Adaptive non European advertising retrieval system amcad

Chart sign (Graph Embedding) It is one of the most popular methods in the field of information retrieval in recent years , But they are often modeled in flat Euclidean space . Recent studies have found that the mapping structure of models under construction in Euclidean space is like hierarchy 、 There are natural errors in the ring structure , In this paper , We will introduce how to use non Euclidean graph representation to improve the modeling accuracy of complex heterogeneous graphs and achieve online revenue under the Alibaba mom search advertising scenario . This work has been recognized by the international data engineering summit ICDE 2022 (International Conference on Data Engineering) Included , The corresponding framework has also been open source .

Paper title ：

AMCAD: Adaptive Mixed-Curvature Representation based Advertisement Retrieval System

Download link ：

https://arxiv.org/abs/2203.14683

Open source address ：

https://github.com/alibaba/Curvature-Learning-Framework

This paper proposes an adaptive non European representation advertisement retrieval system AMCAD, The mixed curvature space is applied to large-scale industrial data for the first time , Make the model automatically learn the representation space with the lowest loss according to the complex data structure end-to-end .

be based on CurvLearn（ Open source address ：https://github.com/alibaba/Curvature-Learning-Framework） Non Euclidean depth learning operator provided ,AMCAD It aims to use a variety of curvature spaces to represent different heterogeneous nodes , By automatically learning spatial curvature and dynamically combining weights, we can achieve high-precision representation of complex heterogeneous graphs . The following figure shows AMCAD The specific design of , On the left is the overall structure of the model , There are three stages from bottom to top ：

AMCAD The overall architecture , It is divided into point level adaptive mixed curvature coding 、 Edge level space mapping and subspace distance fusion , Respectively corresponding to the mixed curvature representation of complex structures 、 Adaptive edge space of heterogeneous nodes and weight fusion of multiple geometric spaces