Explainable Recommendation

Recommendation system is widely used in many fields. The more traditional algorithm that relies mainly on statistics is relatively naive, while the latest AI-based models have the characteristics of black boxes - results hard to predict, algorithms hard to adjust, process hard to explain, etc. The application of recommendation systems is substantially constrained by shortcomings, especially in scenarios where great importance is attached to reliability, where recommendation systems are given little trust. On the other hand, the use of user data in recommendation systems is controversial due to the involvement of private data as the training sample and the input to a recommendation system. As more and more significance is placed on user privacy, recommendation systems might not be able to grasp as much data from users as before, which would create hurdles for such systems.

Combined with technologies like knowledge graph, a more reliable and explainable recommendation system can be established. The data extracted through information extraction system and retrieved via knowledge graph can be exported as some kind of reference or basis of the process, which could to a certain extent explain the final output and could also provide additional information for further optimization. With the help of the knowledge graph, the use of private data could be more restrained, reducing the collection of user’s private information to the smallest scale instead of taking all of user’s operation data as input.

Magi KG offers abundant knowledge graph data, introducing more dimensions of data for reference. Moreover, each piece of knowledge has the original source where it was learned, which is critical for a more precise and more logical output. Modules like Ireul and Leliel are capable of breaking down long texts or questions into the key information, increasing the accuracy of the system. Several modules are suitable to be deployed on the user’s device for the protection of privacy.