Heterogeneous graph neural network for recommendation system problems (ackrec, hfgn)

Continue to organize several articles SIGIR2020 Of Graph+Recommendation The paper of , Other papers can be sorted out according to previous blog posts of bloggers （ Portal ）, This article only arranges two more interesting heterogeneous diagrams + Recommended articles .

Attentional Graph Convolutional Networks for Knowledge Concept Recommendation in MOOCs in a Heterogeneous View（ACKRec）
The background is mooc Recommend videos to students on , The essence is still rating forecast . The difference is to deal with the sparse problem + There will be many videos in the course , Students' interest points / The concept of knowledge may be different , So the author tries to build a multi entity heterogeneous graph （ user , Course , The concept of knowledge , video , teacher ）. There are complex relationships between these entities , Here's the picture , Comparison teachers can give lectures , Lesson contains videos , Students come to class and watch these videos and can use repetitive behavior , The final goal is to learn the knowledge contained in the video .

The specific model architecture is shown in the figure , The main idea is to design meta paths to guide students' interests propagation chart , Because these heterogeneous complex relationships are also too suitable for meta path aggregation . say concretely , Is the development of heterogeneous GCN To aggregate according to the designed meta path , Then because different students have different interests Attention To represent a feature , The last downstream prediction is MF complete .

Hierarchical Fashion Graph Network for Personalized Outfit Recommendation（HFGN）
Personalized clothing recommendation . It is mainly to unify the compatibility of fashion outfit-item（ The same suit is compatible up and down ） And personalized recommendations user-outfit（ Conform to the user's dress style , In particular, I may only like some of them item） These two points .

So follow this line of thought , Heterogeneous graph networks can be used for unified modeling users, items, and outfits, because outfit There will be many item form , And loved by users , Then it will become a layered diagram like the above figure . Specifically, the model contributes a hierarchical graph convolution , There are three kinds of information dissemination ：item Between ,item To outfit,outfit To user, After finishing the communication, do two tasks ： Compatible and personalized recommendations . The structure is as follows

• 1) Cross item dissemination of information , Refine item embedding by combining compatibility modeling . It's good to focus on this part , The formula is as follows ：$$m_{i^{\prime }-i}=w(i,i^{\prime })\sigma (W_1(i\odot i^{\prime }))$$$$i^{\ast }=i+\sum m_{i^{\prime }-i}$$
• 2) Information spreads from items to outfit, It aggregates project semantics into outfit in . This part is based on lightGCN Ideas to aggregate , Bloggers have sorted it out in the past , Portal ： Figure neural network is used to recommend system problems （NGCF,LightGCN）. The formula is as follows ：$$m_{i-o}=\frac{1}{|N_o|}\sigma (W_2{i}^{\ast })$$$$o^{\ast }=o+\sum m_{i-o}$$
• 3) Information from outfit To the user , Integrate historical equipment into user representation . This part is the same as the previous one ：$$m_{o-u}=\frac{1}{|N_u|}\sigma (W_3{o}^{\ast })$$$$u^{\ast }=u+\sum m_{o-u}$$
• Last use BPR Just optimize .