Firstly, it's nice to be part of KDD-BR again!. I started looking at the competition 2 days ago, and prototyped a quick solution. I was planning to implement a greedy or beam search later to enforce a tour, but maybe that won't be necessary anymore: it turns out that, against my expectations, the evaluation accepts any binary adjacency matrix, whether it is a tour or not. Maybe this is intended, as otherwise you'd probably want to use the optimality gap as the metric (I think?).
\nNonetheless, I am still confused about this. Am I the only one? If this adjacency matrix must follow more constraints, I reckon they should be checked for at evaluation time.
", "messages": [ { "id": "1523048", "postDate": "09/24/2021 22:08:13", "content": "Firstly, it's nice to be part of KDD-BR again!. I started looking at the competition 2 days ago, and prototyped a quick solution. I was planning to implement a greedy or beam search later to enforce a tour, but maybe that won't be necessary anymore: it turns out that, against my expectations, the evaluation accepts any binary adjacency matrix, whether it is a tour or not. Maybe this is intended, as otherwise you'd probably want to use the optimality gap as the metric (I think?).
\nNonetheless, I am still confused about this. Am I the only one? If this adjacency matrix must follow more constraints, I reckon they should be checked for at evaluation time.
", "rawMarkdown": "Firstly, it's nice to be part of KDD-BR again!. I started looking at the competition 2 days ago, and prototyped a quick solution. I was planning to implement a greedy or beam search later to enforce a tour, but maybe that won't be necessary anymore: it turns out that, against my expectations, *the evaluation accepts any binary adjacency matrix, whether it is a tour or not*. Maybe this is intended, as otherwise you'd probably want to use the optimality gap as the metric (I think?).\n\nNonetheless, I am still confused about this. Am I the only one? If this adjacency matrix must follow more constraints, I reckon they should be checked for at evaluation time.", "votes": null }, { "id": "1525445", "postDate": "09/27/2021 12:45:19", "content": "Hi Bruno,
\nThis is a Kaggle limitation.
\nEffectively, we are measuring how well you can predict whether an edge is present or not in the optimal solution.
Unfortunately, there are some side effects of such a limitation:
\nI see, too bad there's no built-in way to enforce this. I'll keep on predicting these raw adjacency matrices that are not necessarily tours, if that's alright by you guys.
", "rawMarkdown": "I see, too bad there's no built-in way to enforce this. I'll keep on predicting these raw adjacency matrices that are not necessarily tours, if that's alright by you guys.", "votes": null } ], "comments": [ { "id": 1525445, "author_name": "profverri", "author_url": "", "post_date": "09/27/2021 12:45:19", "content": "Hi Bruno,
\nThis is a Kaggle limitation.
\nEffectively, we are measuring how well you can predict whether an edge is present or not in the optimal solution.
Unfortunately, there are some side effects of such a limitation:
\nI see, too bad there's no built-in way to enforce this. I'll keep on predicting these raw adjacency matrices that are not necessarily tours, if that's alright by you guys.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1523048": "Firstly, it's nice to be part of KDD-BR again!. I started looking at the competition 2 days ago, and prototyped a quick solution. I was planning to implement a greedy or beam search later to enforce a tour, but maybe that won't be necessary anymore: it turns out that, against my expectations, *the evaluation accepts any binary adjacency matrix, whether it is a tour or not*. Maybe this is intended, as otherwise you'd probably want to use the optimality gap as the metric (I think?).\n\nNonetheless, I am still confused about this. Am I the only one? If this adjacency matrix must follow more constraints, I reckon they should be checked for at evaluation time.", "1525445": "Hi Bruno,\n\nThis is a Kaggle limitation.\nEffectively, we are measuring how well you can **predict whether an edge is present or not** in the optimal solution.\n\nUnfortunately, there are some side effects of such a limitation:\n\n- We can not check if the solution is valid (a tour) or not;\n- Given two solutions A and B, it is possible that solution A has a higher F1-Score than solution B in the predictive task while yielding longer tours.", "1526041": "I see, too bad there's no built-in way to enforce this. I'll keep on predicting these raw adjacency matrices that are not necessarily tours, if that's alright by you guys." }, "source": "meta" }