ํ•ด iteration 1๋ฒˆ ์‹œ๊ฐ„๋ณต์žก๋„ = $O(UIK)$
      → user์™€ item์ˆ˜๊ฐ€ ๊ต‰์žฅํžˆ ์ปค์ง€๋ฉด ์‹คํ–‰์–ด๋ ค์›€
    • ⇒ $( S_{u} \subset \bar{O_{u}} ) \ \cup O_{u}$ ์— ํฌํ•จ๋œ item์— ๋Œ€ํ•ด์„œ๋งŒ gradient ๊ณ„์‚ฐํ•˜๋„๋ก.
  • AdaGrad ์ ์šฉ

Recommendation

  • user $u$ $$์˜ $\bar{O_{u}}$ ์˜ item ์ค‘ output layer์—์„œ ๊ฐ€์žฅ ํฐ ์˜ˆ์ธก๊ฐ’๋“ค์„ ๊ฐ€์ง„ item์„ ์ถ”์ฒœ

 

3.2 Discussion

  • CDAE๋Š” top-N recommendation์„ ๊ตฌํ˜„ํ•˜๋Š”๋ฐ ์žˆ์–ด ์œ ์—ฐํ•˜๊ณ  ์ผ๋ฐ˜ํ™”๋œ ๋ชจ๋ธ
    (2.1์—์„œ ์ œ์‹œ๋œ ๋‹ค๋ฅธ ๊ธฐ์กด method๋“ค์„ ์ •์˜ํ•  ์ˆ˜ ์žˆ์Œ)
  • → best model์„ ํ™•์ •์ง€์„ ์ˆ˜ ์—†์ง€๋งŒ ์—ฌ๋Ÿฌ variant๋ฅผ ์‹คํ—˜ํ•ด๋ณด๊ณ  task์— ๊ฐ€์žฅ ์ ํ•ฉํ•œ ๋ชจ๋ธ์„ ๊ณ ๋ฅผ ์ˆ˜ ์žˆ์Œ
  • CDAE์—์„œ๋Š” pair-wise๋ณด๋‹ค point-wise๋ฅผ ์“ฐ๋Š” ๊ฒƒ์ด ์„ฑ๋Šฅ์ด ๋” ์ข‹์Œ

4. Related Work

neural network๊ธฐ๋ฐ˜ ์ถ”์ฒœ ์‹œ์Šคํ…œ ๋ชจ๋ธ

Restricted Boltzmann Machines (RBM)

  • top-N์ด ์•„๋‹Œ ์ˆœ์œ„์˜ˆ์ธก ๋ชจ๋ธ
  • loss function์—์„œ ๊ด€์ฐฐ๋œ ํ‰์ ๋งŒ ๊ณ ๋ ค

AutoRec

  • rating ์˜ˆ์ธก์— Auto-Encoder ์‚ฌ์šฉ
  • ์ฐจ์ด์ 
    1. AutoRec์—ญ์‹œ loss function์—์„œ observed ratings๋งŒ ๊ณ ๋ ค ⇒ ์ด๋Š” top-N ์ถ”์ฒœ์˜ ์„ฑ๋Šฅ์„ ๋ณด์žฅํ•ด์ฃผ์ง€ ๋ชปํ•จ
    2. vanilla Auto-Encoder ๊ตฌ์กฐ๋ฅผ ์‚ฌ์šฉ
    3. denoising ๊ธฐ์ˆ  ์‚ฌ์šฉ X

5. Experimental Results

์‹คํ—˜ ํ‰๊ฐ€๋Š” ๋‘ ํŒŒํŠธ๋กœ ์ด๋ฃจ์–ด์ง

  1. study the effects of various choice of the components of CDAE (CDAE์˜ ์—ฌ๋Ÿฌ ์š”์†Œ๋“ค์— ๋ณ€ํ™”๋ฅผ ์ฃผ์–ด ๊ทธ ํšจ๊ณผ๋ฅผ ๊ฒ€์ฆ)
  2. CDAE์™€ ์ตœ์‹  top-N ์ถ”์ฒœ ๋ชจ๋ธ ๋น„๊ต

5.1 Data Sets and Experimental Setup

3๊ฐ€์ง€ data set

  1. MovieLens 10M (ML)
  2. Netflix
  3. Yelp
  • 4 star ์ด์ƒ์˜ ๋ฐ์ดํ„ฐ๋งŒ ๋‚จ๊ธฐ๊ณ  ๋‚˜๋จธ์ง€๋Š” ์ „๋ถ€ missing data๋กœ ๊ฐ„์ฃผ
  • ๋‚จ๊ฒจ์ง„ ๋ฐ์ดํ„ฐ๋Š” $y_{ui}$ = 1 ๋กœ ๋ณ€ํ™˜
  • (์ด๋Ÿฌํ•œ ์ „์ฒ˜๋ฆฌ๋Š” implicit feedback์„ ๊ฐ€์ง„ ์ถ”์ฒœ์˜ ๊ฒฝ์šฐ ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ๊ณผ์ •์ด๋‹ค)
  • ๋‚จ๊ธด ํ‰์  ์ˆ˜๊ฐ€ 5๊ฐœ ์ดํ•˜์ธ user/item ์ œ๊ฑฐ

