Dear fellow Kagglers, we see how persistently you are storming the leaderboard and how reluctantly and slowly the metric is growing. We have made a decision that will allow us to encourage your efforts.
\nIf the target threshold of HitRate@3 ≥ 0.7 is not reached on the private leaderboard, we will open two additional prize categories funded by the competition bonus pool
\nTop-5 Single Model Award
\nFor the best solutions using one model (see definition below).
Top-5 Insights Award
\nFor the best solutions from teams that did not occupy places in the TOP-3.
A total of 10 bonus prize places are provided - 5 prizes of $500 each in each nomination.
\nWhat is considered a Single Model?
\nSingle Model - is a solution in which the final prediction is generated by one trained estimator with a single set of parameters.
\nSingle Model in the context of the competition includes:
\nSingle Model does not include:
\nVerification rule:
\nIf more than one .fit() call is used for different models in the final pipeline or predictions from several models are combined — this is not considered a Single Model.
Requirements for providing code and solution descriptions are regulated by standard competition conditions (see section 6. Winner's Obligations. All winning solutions must be properly delivered and comply with competition rules
", "messages": [ { "id": 3262729, "postDate": "2025-08-04T07:39:37.670Z", "content": "Dear fellow Kagglers, we see how persistently you are storming the leaderboard and how reluctantly and slowly the metric is growing. We have made a decision that will allow us to encourage your efforts.
\nIf the target threshold of HitRate@3 ≥ 0.7 is not reached on the private leaderboard, we will open two additional prize categories funded by the competition bonus pool
\nTop-5 Single Model Award
\nFor the best solutions using one model (see definition below).
Top-5 Insights Award
\nFor the best solutions from teams that did not occupy places in the TOP-3.
A total of 10 bonus prize places are provided - 5 prizes of $500 each in each nomination.
\nWhat is considered a Single Model?
\nSingle Model - is a solution in which the final prediction is generated by one trained estimator with a single set of parameters.
\nSingle Model in the context of the competition includes:
\nSingle Model does not include:
\nVerification rule:
\nIf more than one .fit() call is used for different models in the final pipeline or predictions from several models are combined — this is not considered a Single Model.
Requirements for providing code and solution descriptions are regulated by standard competition conditions (see section 6. Winner's Obligations. All winning solutions must be properly delivered and comply with competition rules
", "rawMarkdown": "Dear fellow Kagglers, we see how persistently you are storming the leaderboard and how reluctantly and slowly the metric is growing. We have made a decision that will allow us to encourage your efforts. \n\nIf the target threshold of HitRate@3 ≥ 0.7 is not reached on the private leaderboard, we will open two additional prize categories funded by the competition bonus pool\n\n- **Top-5 Single Model Award**\nFor the best solutions using one model (see definition below).\n\n- **Top-5 Insights Award**\nFor the best solutions from teams that did not occupy places in the TOP-3.\n\nA total of 10 bonus prize places are provided - **5 prizes of $500 each** in each nomination.\n\n**What is considered a Single Model?**\n\nSingle Model - is a solution in which the final prediction is generated by one trained estimator with a single set of parameters.\n\nSingle Model in the context of the competition includes:\n\n- Gradient boosting: XGBoost, LightGBM, CatBoost\n- Neural networks (including multi-head architectures)\n- Classical ML algorithms\n- Any model built through one .fit() call and predicting through one .predict()\n\nSingle Model does not include:\n\n- Model ensembles (voting, averaging, weighted combinations)\n- Stacking / Blending with meta-model\n- Using several separately trained models\n- Combining predictions from different estimators\n\n**Verification rule:**\nIf more than one .fit() call is used for different models in the final pipeline or predictions from several models are combined — this is not considered a Single Model. \n\nRequirements for providing code and solution descriptions are regulated by standard competition conditions (see section [6. Winner's Obligations](https://www.kaggle.com/competitions/aeroclub-recsys-2025/rules#6.-winners-obligations). All winning solutions must be properly delivered and comply with competition rules", "votes": 11 }, { "id": 3265330, "postDate": "2025-08-07T10:43:22.627Z", "content": "Thanks for making prize money COUNT! 😄
", "rawMarkdown": "Thanks for making prize money COUNT! 😄 ", "votes": 1 }, { "id": 3265235, "postDate": "2025-08-07T07:44:43.807Z", "content": "I would like to clarify whether a submission that includes a post-processing step similar to the one used in this notebook would still be considered as a \"Single Model\" under the competition rules.
