[Recommendation system and machine learning introduction] Summary of a RD engineer’s recommendation system learning road, more knowledge of recommendation system please pay attention to RecomAI


As a graduate of CS major, I was just a RD development engineer of e-commerce business when I first entered the Internet field. Later, by chance, I turned to the field of big data, and then entered the development work of recommendation system. Review these years of work, summed up some experience, hope to those who are interested in entering the recommendation system related field of students have some help.


Recommendation/search/advertising are the three mature carriages of machine learning at present, which are of great significance to the business realization of various Internet companies. There are some differences between these three businesses in dimension, for example, the goal of search is to reduce the delay and increase the relevance of Query, and the goal of advertising is to increase CTR. The target of video recommendation system is comprehensive indexes such as playback time and click. But in terms of data flow and systems engineering architecture, they are very similar.


For those of you who are interested in or new to recommendation systems, you should have read the following two books





A bigger problem is that if there is no recommended to participate in the actual system development, or no laboratory Buff, cow force feeling each chapter to see clearly, but want to write a simple recommendation system but do not know how to start, unable to imagine like Meituan ali headlines recommendation system is how to fall to the ground.


In my opinion, there are several reasons for this situation


1. Recommendation system is a system

Recommendation system is a system at first. As the core of profitable Internet companies, it is bound to be very complicated in overall architecture, such as involving NLP, recall, Sorting, log washing and log calculation model. The recommendation method mentioned in the book is single, for example, CF(collaborative filtering) is used for recommendation. In practice, recommended recall usually has multiple paths, so as to have diversity. (There’s hardly even a mention of candidate sets, which exist in every recommendation system.)


2. Closely associated with the algorithm

Recommendation system is one of the early fields in which algorithms are implemented. It has large-scale applications to traditional machine learning algorithms such as LR, GBDT and FM. One of the basic capabilities of traditional machine learning is feature engineering, which is very related to individual domain knowledge. The specific industrial realization of each is not the same, especially for reference. As for the deep learning model, if you want tf to write a code and run the CTR will rise several points directly, which is a dream? What effect can specific network design have on your business model and behavioral data is very much related to, it is very likely that your deepFM is worse than your LR+GBDT


3. Data issues

Importance of data, first of all, if it is a small system, such as only tens of thousands of PV, far from using recommendation system improvement effect, industry recommendation system, log is the blood, including exposure log, click on the log, collecting journal and so on behavior, daily log volume will be very big, the logs on the one hand used to update the model, On the other hand, enter the experimental statistical platform to calculate the effect of each algorithm in the past time. In addition, for example, FM factorization machine model, if you have very few logs, it will be a super, super sparse matrix, and the final model effect will not be very good, especially now that Deep learning is widely implemented in the industrial street (Wide & Deep), the recommendation system without large-scale logs will be a source of water and a tree without roots.


Because of these reasons, I created this public account [Recommendation System and Machine Learning], hoping to share some common architectures and basic knowledge of industrial recommendation system, as well as some research and progress in the forefront of information retrieval field, hoping to make progress together with you.


,

Long press the QR code to follow

Recommendation systems and machine learning

ID: RecomAI