chenchongthu
Confusing with Table 3. in your paper
Hi, thank you very much for your impressive work. However, I am confused with a little things in your paper and your code, and I carefully read a lot of times but can not find answers.
The first question is about Table 3. in your paper. How do you conduct the experiment? Specifically, I do not find any descriptions of "delete settings" in Table 3 experiments. For example, how much samples you delete in Table 3 experiment, and how do you choose these "delete samples". And what is the delete method? (if you choose 10 samples to delete, you delete them sequentially or just delete all of them at one time? ). Or, you just run recommendation experiments for these method with different unlearning strategies, but you do not consider unlearning in Table 2. experiment.
An other question is , could you please give a brief yet very helpful description of your code?
Thank you very much for your contributions!
Hello, I have the same problem here. However, after thinking about it, I believe that "deletion" is a hypothetical situation, and the partitioning and aggregation strategies proposed in the paper are the main contribution points. Therefore, Table 3 mainly compares the final performance of different methods after dividing data and training different sub models. Retrain can be seen as training a single model with complete data.
My opinion may not be correct, and the author may have omitted an introduction to deletion operations. We can email him for inquiries.
Hello, I have the same problem here. However, after thinking about it, I believe that "deletion" is a hypothetical situation, and the partitioning and aggregation strategies proposed in the paper are the main contribution points. Therefore, Table 3 mainly compares the final performance of different methods after dividing data and training different sub models. Retrain can be seen as training a single model with complete data.
My opinion may not be correct, and the author may have omitted an introduction to deletion operations. We can email him for inquiries.
My understanding is consistent with yours. However, in my opinion, when interactions that need to be forgotten come from multiple shards, this architecture may not be as effective as the author claims.
