当前位置：网站首页>This article takes you to understand the commonly used models and frameworks of recommender systems

This article takes you to understand the commonly used models and frameworks of recommender systems

2022-08-02 09:34:00 【timerring】

可以看KDD会议,Recent Recommender Systems Papers.

推荐系统概述

传统推荐模型Old school Model

协同过滤模型

through the relationship between users,The information is filtered together with the user's evaluation feedback on the item,So as to find the information that the target user is interested in.

用户—A rating matrix for the item（The matrix is likely to be sparse）


x		x
x	x
	x	x

Row vectors represent each user's preferences,A column vector indicating the attributes of each item

Based on a scoring matrix（行列）计算相似度,Here are some ways to calculate similarity：

余弦相似度
皮尔逊相关系数
欧氏距离
曼哈顿距离

There are mainly user-based collaborative filtering and item-based collaborative filtering.

矩阵分解模型

Matrix decomposition is the product of two low-rank matrices,By the inner product of the two matrices after decomposition,来填补缺失的数据.

优点:思路简单,Prediction can be done easily;

缺点:It is difficult to train incrementally（when the sample surges,May have to rebuild the matrix）,Feature fusion is difficult;

这里kis a hidden factor,Equivalent to a hyperparameter.

逻辑回归模型

Predict whether users will“点击商品”进行分类.into a classification model.

$\begin{aligned} \phi(x) &=w_{0}+w_{1} x_{1}+\cdots+w_{n} x_{n} \\ &=w_{0}+\sum_{i=1}^{n} w_{i} x_{i} \end{aligned}$

优点:模型简单,可解释性强,训练速度快(SGD梯度下降);

缺点:Model modeling capabilities are limited（没有考虑特征之间的相关性,and the intersection between features）,Manual feature engineering is required;

特征交叉模型

PLOY2

$\phi(x) = w_{0}+\sum_{i = 1}^{n} w_{i} x_{i}+\sum_{i = 1}^{n-1} \sum_{j = i+1}^{n} w_{i j} x_{i} x_{j}$

Violence is added to logistic regression二阶特征交叉.

优点:Add second-order features,Enhanced modeling capabilities;

缺点:时间复杂度高 $n-->n^2$ ;

Factorization Machine

$\phi(x)=w_{0}+\sum_{i=1}^{n} w_{i} x_{i}+\sum_{i=1}^{n-1} \sum_{j=i+1}^{n}\left\langle v_{i}, v_{j}\right\rangle x_{i} x_{j}$

Add implicit weights to each feature(Inner product between two vectors),as the weight of feature intersection.

优点∶相比于PLOY2Reduced the amount of model parameters（ $n^2-->nK$ ）,自动特征工程

缺点︰Feature intersection is limited(二阶)

GBDT+LR

GBDT:作为特征编码器;It is mainly used for feature filtering and feature encoding of input data,Generate discrete feature vectors

LR（逻辑回归）︰Use the encoded results for training

优点︰灵活,Suitable for adding new features（Use tree model for feature combination）

缺点：The tree model has high complexity

深度推荐模型

Deep Collaborative Filtering(Neural CF )

Treat user ratings of items as a classification problem.

Learn user interactions with items using fully connected layers.

Replacing the matrix factorization operation with a multi-layer neural network

Using a fully connected network may be a little more efficient than multiplying.

Wide & Deep

基本淘汰

Wide为线性模型,Deepis a deep model

浅层模型(记忆能力）and deep model models（泛化能力）,

WideParts can be rememberedid,Make a model of this.类似于LR.

DeepIt can be regarded as a fully connected network,类似于NCF.

DeepFM

DeepFM包含FM和DNN两部分,两部分共享输入特征.使用FM替换wide & Deep中的wide部分.

DeepFM:一阶特征+二阶特征+深度特征

Abandon the previous orderWide部分,用FM代替,Enhance the ability to combine shallow features,Substitute first and second order.

DIN

首个加入Attention机制

Adjust weights based on users and items

文本编码方法Text Encoding

Count:Count the number of text characters、单词个数
LabelEncoder:Unified labeling
Multi One-Hot:Multi-value label encoding(例如one-hotAdd after encoding)
AB : 011 BC : 110 AC : 101
One-Hot：eg：A： 0 0 1 B：010 C：100
CounterVector:与Multi One-Hot,But join the count
TfidfVectorizer: 次数和词频统计
Word2Vec:词向量映射,然后聚合

原网站

版权声明
本文为[timerring]所创，转载请带上原文链接，感谢
https://yzsam.com/2022/214/202208020919396000.html