# Features¶

## Overview¶

With the great success of deep learning,DNN-based techniques have been widely used in CTR prediction task.

DNN based CTR prediction models usually have following 4 modules:
`Input,Embedding,Low-order&High-order Feature Extractor,Prediction`

- Input&Embedding

The data in CTR estimation task usually includes high sparse,high cardinality categorical features and some dense numerical features.

Since DNN are good at handling dense numerical features,we usually map the sparse categorical features to dense numerical through`embedding technique`

.

For numerical features,we usually apply`discretization`

or`normalization`

on them.

- Feature Extractor

Low-order Extractor learns feature interaction through product between vectors.Factorization-Machine and it’s variants are widely used to learn the low-order feature interaction.

High-order Extractor learns feature combination through complex neural network functions like MLP,Cross Net,etc.

## Feature Columns¶

### SparseFeat¶

`SparseFeat`

is a namedtuple with signature `SparseFeat(name, vocabulary_size, embedding_dim, use_hash, dtype,embedding_name, group_name)`

- name : feature name
- vocabulary_size : number of unique feature values for sprase feature or hashing space when
`use_hash=True`

- embedding_dim : embedding dimension
- use_hash : defualt
`False`

.If`True`

the input will be hashed to space of size`vocabulary_size`

. - dtype : default
`float32`

.dtype of input tensor. - embedding_name : default
`None`

. If None, the embedding_name will be same as`name`

. - group_name : feature group of this feature.

### DenseFeat¶

`DenseFeat`

is a namedtuple with signature `DenseFeat(name, dimension, dtype)`

- name : feature name
- dimension : dimension of dense feature vector.
- dtype : default
`float32`

.dtype of input tensor.

### VarLenSparseFeat¶

`VarLenSparseFeat`

is a namedtuple with signature `VarLenSparseFeat(sparsefeat, maxlen, combiner, length_name, weight_name,weight_norm)`

- sparsefeat : a instance of
`SparseFeat`

- maxlen : maximum length of this feature for all samples
- combiner : pooling method,can be
`sum`

,`mean`

or`max`

- length_name : feature length name,if
`None`

, value 0 in feature is for padding. - weight_name : default
`None`

. If not None, the sequence feature will be multiplyed by the feature whose name is`weight_name`

. - weight_norm : default
`True`

. Whether normalize the weight score or not.

## Models¶

### CCPM (Convolutional Click Prediction Model)¶

CCPM can extract local-global key features from an input instance with varied elements, which can be implemented for not only single ad impression but also sequential ad impression.

**CCPM Model API**
CCPM

### FNN (Factorization-supported Neural Network)¶

According to the paper,FNN learn embedding vectors of categorical data via pre-trained FM. It use FM’s latent vector to initialiaze the embedding vectors.During the training stage,it concatenates the embedding vectors and feeds them into a MLP(MultiLayer Perceptron).

FNN

### PNN (Product-based Neural Network)¶

PNN concatenates sparse feature embeddings and the product between embedding vectors as the input of MLP.

PNN

### Wide & Deep¶

WDL’s deep part concatenates sparse feature embeddings as the input of MLP,the wide part use handcrafted feature as input. The logits of deep part and wide part are added to get the prediction probability.

WDL

### DeepFM¶

DeepFM can be seen as an improvement of WDL and FNN.Compared with WDL,DeepFM use FM instead of LR in the wide part and use concatenation of embedding vectors as the input of MLP in the deep part. Compared with FNN,the embedding vector of FM and input to MLP are same. And they do not need a FM pretrained vector to initialiaze,they are learned end2end.

DeepFM

### MLR(Mixed Logistic Regression/Piece-wise Linear Model)¶

MLR can be viewed as a combination of $2m$ LR model, $m$ is the piece(region) number. $m$ LR model learns the weight that the sample belong to each region,another m LR model learn sample’s click probability in the region. Finally,the sample’s CTR is a weighted sum of each region’s click probability.Notice the weight is normalized weight.

