Homework1-XAI2022-Pawel-Fijalkowski
+Paweł Fijałkowski
+2022-03-30
+ +Initial:
+As the package was removed from CRAN, we needed to download it +directly from GitHub (using Rstudio Cloud, as it only worked +somehow).
+EPL_2021 <- read.csv("EPL_2021.csv")
+head(EPL_2021)Each row in the data set is a shot attempt. Provided features +(“league”, “shotType”, “home_team”, “away_team”, …) are supposed to help +us in predicting “xG” value which essentially is a probability that this +attempt will end up with a positive result (a goal).
+Here we select features that are important in the context of our +problem and do proper transformations ((X,Y) -> (distance, +angle))
+library(dplyr)
+data <- EPL_2021 %>% select(result, X, Y, xG, h_a, situation, shotType, home_goals, away_goals, lastAction) %>%
+ mutate(status = ifelse(result == "Goal", "1", "0")) %>%
+ mutate(distanceToGoal = sqrt((105 - (X * 105)) ^ 2 + (32.5 - (Y * 68)) ^ 2)) %>%
+ mutate(angleToGoal = abs(atan((7.32 * (105 - (X * 105))) / ((105 - (X * 105))^2 +
+ (32.5 - (Y * 68)) ^ 2 - (7.32 / 2) ^ 2)) * 180 / pi)) %>%
+ select(-X, -Y, -result)Models:
+We will be predicting probability of scoring a goal (status) for each +shot attempt.
+library(DALEX)
+library(DALEXtra)
+library(ranger)data_processed <- data %>% mutate(h_a = as.factor(h_a),
+ situation = as.factor(situation),
+ shotType = as.factor(shotType),
+ lastAction = as.factor(lastAction),
+ status = as.numeric(status))
+
+model <- ranger::ranger(status ~., data = data_processed, classification = TRUE, probability = TRUE)Explanations
+Making prediction:
+obs40 <- data_processed[40, !(colnames(data_processed) %in% c("status"))]
+predict_40 <- predict(model, obs40)
+ifelse(predict_40$predictions[2] <= 0.5, 0, 1)## [1] 0data_processed[12, "status"]## [1] 0Model correctly predicted that attempt #40 will not end up with the +goal.
+Let’s create explainer!
+explainer <- explain(model, data = data_processed, y = as.numeric(data_processed$status))## Preparation of a new explainer is initiated
+## -> model label : ranger ( default )
+## -> data : 7389 rows 10 cols
+## -> target variable : 7389 values
+## -> predict function : yhat.ranger will be used ( default )
+## -> predicted values : No value for predict function target column. ( default )
+## -> model_info : package ranger , ver. 0.13.1 , task classification ( default )
+## -> predicted values : numerical, min = 0 , mean = 0.1066805 , max = 0.9838016
+## -> residual function : difference between y and yhat ( default )
+## -> residuals : numerical, min = -0.7995832 , mean = -0.002200902 , max = 0.6997594
+## A new explainer has been created!Break Down for observation #40
+plot(predict_parts(explainer = explainer, new_observation = obs40))
The angle to the goal contributes heavily negatively to the result +(scoring goal).
+Shap on observation 40
+plot(predict_parts(explainer = explainer, new_observation = obs40, type = 'shap', B = 4))
We clearly see that low quality of the attempt (xG) and angleToGoal +contributed negatively.
+We will find 2 observations in which the same feature have different +effect.
+obs37 <- data_processed[37, !(colnames(data_processed) %in% c("status"))]
+plot(predict_parts(explainer = explainer, new_observation = obs37))
obs120 <- data_processed[120, !(colnames(data_processed) %in% c("status"))]
+plot(predict_parts(explainer = explainer, new_observation = obs120))
As we can see “xG” value can contribute negatively(120) or +positively(37) to the status variable.
+
+
+
+