Recommendation systems have now penetrated into all kinds of products on the Internet. Whether it is shopping on an e-commerce website, reading a news website for information, or even wanting to hear different music when traveling, different kinds of recommendation systems play a pivotal role in our lives.
In order to give you an all-round interpretation of the core technology of the recommendation system, the author has carefully organized 21 cards and recommends your collection.
1. Analysis of modern recommendation architecture
Before discussing any kind of architecture, let's first look at what kind of problems this architecture needs to solve. Then under the guidance of these problems, we can analyze the advantages and disadvantages of different architectures in solving these problems.
So, for a recommendation architecture, what kind of problems do we need to solve? There are the following 3 points:
It can provide users with the current recommendation results within one or two hundred milliseconds;
Respond to the results of user interactions with the system;
Consider the issue of coverage of user groups.
The following three cards explain the three most common modern recommendation architectures from shallow to deep.
2. Simple recommendation model
We start with 3 simple recommendation models. Be
3. Models based on latent variables
Through the assumptions of the model, we know the relationship between the hidden variables, but we do not know the value of the hidden variables for the time being. Therefore, it is necessary to determine the actual value of the hidden variable through the "inference" process. When we know the values of these hidden variables, we can predict and analyze future data based on these values.
Hidden variables often also carry the assumption of "statistical distribution". The simplest latent variable model is the Gaussian mixture model.
4. Advanced Recommendation Model
Below, are three common high-level recommendation models, namely: tensor decomposition model, collaborative matrix decomposition and optimization of complex objective functions. Be
5. Recommended Exploit and Explore algorithms
A recommendation system, if one-sidedly optimizes user preferences, is likely to lead to stereotyped recommendation results. Therefore, EE algorithm is needed to realize personalized recommendation. Here are the two most common EE algorithms - UCB algorithm and Thompson sampling algorithm.
6. Recommendation model based on deep learning
After talking about the EE algorithm, let's talk about a more cutting-edge topic in the field of recommendation system research, that is, how to use deep learning to improve the accuracy of recommendation systems. This paper mainly introduces the application of Restricted Boltzmann Machine (RBM), Recurrent Neural Network (RNN) and Multi-layer Neural Network in recommender system.
7. Evaluation of recommender systems
Finally, I will share with you how to evaluate the recommendation system. This topic is very important and involves how to continuously evaluate a recommender system so as to improve the accuracy of the recommender system. Evaluation is divided into: offline evaluation, online evaluation and unbiased estimation.
∞∞∞∞∞
The official account replied "Machine Learning" ,
Invite you to join {IT AI Machine Learning Group}