MLR

### NFM (Neural Factorization Machine)¶

NFM use a bi-interaction pooling layer to learn feature interaction between embedding vectors and compress the result into a singe vector which has the same size as a single embedding vector. And then fed it into a MLP.The output logit of MLP and the output logit of linear part are added to get the prediction probability.

NFM

### AFM (Attentional Factorization Machine)¶

AFM is a variant of FM,tradional FM sums the inner product of embedding vector uniformly. AFM can be seen as weighted sum of feature interactions.The weight is learned by a small MLP.

AFM

### DCN (Deep & Cross Network)¶

DCN use a Cross Net to learn both low and high order feature interaction explicitly,and use a MLP to learn feature interaction implicitly. The output of Cross Net and MLP are concatenated.The concatenated vector are feed into one fully connected layer to get the prediction probability.

DCN

### DIN (Deep Interest Network)¶

DIN introduce a attention method to learn from sequence(multi-valued) feature. Tradional method usually use sum/mean pooling on sequence feature. DIN use a local activation unit to get the activation score between candidate item and history items. User’s interest are represented by weighted sum of user behaviors. user’s interest vector and other embedding vectors are concatenated and fed into a MLP to get the prediction.

DIN

### DIEN (Deep Interest Evolution Network)¶

Deep Interest Evolution Network (DIEN) uses interest extractor layer to capture temporal interests from history behavior sequence. At this layer, an auxiliary loss is proposed to supervise interest extracting at each step. As user interests are diverse, especially in the e-commerce system, interest evolving layer is proposed to capture interest evolving process that is relative to the target item. At interest evolving layer, attention mechanism is embedded into the sequential structure novelly, and the effects of relative interests are strengthened during interest evolution.

DIEN

### xDeepFM¶

xDeepFM use a Compressed Interaction Network (CIN) to learn both low and high order feature interaction explicitly,and use a MLP to learn feature interaction implicitly. In each layer of CIN,first compute outer products between $x^k$ and $x_0$ to get a tensor $Z_{k+1}$,then use a 1DConv to learn feature maps $H_{k+1}$ on this tensor. Finally,apply sum pooling on all the feature maps $H_k$ to get one vector.The vector is used to compute the logit that CIN contributes.

CIN

xDeepFM

### AutoInt(Automatic Feature Interaction)¶

AutoInt use a interacting layer to model the interactions between different features. Within each interacting layer, each feature is allowed to interact with all the other features and is able to automatically identify relevant features to form meaningful higher-order features via the multi-head attention mechanism. By stacking multiple interacting layers,AutoInt is able to model different orders of feature interactions.

InteractingLayer

AutoInt

### ONN(Operation-aware Neural Networks for User Response Prediction)¶

ONN models second order feature interactions like like FFM and preserves second-order interaction information as much as possible.Further more,deep neural network is used to learn higher-ordered feature interactions.

ONN

### FGCNN(Feature Generation by Convolutional Neural Network)¶

FGCNN models with two components: Feature Generation and Deep Classifier. Feature Generation leverages the strength of CNN to generate local patterns and recombine them to generate new features. Deep Classifier adopts the structure of IPNN to learn interactions from the augmented feature space.

FGCNN

### DSIN(Deep Session Interest Network)¶

Deep Session Interest Network (DSIN) extracts users’ multiple historical sessions in their behavior sequences. First it uses self-attention mechanism with bias encoding to extract users’ interests in each session. Then apply Bi-LSTM to model how users’ interests evolve and interact among sessions. Finally, local activation unit is used to adaptively learn the influences of various session interests on the target item.

DSIN

### FiBiNET(Feature Importance and Bilinear feature Interaction NETwork)¶

Feature Importance and Bilinear feature Interaction NETwork is proposed to dynamically learn the feature importance and fine-grained feature interactions. On the one hand, the FiBiNET can dynamically learn the importance of fea- tures via the Squeeze-Excitation network (SENET) mechanism; on the other hand, it is able to effectively learn the feature interactions via bilinear function.

FiBiNET