Thank you for organizing this very interesting competition. And thank you to all participants.
\nThe 22nd place solution is a simple solution using only “title” and “cpc”.
Step 1 may not reach 60 because the title is too short. In that case, add Step 2 to the candidates.
\nThe sum of candidates in Step 1 and Step 2 may not reach 30 because some publication_numbers have very little information on title and cpc. In that case, add Step 3 to the candidates.
\nUsing these candidates, run the Annealing algorithm. However, I am unable to understand the annealing algorithm and use it directly from the following two excellent public notebooks.
\n@tubotubo @andrey67 I really appreciate it.
https://www.kaggle.com/code/tubotubo/uspto-simulated-annealing-baseline
\nhttps://www.kaggle.com/code/andrey67/uspto-annealing-lb-0-31
Of course, I could not find MAGIC at all. It is very amazing who found it!
", "messages": [ { "id": "2935831", "postDate": "07/25/2024 15:18:55", "content": "Thank you for organizing this very interesting competition. And thank you to all participants.
\nThe 22nd place solution is a simple solution using only “title” and “cpc”.
Step 1 may not reach 60 because the title is too short. In that case, add Step 2 to the candidates.
\nThe sum of candidates in Step 1 and Step 2 may not reach 30 because some publication_numbers have very little information on title and cpc. In that case, add Step 3 to the candidates.
\nUsing these candidates, run the Annealing algorithm. However, I am unable to understand the annealing algorithm and use it directly from the following two excellent public notebooks.
\n@tubotubo @andrey67 I really appreciate it.
https://www.kaggle.com/code/tubotubo/uspto-simulated-annealing-baseline
\nhttps://www.kaggle.com/code/andrey67/uspto-annealing-lb-0-31
Of course, I could not find MAGIC at all. It is very amazing who found it!
", "rawMarkdown": "Thank you for organizing this very interesting competition. And thank you to all participants.\nThe 22nd place solution is a simple solution using only “title” and “cpc”.\n\n# Creating candidates\n\n### Step 1: 90 candidates by TF-IDF (title)\n- stopwords='english'\n- ngram=(1, 3)\n- max_df=100\n- min_df=3\n- If a word is in another word set, exclude it\n- Exclude words with less than 5 letters\n\nStep 1 may not reach 60 because the title is too short. In that case, add Step 2 to the candidates.\n\n### Step 2: 60 candidates by TF-IDF (cpc)\n- max_df=0.01\n\nThe sum of candidates in Step 1 and Step 2 may not reach 30 because some publication_numbers have very little information on title and cpc. In that case, add Step 3 to the candidates.\n\n### Step 3: 10 candidates by word occurrence of title\n- Split title into words and sorted by number of occurrences (max 10)\n\n# Annealing algorithm\n\nUsing these candidates, run the Annealing algorithm. However, I am unable to understand the annealing algorithm and use it directly from the following two excellent public notebooks.\n@tubotubo @andrey67 I really appreciate it.\n\nhttps://www.kaggle.com/code/tubotubo/uspto-simulated-annealing-baseline\nhttps://www.kaggle.com/code/andrey67/uspto-annealing-lb-0-31\n\nOf course, I could not find MAGIC at all. It is very amazing who found it!", "votes": null }, { "id": "2936358", "postDate": "07/26/2024 04:57:37", "content": "Thank you for sharing your 22nd place solution! It's impressive how you utilized a simple yet effective TF-IDF approach with titles and CPC, and your step-by-step candidate creation process is very insightful. Great job! @thajime
", "rawMarkdown": "Thank you for sharing your 22nd place solution! It's impressive how you utilized a simple yet effective TF-IDF approach with titles and CPC, and your step-by-step candidate creation process is very insightful. Great job! @thajime", "votes": null } ], "comments": [ { "id": 2936358, "author_name": "aaditshukla", "author_url": "", "post_date": "07/26/2024 04:57:37", "content": "Thank you for sharing your 22nd place solution! It's impressive how you utilized a simple yet effective TF-IDF approach with titles and CPC, and your step-by-step candidate creation process is very insightful. Great job! @thajime
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2935831": "Thank you for organizing this very interesting competition. And thank you to all participants.\nThe 22nd place solution is a simple solution using only “title” and “cpc”.\n\n# Creating candidates\n\n### Step 1: 90 candidates by TF-IDF (title)\n- stopwords='english'\n- ngram=(1, 3)\n- max_df=100\n- min_df=3\n- If a word is in another word set, exclude it\n- Exclude words with less than 5 letters\n\nStep 1 may not reach 60 because the title is too short. In that case, add Step 2 to the candidates.\n\n### Step 2: 60 candidates by TF-IDF (cpc)\n- max_df=0.01\n\nThe sum of candidates in Step 1 and Step 2 may not reach 30 because some publication_numbers have very little information on title and cpc. In that case, add Step 3 to the candidates.\n\n### Step 3: 10 candidates by word occurrence of title\n- Split title into words and sorted by number of occurrences (max 10)\n\n# Annealing algorithm\n\nUsing these candidates, run the Annealing algorithm. However, I am unable to understand the annealing algorithm and use it directly from the following two excellent public notebooks.\n@tubotubo @andrey67 I really appreciate it.\n\nhttps://www.kaggle.com/code/tubotubo/uspto-simulated-annealing-baseline\nhttps://www.kaggle.com/code/andrey67/uspto-annealing-lb-0-31\n\nOf course, I could not find MAGIC at all. It is very amazing who found it!", "2936358": "Thank you for sharing your 22nd place solution! It's impressive how you utilized a simple yet effective TF-IDF approach with titles and CPC, and your step-by-step candidate creation process is very insightful. Great job! @thajime" }, "source": "meta" }