How can machine learning help to predict changes in size of Atlantic herring? Olga Lyashevska ∗ Clementine Harma Deirdre Brophy Coilin Minto Maurice Clarke * Marine and Freshwater Research Centre Galway-Mayo Institute of Technology (GMIT) Galway, Ireland olga.lyashevska@gmit.ie July, 22 2016 1 / 15
Background Avg decline of 4 cm 35 30 Fish length, cm 25 20 15 1960 1970 1980 1990 2000 2010 Years 2 / 15
Problem ◮ Herring are one of the most important pelagic species exploited by fisheries; 3 / 15
Problem ◮ Herring are one of the most important pelagic species exploited by fisheries; ◮ Reductions in growth have consequences for stock productivity; 3 / 15
Problem ◮ Herring are one of the most important pelagic species exploited by fisheries; ◮ Reductions in growth have consequences for stock productivity; ◮ The cause of the decline remains largely unexplained; 3 / 15
Problem ◮ Herring are one of the most important pelagic species exploited by fisheries; ◮ Reductions in growth have consequences for stock productivity; ◮ The cause of the decline remains largely unexplained; ◮ Likely to be driven by the interactive effect of various factors: ◮ sea surface temperature; ◮ zooplankton abundance; ◮ fish abundance; ◮ fishing pressure; 3 / 15
Data ◮ 1959 – 2012; ◮ throughout the year; ◮ random sampling (n = 50 to 100) from commercial vessels; ◮ pelagic trawling; ◮ age and weight-at-length; ◮ total sample size 50,000; 4 / 15
Study Area Celtic Sea 5 / 15
Objective To identify important variables underlying changes in growth using Gradient Boosting Regression Trees (GBRT) residuals residuals 6 / 15
GBRT ◮ Advantages: ◮ Detection of (non-linear) feature interactions; ◮ Resistance to inclusion of irrelevant features; ◮ Heterogeneous data (features measured on different scale); ◮ Robustness to outliers; ◮ Accuracy; ◮ Different loss functions 7 / 15
GBRT ◮ Advantages: ◮ Detection of (non-linear) feature interactions; ◮ Resistance to inclusion of irrelevant features; ◮ Heterogeneous data (features measured on different scale); ◮ Robustness to outliers; ◮ Accuracy; ◮ Different loss functions ◮ Disdvantages: ◮ Requires careful tuning; ◮ Slow to train (but fast to predict); 7 / 15
Formal Specification M � F m ( x ) = γ m h m ( x ) (1) m =1 where γ m is a weight and h m ( x ) are weak learners. 8 / 15
Formal Specification M � F m ( x ) = γ m h m ( x ) (1) m =1 where γ m is a weight and h m ( x ) are weak learners. GBRT builds the additive model in a forward stagewise fashion: F m ( x ) = F m − 1 ( x ) + ǫγ m h m ( x ) (2) where ǫ is a shrinkage. 8 / 15
Formal Specification M � F m ( x ) = γ m h m ( x ) (1) m =1 where γ m is a weight and h m ( x ) are weak learners. GBRT builds the additive model in a forward stagewise fashion: F m ( x ) = F m − 1 ( x ) + ǫγ m h m ( x ) (2) where ǫ is a shrinkage. At each stage the weak learner h m ( x ) is chosen to minimize the loss function L given the current model F m − 1 and its fit F m − 1 ( x i ) n � F m ( x ) = F m − 1 ( x ) + arg min L ( y i , F m − 1 ( x i ) − h ( x )) (3) h i =1 8 / 15
GBRT hyperparameters ◮ number of iterations = 500; ◮ shrinkage (learning rate) = 0.05; ◮ max tree depth = 6; ◮ subsample = 0.75; ◮ loss function = Least Squares; scikit 9 / 15
Model estimation 2.8 2.6 Train 2.4 Test Validation ◮ MSE: 1.31 2.2 ◮ R 2 train: 54.5% MSE ◮ R 2 test: 51.7% 2.0 ◮ R 2 val: 52.6% 1.8 1.6 1.4 1.2 0 100 200 300 400 500 Boosting Iterations Low R 2 due to individual variability 10 / 15
True vs Predicted 34 32 30 Predicted length, cm 28 26 24 22 20 18 18 20 22 24 26 28 30 32 34 True length, cm 11 / 15
Variable Importance Plot nao sal 0.0 0.2 0.4 0.6 0.8 1.0 trend temperature abundance food 12 / 15
Partial Dependence Plots 0.5 0.5 0.5 Partial dependence Partial dependence Partial dependence 0.0 0.0 0.0 − 0.5 − 0.5 − 0.5 − 1.0 − 1.0 − 1.0 2 4 6 8 10 12 13.0 13.5 14.0 14.5 15.0 15 30 45 60 75 xmonth sst chel1 70 70 - 14.5 -1. 3 1 4 . 60 60 ✁ � . 0 1 1 -0.39 - 0 . 8 0 0 . 0 50 50 14.0 ✂ chel1 chel1 40 40 0.43 sst 0.43 0.43 30 30 13.5 0.85 ✷ 0 ✷ 0 0 . 4 3 10 10 13.0 -0.39 - -1. 0.0 0.0 0 0.43 0.0 0 0 . ☎ 8 ✆ 4 6 8 10 1 4 6 8 10 1 1 ✷ .8 13. 13.6 14.0 ✝ 14.4 ✄ 0 1 ✷ ✷ ✷ ✷ ✷ xmonth xmonth sst 13 / 15
Conclusions ◮ trend, sea surface temperature and food availability are three most importans features; 14 / 15
Conclusions ◮ trend, sea surface temperature and food availability are three most importans features; ◮ sea surface temperature above 14 degrees negatively relates to fish length, whereas food availability is invariant; 14 / 15
Conclusions ◮ trend, sea surface temperature and food availability are three most importans features; ◮ sea surface temperature above 14 degrees negatively relates to fish length, whereas food availability is invariant; ◮ there is a high degree of interaction between all features; 14 / 15
Conclusions ◮ trend, sea surface temperature and food availability are three most importans features; ◮ sea surface temperature above 14 degrees negatively relates to fish length, whereas food availability is invariant; ◮ there is a high degree of interaction between all features; ◮ not a cause-effect relationship, but a relative importance of the variables; 14 / 15
lyashevska linked.in/lyashevska s ci kit Acknowledgements: This research was carried out with the support of the Irish Environmental Protection Agency grant (Ecosystem tipping points: learning from the past to manage for the future, project code 2015-NC-MS-3) and the support of the Marine Institute under the Marine Research Sub-programme funded by the Irish Government. 15 / 15
Recommend
More recommend