 
 

5.2 Implementation Details

  • ๋ชจ๋ธ๋ณ„๋กœ 5-fold CV ์ง„ํ–‰ํ•˜์—ฌ best hyperparameter ์„ ์ • → ํ•™์Šต
  • SGD๋ฅผ ์ ์šฉํ•˜์—ฌ parameter ํ•™์Šต (์ œ์•ˆ ๋ชจ๋ธ๊ณผ ๋น„๊ต๋ฅผ ์œ„ํ•ด)
  • AdaGrad ์ ์šฉ
  • negative sampling์˜ ๊ฒฝ์šฐ, ์—ฌ๋Ÿฌ ์ˆซ์ž๋ฅผ ์‹คํ—˜ํ•ด๋ณธ ๊ฒฐ๊ณผ, NS = 5 ์ผ ๋•Œ ์ผ๊ด€์„ฑ์žˆ๋Š” ์ข‹์€ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์ž„
    → ์ด๋Š” ๊ฐuser๋Š” negative sample์ด observed rating ์ˆ˜์˜ 5๋ฐฐ์ž„์„ ์˜๋ฏธ ..(?)

5.3 Evaluation Metrics

  • precision
  • recall

์ถ”์ฒœ๋ชจ๋ธ์˜ precision๊ณผ recall์€ ๋ชจ๋“  ์œ ์ €์˜ precision๊ณผ recall์˜ ํ‰๊ท 

  • MAP (Mean Average Precision)
    • AP (Average Precision)
    • ranked precision metric์œผ๋กœ, ์˜ณ๊ฒŒ(?) ์ถ”์ฒœํ•œ item์ด top rank์— ์žˆ์„์ˆ˜๋ก ๋” ํฐ ๊ฐ€์ค‘์น˜ ๋ถ€์—ฌํ•˜๋Š” ๋ฐฉ์‹
    • MAP
    • ๋ชจ๋“  user์˜ AP score์˜ ํ‰๊ท 
  • N์ด 1,5,10 ์ธ ๊ฒฝ์šฐ์— ๋Œ€ํ•ด์„œ๋งŒ ์‹คํ—˜. (N์€ ์ถ”์ฒœ ์ˆ˜)

 

 

5.4 Analysis of CDAE Components

CDAE๋ชจ๋ธ์˜ ์ฃผ์š” ์š”์†Œ

  1. mapping function
    • hidden layer์™€ output layer์— identitify/sigmoid function ์กฐํ•ฉ์‹ค
  2. loss function
    • CDAE์˜ ๊ฒฝ์šฐ pair-wise๋ณด๋‹ค point-wise objective function์˜ ์„ฑ๋Šฅ์ด ๋” ์ข‹์Œ
      ์›์ธ : implicit feedback
  3. level of corruption
    • corrupton level q [ 0, 0.2, 0.4, 0.6, 0.8, 1]

 

์‹คํ—˜ ๊ฒฐ๊ณผ

๊ฒฐ๋ก  : ๋ฐ์ดํ„ฐ์— ๋”ฐ๋ผ ์š”์†Œ๋“ค์„ ๊ฒฐ์ •ํ•ด์•ผ ํ•จ
mapping function

  • non-linear function์„ ์‚ฌ์šฉ → ์„ฑ๋Šฅ ํ–ฅ์ƒ

denoising technique

  • input์— noise๋ฅผ ์ถ”๊ฐ€ → ์„ฑ๋Šฅ ํ–ฅ์ƒ
    M1, M2์—์„œ ๋šœ๋ ท

 

DAE์™€ ๋น„๊ต

  • CDAE๋Š” user specific input์„ ๊ฐ€์ง

์‹คํ—˜

  • Yelp data
    M!, M2์—์„œ user-specific vector๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ํฐ ์„ฑ๋Šฅํ–ฅ์ƒ์„ ๋ณด์ž„
    M3, M4์—์„œ๋Š” ์„ฑ๋Šฅ ํ–ฅ์ƒ์˜ ์ •๋„๊ฐ€ ์ƒ๋Œ€์ ์œผ๋กœ ์ž‘์Œ
  • MovieLens data
    ๋ชจ๋“  variant์— ๋Œ€ํ•˜์—ฌ DAE๋ณด๋‹ค CDAE์—์„œ MAP ๊ฐ’์ด ๋” ํผ

 
 

Tied Weights

(์—ฌ๊ธฐ์„œ tied weights์˜ ํšจ๊ณผ์„ฑ ํ™•์ธ)

  • tied weights๋ฅผ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ์™€ ์•ˆ ํ–ˆ์„ ๋•Œ์˜ MAP ๋น„๊ต

์‹คํ—˜ ๊ฒฐ๊ณผ

  • MovieLens์˜ M2๋ชจ๋ธ์„ ์ œ์™ธํ•˜๊ณ  ์ „๋ถ€ tied weights๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์•˜์„ ๋•Œ(NTW) ๋” ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์ž„
    → CDAE์—์„œ tied weights๋ฅผ ์‚ฌ์šฉ ์ถ”์ฒœ X

