{ "id": 557357, "title": "Solution Overview", "content": "

Solution Overview

\n

Hey everyone! We wanted to share how we approached this customer experience prediction challenge. Let us walk you through what we did:

\n

Our Approach

\n

For this challenge, we selected XGBoost as our primary model. XGBoost's ability to handle both numerical and categorical data, coupled with its robust performance, made it a natural choice for the structured dataset we worked with.

\n

Data Preprocessing

\n

We dedicated significant effort to understanding and preparing the data. Below are the key steps we followed:

\n

1. Handling Missing Values

\n\n
\n col  numeric_cols:\n    data[col] = data[col].fillna(data[col].median())\n\n\n col  categorical_cols:\n    data[col] = data[col].fillna(data[col].mode()[])\n
\n

2. Date Features

\n

We engineered date-related features, breaking down date columns into components like year, month, day, and weekend indicators. Additionally, we calculated meaningful time differences:

\n
data[] = data[col].dt.year.fillna(-).astype()\ndata[] = data[col].dt.month.fillna(-).astype()\ndata[] = (data[col].dt.dayofweek >= ).astype()\n
\n

3. Feature Engineering

\n

To enhance the dataset, we created new features:

\n\n

4. Encoding Categorical Variables

\n

We converted text-based categories into numeric representations using label encoding and one-hot encoding.

\n

Modeling

\n

1. Model Selection

\n

We utilized XGBoost, a gradient boosting algorithm, for its suitability to structured datasets.

\n

2. Hyperparameter Tuning

\n

We optimized the following hyperparameters through trial and error:

\n
params = {\n    : ,\n    : ,\n    : ,\n    : ,\n    : ,\n    : ,\n    : \n}\n
\n

3. Cross-Validation

\n

To validate the model's performance, we implemented 5-fold stratified cross-validation:

\n
skf = StratifiedKFold(n_splits=, shuffle=, random_state=)\n train_idx, val_idx  skf.split(X_train, y_train):\n    model = xgb.XGBClassifier(**params)\n    model.fit(X_train[train_idx], y_train[train_idx],\n              eval_set=[(X_train[val_idx], y_train[val_idx])],\n              early_stopping_rounds=, verbose=)\n
\n

4. PCA Features

\n

We applied Principal Component Analysis (PCA) to capture latent patterns in the data:

\n
pca = PCA(n_components=)\npca_features = pca.fit_transform(X_numeric_transformed)\n i  (pca_features.shape[]):\n    X[] = pca_features[:, i]\n
\n

5. Early Stopping

\n

To prevent overfitting, we used early stopping based on validation loss during training.

\n

How We Evaluated It

\n

Our primary evaluation metric was the weighted F1 score to account for imbalanced class distributions. We also:

\n\n
weighted_f1 = f1_score(y_val, y_pred, average=)\n
\n

Biggest Challenges

\n
    \n
  1. Date Handling: Engineering meaningful features from date columns required careful thought.
    2. \n
  2. Categorical Variables: Balancing effective encoding with model performance.
    3. \n
  3. Feature Selection: Ensuring that all new features contributed positively to predictions.
    4. \n
\n

What We Learned

\n
    \n
  1. Feature Engineering: Time spent on creating meaningful features, especially around dates and loyalty programs, greatly improved model performance.
    2. \n
  2. Importance of Pricing Features: Discount-related features were highly predictive.
    3. \n
  3. Cross-Validation: Ensured that our model was generalizable.
    4. \n
  4. PCA Features: Helped uncover patterns missed by manual feature creation.
    5. \n
", "comments": [], "comment_count": 0, "messages": [ { "id": 3099970, "postDate": "2025-01-18T15:32:07.020Z", "content": "

Solution Overview

\n

Hey everyone! We wanted to share how we approached this customer experience prediction challenge. Let us walk you through what we did:

\n

Our Approach

\n

For this challenge, we selected XGBoost as our primary model. XGBoost's ability to handle both numerical and categorical data, coupled with its robust performance, made it a natural choice for the structured dataset we worked with.

\n

Data Preprocessing

\n

We dedicated significant effort to understanding and preparing the data. Below are the key steps we followed:

\n

1. Handling Missing Values

\n\n
\n col  numeric_cols:\n    data[col] = data[col].fillna(data[col].median())\n\n\n col  categorical_cols:\n    data[col] = data[col].fillna(data[col].mode()[])\n
\n

2. Date Features

\n

We engineered date-related features, breaking down date columns into components like year, month, day, and weekend indicators. Additionally, we calculated meaningful time differences:

\n
data[] = data[col].dt.year.fillna(-).astype()\ndata[] = data[col].dt.month.fillna(-).astype()\ndata[] = (data[col].dt.dayofweek >= ).astype()\n
\n

3. Feature Engineering

\n

To enhance the dataset, we created new features:

\n\n

4. Encoding Categorical Variables

\n

We converted text-based categories into numeric representations using label encoding and one-hot encoding.

\n

Modeling

\n

1. Model Selection

\n

We utilized XGBoost, a gradient boosting algorithm, for its suitability to structured datasets.