\nSpecifically, the approach uses a single trained XGBoost ranking model (with only one .fit() call), and then applies a rule-based re-ranking on the model's output scores to slightly penalize duplicated or similar flight options. This re-ranking does not involve training or combining multiple models—it's purely a transformation of the prediction scores before generating the final submission.
\nCould you please confirm if such a method is still eligible for the \"Top-5 Single Model Award\"?
\nThank you very much for your time and clarification.
Thanks for your question @bestwater. Yes, solutions with post-processing could be applied to the \"Single Model.\" I think that's quite fair. The main goal of that nomination is to exclude the overcomplicated multi-model approaches that we sometimes build on Kaggle when squeezing LB scores, but never implement in production.
\nEasygoing, elegant, and simple solutions are more than welcome. We're solving a production-related problem, not an academia one.
What about k-fold training for one type of model?
", "rawMarkdown": "What about k-fold training for one type of model?", "replies": [ { "id": 3270775, "postDate": "2025-08-17T10:06:01.397Z", "content": "It's fine. Considered.
", "rawMarkdown": "It's fine. Considered. ", "replies": [ { "id": 3270881, "postDate": "2025-08-17T14:45:02.560Z", "content": "Then what if the notebook submitted infers directly with models trained and saved with k-fold strategy, without calling \"train()\" or \"fit()\"? Thx.
", "rawMarkdown": "Then what if the notebook submitted infers directly with models trained and saved with k-fold strategy, without calling \"train()\" or \"fit()\"? Thx." }, { "id": 3270906, "postDate": "2025-08-17T15:35:28.247Z", "content": "@godgod3 first of all we need high private lb score, SM marked selected submission from the competitor and shared notebook. If the shared notebook holds just an inference part, or for example can't be executed in Kaggle environment it is fine for a while. Based on the scores we'll contact real pretenders and ask them to share full solution according to the Competion Rules - Winner Obligations.
\nMay I ask you, are you talking about one of your submissions? Feel free to contact me directly to clarify the details.
", "rawMarkdown": "@godgod3 first of all we need high private lb score, _SM_ marked selected submission from the competitor and shared notebook. If the shared notebook holds just an inference part, or for example can't be executed in Kaggle environment it is fine for a while. Based on the scores we'll contact real pretenders and ask them to share full solution according to the Competion Rules - Winner Obligations.\n\nMay I ask you, are you talking about one of your submissions? Feel free to contact me directly to clarify the details. " }, { "id": 3271086, "postDate": "2025-08-18T03:40:47.420Z", "content": "So top participants will be contacted and they do have to make notebooks public by themselves？
", "rawMarkdown": "So top participants will be contacted and they do have to make notebooks public by themselves？" } ] } ] }, { "id": 3270020, "postDate": "2025-08-15T14:36:20.940Z", "rawMarkdown": "", "votes": 1, "isDeleted": true, "replies": [ { "id": 3270119, "postDate": "2025-08-15T18:30:31.050Z", "content": "Same as others: Private LB Metric + Winner obligations
", "rawMarkdown": "Same as others: Private LB Metric + [Winner obligations](https://www.kaggle.com/competitions/aeroclub-recsys-2025/rules#6.-winners-obligations) " } ] } ], "comments": [ { "id": 3265330, "author_name": "ducnh279", "author_url": "", "post_date": "2025-08-07T10:43:22.627000", "content": "Thanks for making prize money COUNT! 😄
", "votes": 1, "replies": [] }, { "id": 3265235, "author_name": "bestwater", "author_url": "", "post_date": "2025-08-07T07:44:43.807000", "content": "I would like to clarify whether a submission that includes a post-processing step similar to the one used in this notebook would still be considered as a \"Single Model\" under the competition rules.
\nSpecifically, the approach uses a single trained XGBoost ranking model (with only one .fit() call), and then applies a rule-based re-ranking on the model's output scores to slightly penalize duplicated or similar flight options. This re-ranking does not involve training or combining multiple models—it's purely a transformation of the prediction scores before generating the final submission.
\nCould you please confirm if such a method is still eligible for the \"Top-5 Single Model Award\"?
\nThank you very much for your time and clarification.
Thanks for your question @bestwater. Yes, solutions with post-processing could be applied to the \"Single Model.\" I think that's quite fair. The main goal of that nomination is to exclude the overcomplicated multi-model approaches that we sometimes build on Kaggle when squeezing LB scores, but never implement in production.