 

Number of latent dimension

  • latent dimension $K$ ์˜ ํšจ๊ณผ์„ฑ ํ™•์ธ

์‹คํ—˜ ๊ฒฐ๊ณผ

  • $K$๊ฐ’์ด ์ปค์งˆ์ˆ˜๋ก (์ผ์ • ์ˆ˜์ค€๊นŒ์ง€๋งŒ) ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ (K๊ฐ’์„ ๋” ์ฆ๊ฐ€์‹œํ‚ค๋ฉด ๋” ์ด์ƒ ์„ฑ๋Šฅ์ด ์ข‹์•„์ง€์ง€ ์•Š๊ณ  overfitting์— ์˜ํ•ด ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง)

 
 

5.5 Experimental Comparisons woth Previous Models

  • CDAE์™€ ๋‹ค๋ฅธ ์œ ๋ช…ํ•œ top-N recommendation ๋ฐฉ์‹ baseline์œผ๋กœ ์‚ผ๊ณ  ๋น„๊ต
  • baseline ๋ชจ๋ธ๋“ค์˜ hyperparameter๋Š” cross validation์„ ์ ์šฉํ•˜์—ฌ ์„ ์ •
  • latent dimension = 50
  • ํ‰๊ฐ€์ง€ํ‘œ : MAP์™€ recall

 
baseline method

  • POP
  • ITEMCF
  • MK
  • BPR
  • FISM

์‹คํ—˜๊ฒฐ๊ณผ

  • MAP์™€ recall์˜ ๊ฒฐ๊ณผ ์ผ์น˜ (๊ทธ๋ž˜ํ”„์˜ ์„ฑํ–ฅ์ด ๋น„์Šทํ•œ๊ฑธ ์•Œ์ˆ˜์žˆ์Œ)
  • MAP@10๊ณผ Recall@10์˜ ๊ฒฐ๊ณผ์—์„œ CDAE๊ฐ€ ์ผ๊ด€์„ฑ์žˆ๊ฒŒ ๋‹ค๋ฅธ ๋ชจ๋ธ๋“ค์„ ๋Šฅ๊ฐ€ํ•จ
  • Yelp data์—์„œ๋Š” CDAE๊ฐ€ ๋‹ค๋ฅธ ๋ชจ๋ธ๋“ค๊ณผ ํฐ ์ฐจ์ด๋ฅผ ์ด๋ฃจ๋ฉฐ ๋Šฅ๊ฐ€, ๊ทธ ์ค‘ ๋‘๋ฒˆ์งธ๋กœ ์„ฑ๋Šฅ์ด ์ข‹์€ ๋ชจ๋ธ์ธ MF์˜ ์„ฑ๋Šฅ๋ณด๋‹ค 15% ์ข‹์Œ
  • Netflix data์—์„œ ํฅ๋ฏธ๋กœ์šด ์ ์€ ITEMCF๋ชจ๋ธ์ด (CDAE์ œ์™ธ)๋‹ค๋ฅธ ๋ชจ๋ธ๋“ค์— ๋น„ํ•ด ๋” ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์ž„. ๊ทธ๋Ÿผ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ์—ฌ๊ธฐ์„œ๋„ ์—ฌ์ „ํžˆ CDAE๊ฐ€ ๋” ์ข‹์Œ

6. Conclusion

  • top-n ์ถ”์ฒœ๋ชจ๋ธ๋กœ CDAE๋ฅผ ์ œ์‹œ
  • CDAE๋Š” DAE๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ user-item feedback ๋ฐ์ดํ„ฐ ํ˜•์„ฑ
  • 3๊ฐ€์ง€ ๋ฐ์ดํ„ฐ์…‹์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ชจ๋ธ์˜ ์š”์†Œ๋“ค์ด ์„ฑ๋Šฅ์— ์–ด๋–ค ์˜ํ–ฅ์„ ์ฃผ๋Š”์ง€ ์‹คํ—˜
  • CDAE์™€ ์ตœ์‹  top-N ์ถ”์ฒœ๋ชจ๋ธ๋“ค ๋น„๊ต (CDAE๊ฐ€ ํฐ ์ฐจ์ด๋กœ ์„ฑ๋Šฅ์ด ๋” ์ข‹๋‹ค๋Š” ๊ฒƒ๋„ ํ™•์ธ)
  • q = 0, q = 1
    → M1์—์„œ q = 0์ผ ๋•Œ MAP๊ฐ’์ด ํ›จ์”ฌ ๋‚ฎ์Œ
  • input vector๋ฅผ ๊ทธ๋Œ€๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด ์ „์ฒด๋ฅผ drop outํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ๋” ์ข‹์€ ๋ชจ๋ธ์„ ๋‚ด์ง€ ๋ชปํ•จ
  • non-linear function์€ ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ์™„ํ™”์‹œํ‚ด

mapping function

  • hidden layer์™€ output layer์— identitify/sigmoid function ์กฐํ•ฉ์‹คํ—˜

 

728x90
๋ฐ˜์‘ํ˜•