Update app.R

app.R

@@ -152,6 +152,9 @@ team_leagues <- download_private_csv("HyannisHarborHawksCCBL/CleanDataObjects",
 college_guts <- download_private_csv("HyannisHarborHawksCCBL/CleanDataObjects", "college_GUTS.csv")
 college_join <- download_private_csv("HyannisHarborHawksCCBL/CleanDataObjects", "college_join.csv")
 
+#Reference arsenal data (only pitches thrown over 5% for retagging)
+pitcher_arsenals_over_5perc <- download_private_csv("HyannisHarborHawksCCBL/CleanDataObjects", "pitcher_arsenals_over_5perc.csv")
+
 #Team logo
 logo_b64 <- base64enc::base64encode("HHLogo.png")
 logo_uri <- paste0("data:image/png;base64,", logo_b64)
@@ -160,6 +163,102 @@ logo_uri <- paste0("data:image/png;base64,", logo_b64)
 #Clean data 
 # ============================================================
 
+#Function that retags pitches into a column of a Trackman dataset
+retag_pitches <- function(data) {
+  
+  #Uses create_pitcher_arsenals function to get arsenals from reference set (college season)
+  arsenal_percentages <- pitcher_arsenals_over_5perc %>%
+    mutate(in_reference = TRUE)   
+  
+  # get row ID for joins later
+  data <- data %>% mutate(row_id = row_number())
+  
+  #aggregation that prepares data for prediction (removes NAs, creates features, flips HorzBreak and RelSide)
+  predict_data <- data %>%
+    filter(
+      !Pitcher %in% unique(tagged_sample$Pitcher),
+      !TaggedPitchType %in% c("Other", "Undefined", "FastBall", "Knuckleball"),
+      !is.na(TaggedPitchType), !is.na(RelSpeed),
+      !is.na(InducedVertBreak), !is.na(Extension)
+    ) %>%
+    mutate(
+      max_velo_diff = max(RelSpeed, na.rm = TRUE) - RelSpeed,
+      max_vert_diff = InducedVertBreak[which.max(RelSpeed)] - InducedVertBreak,
+      max_ext_diff  = Extension[which.max(RelSpeed)] - Extension,
+      .by = c(GameUID, Pitcher)
+    ) %>%
+    mutate(
+      RelSide = case_when(
+        PitcherThrows == "Right" ~ RelSide,
+        PitcherThrows == "Left"  ~ -RelSide,
+        PitcherThrows %in% c("Both", "Undefined") & RelSide > 0 ~ RelSide,
+        PitcherThrows %in% c("Both", "Undefined") & RelSide < 0 ~ -RelSide),
+      HorzBreak = case_when(
+        PitcherThrows == "Right" ~ HorzBreak,
+        PitcherThrows == "Left"  ~ -HorzBreak,
+        PitcherThrows %in% c("Both", "Undefined") & HorzBreak > 0 ~ HorzBreak,
+        PitcherThrows %in% c("Both", "Undefined") & HorzBreak < 0 ~ -HorzBreak)
+    )
+  
+  #Predcted probabilities from model
+  probs          <- predict(final_fit, predict_data, type = "prob")
+  pred_col_names <- names(probs)
+  
+  #Get row_id for join
+  probs$row_id   <- predict_data$row_id
+  
+  #Join over probabilities
+  data <- data %>%
+    left_join(arsenal_percentages, by = "Pitcher") %>%
+    left_join(probs, by = "row_id") %>%
+    select(-row_id)
+  
+  #Determine if pitcher has reference data (pitched in college season)
+  has_arsenal <- !is.na(data$in_reference)  
+  
+  # mask only pitchers who are in the reference, everyone else predicts normally, meaning a pitcher not in the reference has no 
+  #arsenal constraint and is just given the model prediction. everyone else moves to the next steps:
+  pitch_to_perc <- c(
+    .pred_Fastball = "perc_FF", .pred_Cutter    = "perc_CT",
+    .pred_Sinker   = "perc_SI", .pred_Slider    = "perc_SL",
+    .pred_Sweeper  = "perc_SW", .pred_Curveball = "perc_CU",
+    .pred_Changeup = "perc_CH", .pred_Splitter  = "perc_SP"
+  )
+  for (pc in intersect(names(pitch_to_perc), names(data))) {
+    ac   <- pitch_to_perc[[pc]]
+    mult <- ifelse(has_arsenal, as.numeric(coalesce(data[[ac]], 0) > 0), 1)
+    data[[pc]] <- data[[pc]] * mult 
+  }
+  
+  #Build a matrix of the (now masked) class probabilities 
+  prob_mat  <- as.matrix(data[pred_col_names])
+  
+  #Flag rows that actually got scored, unscored rows stay all-NA and are skipped
+  predicted <- rowSums(!is.na(prob_mat)) > 0
+  
+  #Start everyone as NA, only scored rows get a tag
+  pt <- rep(NA_character_, nrow(data))
+  
+  #Pull just the scored rows, zero out NAs so max.col can run
+  m  <- prob_mat[predicted, , drop = FALSE]
+  m[is.na(m)] <- 0
+  
+  #Pick the highest surviving probability per row as the predicted pitch
+  pt[predicted] <- pred_col_names[max.col(m, ties.method = "first")]
+  
+  #Strip the ".pred_" prefix to get a clean pitch type, drop the helper column
+  data$TaggedPitchType <- sub("^\\.pred_", "", pt)
+  data %>% select(-in_reference)
+  
+  #make all splitters change ups
+  data <- data %>% 
+    mutate(TaggedPitchType = ifelse(TaggedPitchType == "Splitter", "Changeup", TaggedPitchType))
+    
+  return(data)
+  
+}
+
+
 #Function that predicts stuffplus 
 predict_stuffplus <- function(data) {
 
@@ -282,6 +381,8 @@ clean_college_data <- function(data, teams = NA) {
       TaggedPitchType == "Undefined" ~ "Other", 
       T ~ TaggedPitchType
     ))
+
+  data <- retag_pitches(data)
   
   #Main aggregations
   data <- data %>%