\nEasygoing, elegant, and simple solutions are more than welcome. We're solving a production-related problem, not an academia one.
What about k-fold training for one type of model?
", "votes": 0, "replies": [ { "id": 3270775, "author_name": "Samvel Kocharyan", "author_url": "", "post_date": "2025-08-17T10:06:01.397000", "content": "It's fine. Considered.
", "votes": 0, "replies": [ { "id": 3270881, "author_name": "GodGod3", "author_url": "", "post_date": "2025-08-17T14:45:02.560000", "content": "Then what if the notebook submitted infers directly with models trained and saved with k-fold strategy, without calling \"train()\" or \"fit()\"? Thx.
", "votes": 0, "replies": [] }, { "id": 3270906, "author_name": "Samvel Kocharyan", "author_url": "", "post_date": "2025-08-17T15:35:28.247000", "content": "@godgod3 first of all we need high private lb score, SM marked selected submission from the competitor and shared notebook. If the shared notebook holds just an inference part, or for example can't be executed in Kaggle environment it is fine for a while. Based on the scores we'll contact real pretenders and ask them to share full solution according to the Competion Rules - Winner Obligations.
\nMay I ask you, are you talking about one of your submissions? Feel free to contact me directly to clarify the details.
", "votes": 0, "replies": [] }, { "id": 3271086, "author_name": "GodGod3", "author_url": "", "post_date": "2025-08-18T03:40:47.420000", "content": "So top participants will be contacted and they do have to make notebooks public by themselves？
", "votes": 0, "replies": [] } ] } ] }, { "id": 3270020, "author_name": "", "author_url": "", "post_date": "2025-08-15T14:36:20.940000", "content": "", "votes": 1, "replies": [ { "id": 3270119, "author_name": "Samvel Kocharyan", "author_url": "", "post_date": "2025-08-15T18:30:31.050000", "content": "Same as others: Private LB Metric + Winner obligations
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "3262729": "Dear fellow Kagglers, we see how persistently you are storming the leaderboard and how reluctantly and slowly the metric is growing. We have made a decision that will allow us to encourage your efforts. \n\nIf the target threshold of HitRate@3 ≥ 0.7 is not reached on the private leaderboard, we will open two additional prize categories funded by the competition bonus pool\n\n- **Top-5 Single Model Award**\nFor the best solutions using one model (see definition below).\n\n- **Top-5 Insights Award**\nFor the best solutions from teams that did not occupy places in the TOP-3.\n\nA total of 10 bonus prize places are provided - **5 prizes of $500 each** in each nomination.\n\n**What is considered a Single Model?**\n\nSingle Model - is a solution in which the final prediction is generated by one trained estimator with a single set of parameters.\n\nSingle Model in the context of the competition includes:\n\n- Gradient boosting: XGBoost, LightGBM, CatBoost\n- Neural networks (including multi-head architectures)\n- Classical ML algorithms\n- Any model built through one .fit() call and predicting through one .predict()\n\nSingle Model does not include:\n\n- Model ensembles (voting, averaging, weighted combinations)\n- Stacking / Blending with meta-model\n- Using several separately trained models\n- Combining predictions from different estimators\n\n**Verification rule:**\nIf more than one .fit() call is used for different models in the final pipeline or predictions from several models are combined — this is not considered a Single Model. \n\nRequirements for providing code and solution descriptions are regulated by standard competition conditions (see section [6. Winner's Obligations](https://www.kaggle.com/competitions/aeroclub-recsys-2025/rules#6.-winners-obligations). All winning solutions must be properly delivered and comply with competition rules", "3265330": "Thanks for making prize money COUNT! 😄 ", "3265235": "\nI would like to clarify whether a submission that includes a post-processing step similar to the one used in [this notebook](https://www.kaggle.com/code/mango789/xgboost-ranker-rule-based-rerank/notebook) would still be considered as a \"Single Model\" under the competition rules.\n\nSpecifically, the approach uses a single trained XGBoost ranking model (with only one `.fit()` call), and then applies a rule-based re-ranking on the model's output scores to slightly penalize duplicated or similar flight options. This re-ranking does **not** involve training or combining multiple models—it's purely a transformation of the prediction scores before generating the final submission.\n\nCould you please confirm if such a method is still eligible for the \"Top-5 Single Model Award\"?\n\nThank you very much for your time and clarification.\n\n", "3270638": "What about k-fold training for one type of model?", "3270020": "" } }