\

Algorithm and model are the core of recommendation system, which directly determine the effect of recommendation. In recommendation system, algorithm should be considered from two aspects: algorithm accuracy and algorithm efficiency. Compared with algorithm accuracy, recommendation system engineering pays more attention to algorithm efficiency. The algorithm and model levels are mainly included as shown in the figure.

From the theory of recommendation algorithm, it can be mainly from the recall layer and the sorting layer. Although the recall layer and the sorting layer are both done by using algorithms and models, there are certain differences in the models used for different stages.

In the recall layer, we can divide into three types of methods, namely, traditional algorithm, machine learning and deep learning model.

So-called traditional algorithm is adopted by using the theory of traditional recommendation system in the algorithm, such as computing the similarity between the user, similarity between objects, etc., and from these two kinds of algorithm in general, can be said with the collaborative filtering algorithm, only one is user-based collaborative filtering algorithm, a collaborative filtering algorithm based on item.

The so-called machine learning Model refers to the use of machine learning correlation algorithm to get recall layer results, such as NMF (Nonnegative Matrix Factor, non-negative Matrix decomposition) algorithm, LFM Model (Latent Factor Model, Based on latent hidden factor model), FTRL (Follow The Regularized Leader) algorithm, PersonalRank algorithm, etc.

The so-called deep learning model refers to the methods that use deep learning correlation algorithms to get recall layer results, including deep neural network (DNN) and Embedding.

In the sorting layer, we can also use machine learning and deep learning to sort the recall results. Click rate estimation is generally used in the sorting, and the probability descending sorting is calculated by taking the first N to get the final sorting result.

When using machine learning for sorting layer sorting, we can also divide it into two categories, namely linear model and tree model. In fact, we will combine these two types of models in the real application process.

  • In linear model, logistic regression is mainly used as the main sorting algorithm, and logistic regression is a kind of deformation of linear regression. Therefore, if you want to learn logistic regression well, you must have a certain understanding of linear regression.
  • Tree model is a general name of the most widely used model in sorting algorithm. The tree models commonly used in sorting layer are divided into decision tree model and integrated learning model, and they are random forest, Boosting, GBDT, XGB and LGB in recommendation system.

Although machine learning algorithms of sorting layer can be divided into linear model and tree model, they are generally combined in sorting, among which the common combination method is GBDT+LR for CTR prediction and sorting layer.

Deep learning model for CTR prediction and ranking layer ranking is also commonly used in recommendation systems, especially for a large number of data and features, which has gradually become the mainstream ranking method of major enterprises. The common methods are DeepFM and xDeepFM. As the name implies, DeepFM is a combination of depth model (Deep) and factor factorizer (FM). XDeepFM is an upgraded version of DeepFM, mainly improving the shortcomings of DCN model in DeepFM model.

END

Machine Learning Online Manual Deep Learning online Manual AI Basics download (PDF updated to25Set) site QQ group1003271085To join the wechat group, please reply to "add group" to get a discount station knowledge planet coupon, please reply to "knowledge planet" like the article, click on itCopy the code