2021.SIGIR.ConsisRec Enhancing GNN for Social Recommendation via Consistent Neighbor Aggregation

[soroush-ziaeinejad]

Main problem:

The main problem is trying to find a way to alleviate the cold start issue. This paper proposed a method to empower the GNN (that is used in other works) to resolve the social inconsistency problem.

Social inconsistency problem:

Two connected users in the social graph are not necessarily interested in the same topic. Also, being connected with an item does not always mean that the user is interested in the item because of the rating value. In other words, two users who both rated the same item can have completely different ideas about the item. (User A: 5 stars, user B: 1 star)

Existing shortcomings:

• Ignoring the social inconsistency problem and aggregating the information from two neighbors who have inconsistent relations or links.

ConsisRec:

• Embedding Layer: concatenating user and item instances and embedding them to a single entity called 'query'.

• Query Layer: extracting the consistent neighbors of the user, and the consistent neighbors of the item

• Neighbor Sampling: Dynamically samples social neighbors based on social items. This can boost the GNN power and improve ranking performance.

• Relation Attention: Aggregating the neighbor sampling output, the embeddings.

• Finally, the inner product of the (user embedding + user) and (item embedding + item) outputs the rating prediction and final loss can be calculated.

Inputs:

• Users
• Items
• Rating information

Outputs:

• Rating score: recommendation score of each item for each user, considering their consistent neighbors.

Experiments:

• Dataset:
  • Ciao [7, 317 users, 104, 975 items with 111, 781 social links]
  • Epinions [18, 069 users, 261, 246 items with 355, 530 social links]
• Metrics: RMSE, MAE

Baselines:

• SoRec: matrix factorization / CIKM08
• SoReg: social regularization / WSDM11
• SocialMF: matrix factorization/ RecSys10
• GCMC+SN: Graph Convolutional NN / ???17
• GraphRec: Graph NN / WWW19
• CUNE: Collaborative User Network Embedding / SIAM17

ConsisRec outperforms all these baselines on both datasets and improves both RMSE and MAE metrics at least for 1.7 percent.

Code:

The code of this paper is available here

Presentation:

There is no available presentation for this paper.

[hosseinfani]

@soroush-ziaeinejad 1) Where do they address the cold start item/user? Do they have a specific experiments for these scenarios? 2) I am thinking of using inconsistent neighbors as negative samples (in addition to positive/consistent neighbors) 3) One item your awesome reviews misses is the gaps of the current paper. Please add this item to this and previous summaries. Thank you.

[soroush-ziaeinejad]

@hosseinfani

1. They don't specifically address the cold-start problem. They mention that because of the cold-start problem, we have to use social information and connections to improve recommender systems. The main issue in social connection extracting that they want to address is social inconsistency.
2. Inconsistency is defined as existing irrelevant or opposite relation between users with explicit connections. Yes, finding users with opposite ideas can help to use them as negative samples.
3. Thanks Hossein, sure I will add this part to my summaries.

Thanks for the feedback.