Student_Placement_Prediction.Rmd

---
output:
  pdf_document: default
  html_document: default
---
```{r}
library(readxl)
library(dplyr)
library(caret)
library(pROC)
library(rpart)
```

```{r}
data <- read_excel("/file/path")
```

# (2)
# best approach is to exclude such parameters from the model
# parametes that are not available at the time of decision making are not present in the model

```{r}
data$Gender <- as.factor(data$Gender)
data$Board_SSC <- as.factor(data$Board_SSC)
data$Board_HSC <- as.factor(data$Board_HSC)
data$Stream_HSC <- as.factor(data$Stream_HSC)
data$Course_Degree <- as.factor(data$Course_Degree)
data$Entrance_Test <- as.factor(data$Entrance_Test)
data$Specialization_MBA <- as.factor(data$Specialization_MBA)

data <- na.omit(data)
```

```{r}
# (3) logistic regression model using only SSC percentage
model_ssc <- glm(Placement_B ~ Percent_SSC, data = data, family = "binomial")
summary(model_ssc)
```

```{r}
# (4) placement probabilities for  SSC percentages
new_data <- data.frame(Percent_SSC = c(60, 80))
predicted_probabilities <- predict(model_ssc, newdata = new_data, type = "response")
predicted_probabilities
```

```{r}
# (5) best cutoff probability = 0.780510400018327
roc_curve <- roc(data$Placement_B, fitted(model_ssc))
best_cutoff <- coords(roc_curve, "best", ret = "threshold")
print(paste("Best Cutoff:", best_cutoff))
```

```{r}
# (6) logistic regression model with all appropriate parameters
model_full <- glm(Placement_B ~ Percent_SSC + Percent_HSC + Percent_Degree + Experience_Yrs + Percentile_ET + Percent_MBA, data = data, family = "binomial")
summary(model_full)
```

```{r}
# (7) sensitivity and specificity at 0.7 cut-off
predicted_class <- ifelse(predict(model_full, type = "response") > 0.7, 1, 0)
conf_matrix <- confusionMatrix(factor(predicted_class), factor(data$Placement_B))
sensitivity_specificity_0_7 <- conf_matrix$byClass[c("Sensitivity", "Specificity")]
```

```{r}
# (8) optimal threshold considering cost
calculate_cost <- function(threshold, data, model) {
    predicted_class <- ifelse(predict(model, type = "response") > threshold, 1, 0)
    table <- table(factor(predicted_class, levels = c(0, 1)), factor(data$Placement_B, levels = c(0, 1)))
    fp_cost <- 4 * table[2, 1]
    fn_cost <- table[1, 2]
    total_cost <- fp_cost + fn_cost
    return(total_cost)
}

thresholds <- seq(0, 1, by = 0.01)

costs <- sapply(thresholds, calculate_cost, data = data, model = model_full)

for (i in 1:length(thresholds)) {
    print(paste("Threshold:", thresholds[i], "Cost:", costs[i]))
}

optimal_threshold <- thresholds[which.min(costs)]
print(paste("Optimal Threshold:", optimal_threshold))
# Optimal Threshold: 0.76
```

```{r}
# (9) decision tree model
tree_model <- rpart(Placement_B ~ ., data = data, method = "class")
tree_pred <- predict(tree_model, newdata = data, type = "class")
tree_conf_matrix <- confusionMatrix(tree_pred, factor(data$Placement_B))
```

```{r}
# model comparison
log_conf_matrix <- conf_matrix
comparison <- data.frame(
    Model = c("Logistic Regression", "Decision Tree"),
    Accuracy = c(log_conf_matrix$overall["Accuracy"], tree_conf_matrix$overall["Accuracy"]),
    Sensitivity = c(log_conf_matrix$byClass["Sensitivity"], tree_conf_matrix$byClass["Sensitivity"]),
    Specificity = c(log_conf_matrix$byClass["Specificity"], tree_conf_matrix$byClass["Specificity"])
)
comparison
```

```{r}
print(paste("Best cutoff (ROC):", best_cutoff))
print(paste("Optimal threshold (Cost):", optimal_threshold))
print("Model Comparison:")
print(comparison)
```


# (10) recommendation

# logistic reg had good accuracy (0.7654321) and good specificity (0.8965517) , but low sensitivity (0.2222222)
# model needs improvement with sensitivity
# to improve sensitivity can:
# use more relevant predictors, adjust model parameters, or use different modeling techniques


# decsion tree had perfect (1.0) accuracy, sensitivity, and sepecificity
# however, these perfect results indicate overfitting to the training data (may not be good for new unseen data)
#
# model choice = logistic regression
# but with logistic reg, be sure to include
# regular model updates, sensitivity improvements, and consideration of non-quantifiable factors