Kdd2022 | what features are effective for interaction?

Paper title ：

Detecting Arbitrary Order Beneficial Feature Interactions for Recommender Systems

Thesis link ：

https://arxiv.org/pdf/2206.13764.pdf

Code ：

https://github.com/ruizhangai/HIRS_Hypergraph_Infomax_Recommender_System.git

This paper mainly puts forward the relevant methods for feature interaction in Recommendation System , If all possible features are interacted , That consumption is great , In this paper, HIRS Used to directly generate beneficial feature interactions . The number of generated feature interactions can be specified to be much smaller than the number of all possible interactions , Therefore, the running time of the model is shorter .

In this paper, the concepts of hypergraph and mutual information are used to automatically generate effective feature interaction . Take the features in the sample as nodes , If there is interaction between features, there are superedges between these feature nodes .

• First , After obtaining the sample , Take the sample characteristics as nodes , Map to the corresponding embedding, The edge set is empty
• Utilization node embedding Predicted super edge set
• Use nodes and super edges to predict the results of the recommendation system
• Some points need to be considered in the process of Super Edge prediction , For example, if the generated interaction is beneficial and effective , The resulting super edge matrix takes 0,1 sparse , So for these problems , The author makes corresponding improvements based on mutual information .