Implementation of KGAT: - Knowledge Graph Attention Network for Recommendation

[karan96]

Summary shared in issue (#148 ).

Code: - https://github.com/xiangwang1223/knowledge_graph_attention_network

Beginning to implement this code side by side to understand how KGAT works. Not sure if we can convert our problem to be implemented in a knowledge graph but since I read the paper I thought it should be good to gain more understanding on this as according to me we can define our problem in knowledge graph. (Requires Discussion, might be wrong).

[karan96]

@hosseinfani Kindly share your views on this and the summary I shared related to the above issue in (#148)

[hosseinfani]

@karan96 this is a good paper from a good venue. I agree that if we consider the side information as the location of users, we can employ this method for our project. Please keep me update on how you model locations and teams and how to feed them into this method.

[karan96]

@hosseinfani Initial run was successful. It went on towards training the model. I stopped the execution after seeing that it has successfully started with its first epoch because it was taking too long. I believe a complete initial run is not required right now because we know that the model is working. I will try to figure out a way to incorporate our use case in this. Furthermore, what learning should I know try to extract from this initial run?

[hosseinfani]

@karan96 thank you. you have to find where the input data is parsed or what file structure is needed. Also, what the predictions mean and what is the intuition behind the predictions?

[karan96]

@hosseinfani I followed the code and the code does not provide a way where they are preprocessing the dataset. They have given the sources of those databases but the files included in their repo which they are using for their approach are already formatted. They are just picking up kg data and cf data for training and testing and then using it for their model. What should I do in this case?

[karan96]

@hosseinfani As an update, I have fixed the code and now the data is exactly what the data of this paper looks like but on running the code on the toy dataset I faced some bugs that I am currently trying to resolve.

[karan96]

@hosseinfani I was able to create the data file as we discussed and when I executed the code, towards the end when the evaluation part starts it requires a knowledge graph of the format User -> Relation -> Item, in our case Member -> BelongsTo/From -> Location. The author provides no information on how this was created. I tried looking up online for solutions involving creation of Knowledge Graphs and every implementation I could find uses text to create that. How should I proceed further since it seems KGs are an important aspect of this implementation?

[hosseinfani]

@karan96 thanks for the update. KG is just a graph! As you said, you have to simply create a graph whose nodes are members and locations and the edges represented the Relation (belongsto/from).

[karan96]

@hosseinfani Hi Dr. Fani, I was able to create the userList, itemlist, and other necessary files, but the way code is working is a the step where it is calculating metrics such as precision, recall and ndcg, it creates a dict named user_ids from test file and then matches them with the data from train file. Hence when the user_ids from test file has a user id 22 which is not present in the train file, it returns a Key Error 22 which we saw earlier. Might have to discuss it further.

[karan96]

I have changed the way we have formatted the dataset earlier. The new dataset now take countries as head, users as tails and relations between them to make a KG. Currently have to run them along KGAT. Will update the results soon.