[Editor’s Note] In this era of rapid development of science and technology and information explosion, it is no exaggeration to say that recommendation system has been completely integrated into our life. What restaurants we go to, what clothes we buy, what kinds of information we read and what kinds of videos we watch are largely determined by the recommendation system behind us.

In this paper, researchers from the social computing group of Microsoft Research Asia forecast the development direction of recommendation system in the future from five aspects: deep learning, knowledge graph, reinforcement learning, user portrait and interpretable recommendation. The full article has been reprinted with permission from Microsoft Research AI Headline (ID: MSRAsia). Article dry goods more, suggest collection!

“Guess you like it”, “Users who have purchased this product have also purchased…” Personalized recommendation is nothing new for modern Internet users who are inseparable from social networking platforms, e-commerce, news reading and life services.

With the development of information technology and Internet industry, information overload has become a challenge for people to deal with information. For users, how to quickly and accurately locate the content they need in the exponential growth of resources is a very important and challenging thing. For businesses, how to present the right items to users in a timely manner, so as to promote trading volume and economic growth, is also a very difficult thing. The birth of recommendation system greatly alleviates this difficulty.

Recommendation system is a kind of information filtering system, which can learn users’ interests and hobbies according to users’ files or historical behavior records, and predict users’ ratings or preferences for a given item. It has changed the way businesses communicate with users and strengthened the interaction between users and businesses.


Recommendations reportedly account for 35% of Amazon’s sales, 75% of Netflix’s spending, and 60% of the views on Youtube’s home page come from recommendations.

Therefore, how to build an effective recommendation system is of far-reaching significance. We will take a look at the future of recommendation systems from the application of deep learning, the application of knowledge graph, the application of reinforcement learning, user portrait and interpretable recommendation. This paper will discuss the application of deep learning in recommendation system.

Recommendation systems and deep learning

In recent years, deep learning technology has achieved great success in speech recognition, computer vision, natural language understanding and other fields. How to apply it to recommendation system is the current research hotspot. The current application of in-depth recommendation system is mainly reflected in the following three aspects:

  1. Improve representational learning ability. The advantage of deep neural network lies in its strong representation learning ability. Therefore, one of the most direct applications is to use deep learning technology to learn effective implicit factor feature representation from complex content data, so that it can be easily used by the recommendation system.
  2. Deep collaborative filtering. The classical matrix decomposition model can be described as a very simple neural network. We can enhance the function of the recommendation model by expanding its structure and introducing more nonlinear elements. For example, in WWW 2017 paper Neural Collaborative Filtering, the author proposed an enhanced matrix decomposition model. On the one hand, it makes up for the weakness that the naive dot product operation of two hidden vectors cannot distinguish the importance difference between dimensions. On the other hand, it introduces an additional multi-layer perceptron module to introduce more nonlinear operations. In addition, deep learning related technologies such as automatic encoding machine, convolutional neural network, memory network and attention network have also been applied to improve the traditional collaborative filtering model and achieved good results.
  3. Deep interaction between features.In order to improve the accuracy of the model, enterprise-level recommendation systems often use rich and even heterogeneous content data. These features present different information from different dimensions, and the combination of features is often very meaningful. Traditional crossover features are manually designed by engineers, which are very limited, expensive, and can’t be extended to crossover patterns that haven’t existed before. So scholars began to studyNeural network is used to automatically learn high-order feature interaction patterns to make up for the limitations of artificial feature engineering.Models related to this layer include Wide&Deep, PNN, DeepFM, DCN, and our recently proposed xDeepFM model (xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems, KDD 2018).

Deep learning technology has a broad application prospect in recommendation system. Here are a few possible future research directions:

1. Efficiency and scalability

For the industrial recommendation system, not only the accuracy of the model, but also the efficiency and maintainability are very important aspects. Efficiency refers to the fact that when a user sends a request, the recommendation system can return the result in near real time, without making the user wait. Maintainability means that the system is easy to deploy and can support periodic or incremental updates. As is known to all, complex neural networks require a huge amount of computation, so how to apply them more efficiently on the super-large recommendation platform is a technical difficulty that needs to be solved urgently.

2. Diversified data fusion

In real platforms, the data of users or objects is often complex and diverse. The contents of items can include text, images, categories and other data; User behavior data can come from multiple domains, such as social networks, search engines, news reading apps, etc. The feedback of users’ behaviors can also be rich and varied. For example, in e-commerce websites, users’ behaviors may include searching, browsing, clicking, favorites and purchasing, etc. Not only that, but within these dimensions, the distribution of data for different users or items varies greatly; The data amount of user’s behavior feedback is also different. The data amount of click behavior is often much larger than that of purchase behavior. Therefore, a single, homogeneous model cannot effectively deal with these diverse data. How to deeply integrate these complex data is a technical difficulty.

3. Capture users’ short-term and long-term preferences

User preferences can be roughly divided into long-term and short-term categories. Long-term preference usually refers to the user’s interests. For example, if she is a Mayday fan, she will be interested in Mayday’s songs and concert tickets for a long time in the future. Short-term preference refers to the user in the current environment of immediate interests, such as the latest week users prefer listening to trill hits, then the recommendation system should capture the user of this interest, or the user in the future a month have plan to move, then the recommendation system can properly push some moving company advertising. At present, some popular methods are to combine recurrent neural networks with deep collaborative filtering technology to achieve both short-term and long-term memory functions. How to combine user’s long-term preference and short-term demand more closely and effectively with the influence of situational factors is also a research hotspot.


In the next article, we will discuss the research of “recommendation system and knowledge graph”. If you want to know more research hotspots about recommendation system, please continue to pay attention. Welcome to share and collect!