[Editor’s Note] 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.

In the previous articles, we introduced deep learning, knowledge graph, reinforcement learning, application of user portrait in recommendation system and possible future research directions respectively. In the final article of the day, we will cover the interpretability of recommendation systems.

Most of the recommendation system studies mentioned above focus on improving the accuracy of recommendations, and the communication with recommended objects is not considered enough. Recently, scholars have begun to pay attention to whether recommendation can fully grasp user psychology and give appropriate examples to communicate with users in a way that users can easily accept. It is found that such a system can not only improve system transparency, but also improve users’ trust and acceptance of the system, the probability of users choosing recommended products and the degree of user satisfaction. Designing such an interpretable recommendation system is our ultimate goal.

As a less explored direction in the field of recommendation, it can explain that many aspects of recommendation are worth studying and exploring. At present, we are considering the following three aspects for research.

1. Use knowledge graph to enhance the interpretation ability of the algorithm. As a highly readable external knowledge carrier, knowledge graph provides great possibilities for improving algorithm interpretation ability.

The existing explainable recommendations generate recommendation interpretations which are usually limited to one of the media of objects, users or features, and the correlation among these three media is not sufficiently mined. We hope to use the knowledge graph to connect these three media and flexibly select the most appropriate media to recommend and explain to users according to the specific situation.

In addition, it is possible to establish readable depth structures between features by using conceptual graphs such as Microsoft Concept Graph, thus replacing the current deep learning network with very weak interpretation to improve the readability and ensure the accuracy of the algorithm.

In an era of increasing importance of explainable artificial intelligence, the combination of symbolic knowledge such as knowledge graphs and deep learning is a promising direction.

2. Model-independent interpretable recommendation frameworks. At present, most interpretable recommendation systems are designed for specific recommendation models, and their extensibility is weak. For emerging recommendation models, such as complex and mixed models with deep neural networks, the interpretable ability is not enough. If there is a model-independent interpretable recommendation framework, it can avoid designing interpretation schemes separately for each recommendation system, thus improving the extensibility of the method.

We have made A preliminary attempt on this (A Reinforcement Learning Framework for Explainable Recommendation, ICDM2018). In this work, we propose the following reinforcement learning framework (Figure 1) to interpret any recommendation model while ensuring extensibility, explanatory power, and explanatory quality.

       

Figure 1: Model-independent explainable recommendation reinforcement learning framework

3. Make conversational recommendation based on generation model. At present, the form of recommendation explanation is often pre-set and monotonous (for example, the pre-set recommendation explanation takes the user as the medium), so although some examples can be cited according to the user’s psychology, it is still too rigid in the way of communication.

If the generation model can be used to make the recommendation system “create” a smooth sentence or even high eq, it can carry out flexible and changeable recommendation and interpretation in the process of chatting with users. Our team worked with Microsoft Xiaoice to try some of this, generating music recommendation explanations for Xiaoice.

      

We believe that future recommendation systems should further consider the efficiency and extensibility of recommendation algorithms, integrate multi-source heterogeneous user behavior data, and capture users’ long-term and short-term preferences. It is an important research direction to combine knowledge graph reasoning in recommendation system, design learning mechanism of general strategy, and obtain effective decision model from limited user interaction data.

In terms of interpretability, we need to use knowledge graph to enhance algorithm interpretation ability, design model-independent interpretable recommendation framework, and consider conversational recommendation combined with generative model.

Finally, we need to pay attention to user privacy, design a mechanism for sharing user data between different platforms, and establish a unified user representation model for recommendation systems.

We believe that personalized recommendation system will continue to evolve in different directions of accuracy, diversity, computational efficiency and interpretability, and finally solve the problem of user information overload.

If you want to know more about the research hotspot of recommendation system, please keep paying attention.

Related reading:

Dry goods | five research hot spots of personalized recommendation system user portrait (4)

Reinforcement learning of five Research hotspots of personalized Recommendation System (III)

Dry goods | personalized recommendation system five research hot spots of knowledge map (2)

Dry goods | personalized recommendation system five research hot spots of deep learning (a)

Build recommendation system quick start, just five steps!

Welcome everyone to like, collect, and share more technical knowledge with your friends.

This account is the official account of the first recommendation of the fourth paradigm intelligent recommendation product. This account is based on the computer field, especially the cutting-edge research related to artificial intelligence. It aims to share more knowledge related to artificial intelligence with the public and promote the public’s understanding of ARTIFICIAL intelligence from a professional perspective. At the same time, we also hope to provide an open platform for discussion, communication and learning for people related to ARTIFICIAL intelligence, so that everyone can enjoy the value created by artificial intelligence as soon as possible.

Every member of the fourth paradigm has contributed to the landing of artificial intelligence. Under this account, you can read the academic frontier, knowledge and industry information from the computer field.

For more information, welcome to search and follow the official Weibo, wechat (ID: DSFSXJ) public account.