Main Problem
The author tries to address the exisiting problem of collaborative filtering methods the inability of modelling side information , i.e., they model each interaction as independent data instance and do not consider their relations. The authors proposes Knowledge Graph Attention Network that provides a solution for above problem.
Related Work
The author categorizes the recent advancements made in collaborative knowledge graph which is a combination of knowledge graphs and User-Item Graph, into two categories: -
Path - based methods: - Which uses path-based algorithms or meta-paths to constrain the paths. The first problem with this two stage approach is that the first stage of path selection has a large impact on the final performance but not optimized for recommendation objective and meta-paths are a very labor intensive approach and requires high level of domain knowledge.
Regularization methods: - These methods lack explicit modeling of high-order relations in which long-range connectivities are often missed out and the results of high-order modeling are not interpretable.
Proposed Architecture
To address the above problem and limitations author proposes KGAT. The KGAT architecture is as follows, the following three points are unabridged text from the original paper.
1) embedding layer, which parameterizes each node as a vector by preserving the structure of CKG
2) attentive embedding propagation layers, which recursively propagate embeddings from a node’s neighbors to update its representation, and employ knowledge-aware attention mechanism to learn the weight of each neighbor during a propagation
3) prediction layer, which aggregates the representations of a user and an item from all propagation layers, and outputs the predicted matching score.
• Input: collaborative knowledge graph G that includes the useritem
bipartite graph G1 and knowledge graph G2.
• Output: a prediction function that predicts the probability yˆui
that user u would adopt item i.
A great paper that clearly explains the architecture.
User provides code and dataset links for us to implement.
Authors reveal the experimentation settings used for the evaluation model and the baselines they have used.
This paper was read because of two reasons: -
To understand this paper better: - Guo, S., Lin, Y., Feng, N., Song, C., & Wan, H. (2019, July). Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. In Proceedings of the AAAI conference on artificial intelligence (Vol. 33, No. 01, pp. 922-929).\
Since our problem can and does falls under high-order modeling domain where we are trying to capture more information related to user, this could be a good approach to begin with.
Title: - KGAT: Knowledge Graph Attention Network for Recommendation Year: - 2019 Venue: - KDD
Main Problem The author tries to address the exisiting problem of collaborative filtering methods the inability of modelling side information , i.e., they model each interaction as independent data instance and do not consider their relations. The authors proposes Knowledge Graph Attention Network that provides a solution for above problem.
Related Work The author categorizes the recent advancements made in collaborative knowledge graph which is a combination of knowledge graphs and User-Item Graph, into two categories: -
Proposed Architecture To address the above problem and limitations author proposes KGAT. The KGAT architecture is as follows, the following three points are unabridged text from the original paper.
1) embedding layer, which parameterizes each node as a vector by preserving the structure of CKG 2) attentive embedding propagation layers, which recursively propagate embeddings from a node’s neighbors to update its representation, and employ knowledge-aware attention mechanism to learn the weight of each neighbor during a propagation 3) prediction layer, which aggregates the representations of a user and an item from all propagation layers, and outputs the predicted matching score.
• Input: collaborative knowledge graph G that includes the useritem bipartite graph G1 and knowledge graph G2. • Output: a prediction function that predicts the probability yˆui that user u would adopt item i.
Code https://github.com/xiangwang1223/knowledge_graph_attention_network
Dataset http://jmcauley.ucsd.edu/data/amazon https://grouplens.org/datasets/hetrec-2011/ https://www.yelp.com/dataset/challenge
Pros
This paper was read because of two reasons: -
To understand this paper better: - Guo, S., Lin, Y., Feng, N., Song, C., & Wan, H. (2019, July). Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. In Proceedings of the AAAI conference on artificial intelligence (Vol. 33, No. 01, pp. 922-929).\
Since our problem can and does falls under high-order modeling domain where we are trying to capture more information related to user, this could be a good approach to begin with.