![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F17113715%2Fb2e193d0c094378486c536f23c256286%2F2.png?generation=1741436626853330&alt=media\")
First of all, I'd like to thank the organizers and Kaggle for hosting such an amazing competition. I would also like to express my special gratitude to my teammates @cody11null and stefanoclss.
\nTo improve the accuracy of the target transformation, we first transformed the survival times of each race_group to follow a log-logistic distribution, and then performed final model prediction on an ensemble of multiple GBDT+TabM and nn_pairwise models using the two types of target transformation.
\n
Assuming that efs_time with long survival time (efs=0) and efs_time with short survival time (efs=1) follow different statistical distributions, found that the efs=1 data follow a unique log-logistic distribution by race_group. Then extracted the data with efs=0 that followed this log-logistic distribution, and corrected these data by the conditional expectation of this distribution. Using these data, two target transformations (Kaplan-Meier and QuantileTransform) were performed.
\nWe performed some feature engineering and built the following model.
\nEach single models were divided into several blocks, and a stacking ensemble using the oof of each model was performed for each block in the first layer. A model was constructed that maximized CV and LB by increasing diversity for each block through repeated trials. In the second layer, a linear model was constructed in which the first layer ensemble was weighted to maximize the c-index of CV.Finally,post-processing was performed to balance the scores for each race_group and maximize the c-index.
\nBelow is our final submission result.
\nBest : CV 0.6875 / public LB 0.694 / private LB 0.694
What was the use of scipy -> minimize here @kazuwhite ?
", "votes": 1, "replies": [ { "id": 3146620, "author_name": "KW", "author_url": "", "post_date": "2025-03-11T04:13:14.690000", "content": "One ensemble block used Nelder-Mead to search for ensemble weights and scipy-minimize to minimize the negative c-index score function.
\nOn the other hand, in another ensemble block, weights that directly maximize the c-index were found by simulation, and these were finally combined to build a predictive model.
Got this @kazuwhite
\nGreat work! This seems like a hill climb of sorts
First of all, I'd like to thank the organizers and Kaggle for hosting such an amazing competition. I would also like to express my special gratitude to my teammates @cody11null and stefanoclss.
\nTo improve the accuracy of the target transformation, we first transformed the survival times of each race_group to follow a log-logistic distribution, and then performed final model prediction on an ensemble of multiple GBDT+TabM and nn_pairwise models using the two types of target transformation.
\n
Assuming that efs_time with long survival time (efs=0) and efs_time with short survival time (efs=1) follow different statistical distributions, found that the efs=1 data follow a unique log-logistic distribution by race_group. Then extracted the data with efs=0 that followed this log-logistic distribution, and corrected these data by the conditional expectation of this distribution. Using these data, two target transformations (Kaplan-Meier and QuantileTransform) were performed.
\nWe performed some feature engineering and built the following model.
\nEach single models were divided into several blocks, and a stacking ensemble using the oof of each model was performed for each block in the first layer. A model was constructed that maximized CV and LB by increasing diversity for each block through repeated trials. In the second layer, a linear model was constructed in which the first layer ensemble was weighted to maximize the c-index of CV.Finally,post-processing was performed to balance the scores for each race_group and maximize the c-index.
\nBelow is our final submission result.
\nBest : CV 0.6875 / public LB 0.694 / private LB 0.694
What was the use of scipy -> minimize here @kazuwhite ?
", "rawMarkdown": "What was the use of scipy -> minimize here @kazuwhite ?", "votes": 1, "replies": [ { "id": 3146620, "postDate": "2025-03-11T04:13:14.690Z", "content": "One ensemble block used Nelder-Mead to search for ensemble weights and scipy-minimize to minimize the negative c-index score function.
\nOn the other hand, in another ensemble block, weights that directly maximize the c-index were found by simulation, and these were finally combined to build a predictive model.
Got this @kazuwhite
\nGreat work! This seems like a hill climb of sorts