\n

2. Hyperparameter Tuning

\n

We optimized the following hyperparameters through trial and error:

\n
params = {\n    : ,\n    : ,\n    : ,\n    : ,\n    : ,\n    : ,\n    : \n}\n
\n

3. Cross-Validation

\n

To validate the model's performance, we implemented 5-fold stratified cross-validation:

\n
skf = StratifiedKFold(n_splits=, shuffle=, random_state=)\n train_idx, val_idx  skf.split(X_train, y_train):\n    model = xgb.XGBClassifier(**params)\n    model.fit(X_train[train_idx], y_train[train_idx],\n              eval_set=[(X_train[val_idx], y_train[val_idx])],\n              early_stopping_rounds=, verbose=)\n
\n

4. PCA Features

\n

We applied Principal Component Analysis (PCA) to capture latent patterns in the data:

\n
pca = PCA(n_components=)\npca_features = pca.fit_transform(X_numeric_transformed)\n i  (pca_features.shape[]):\n    X[] = pca_features[:, i]\n
\n

5. Early Stopping

\n

To prevent overfitting, we used early stopping based on validation loss during training.

\n

How We Evaluated It

\n

Our primary evaluation metric was the weighted F1 score to account for imbalanced class distributions. We also:

\n\n
weighted_f1 = f1_score(y_val, y_pred, average=)\n
\n

Biggest Challenges

\n
    \n
  1. Date Handling: Engineering meaningful features from date columns required careful thought.
    2. \n
  2. Categorical Variables: Balancing effective encoding with model performance.
    3. \n
  3. Feature Selection: Ensuring that all new features contributed positively to predictions.
    4. \n
\n

What We Learned

\n
    \n
  1. Feature Engineering: Time spent on creating meaningful features, especially around dates and loyalty programs, greatly improved model performance.
    2. \n
  2. Importance of Pricing Features: Discount-related features were highly predictive.
    3. \n
  3. Cross-Validation: Ensured that our model was generalizable.
    4. \n
  4. PCA Features: Helped uncover patterns missed by manual feature creation.
    5. \n
", "rawMarkdown": "# Solution Overview\n\nHey everyone! We wanted to share how we approached this customer experience prediction challenge. Let us walk you through what we did:\n\n\n## Our Approach\n\nFor this challenge, we selected **XGBoost** as our primary model. XGBoost's ability to handle both numerical and categorical data, coupled with its robust performance, made it a natural choice for the structured dataset we worked with.\n\n\n## Data Preprocessing\n\nWe dedicated significant effort to understanding and preparing the data. Below are the key steps we followed:\n\n### 1. Handling Missing Values\n- **Numeric Columns:** Filled with the median.\n- **Categorical Columns:** Filled with the mode.\n```python\n# Fill numeric columns with median\nfor col in numeric_cols:\n data[col] = data[col].fillna(data[col].median())\n\n# Fill categorical columns with mode\nfor col in categorical_cols:\n data[col] = data[col].fillna(data[col].mode()[0])\n```\n\n### 2. Date Features\nWe engineered date-related features, breaking down date columns into components like year, month, day, and weekend indicators. Additionally, we calculated meaningful time differences:\n```python\ndata[f'{col}_year'] = data[col].dt.year.fillna(-1).astype(int)\ndata[f'{col}_month'] = data[col].dt.month.fillna(-1).astype(int)\ndata[f'{col}_is_weekend'] = (data[col].dt.dayofweek >= 5).astype(int)\n```\n\n### 3. Feature Engineering\nTo enhance the dataset, we created new features:\n- **Pricing Features:**\n - Discount amounts and percentages:\n ```python\n data['discount_amount'] = data['Product_value'] - data['final_payment']\n data['discount_percentage'] = (data['discount_amount'] / data['Product_value'] * 100).clip(0, 100)\n ```\n- **Loyalty Program Features:**\n - Indicators for redeemed points and their value ratio:\n ```python\n data['has_redeemed_points'] = (data['loyalty_points_redeemed'] > 0).astype(int)\n data['points_to_value_ratio'] = data['loyalty_points_redeemed'] / (data['Product_value'] + 1)\n ```\n\n### 4. Encoding Categorical Variables\nWe converted text-based categories into numeric representations using label encoding and one-hot encoding.\n\n\n## Modeling\n\n### 1. Model Selection\nWe utilized **XGBoost**, a gradient boosting algorithm, for its suitability to structured datasets.\n\n### 2. Hyperparameter Tuning\nWe optimized the following hyperparameters through trial and error:\n```python\nparams = {\n 'n_estimators': 1000,\n 'max_depth': 9,\n 'learning_rate': 0.02,\n 'subsample': 0.9,\n 'colsample_bytree': 0.9,\n 'min_child_weight': 1,\n 'gamma': 0.04\n}\n```\n\n### 3. Cross-Validation\nTo validate the model's performance, we implemented 5-fold stratified cross-validation:\n```python\nskf = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)\nfor train_idx, val_idx in skf.split(X_train, y_train):\n model = xgb.XGBClassifier(**params)\n model.fit(X_train[train_idx], y_train[train_idx],\n eval_set=[(X_train[val_idx], y_train[val_idx])],\n early_stopping_rounds=50, verbose=False)\n```\n\n### 4. PCA Features\nWe applied Principal Component Analysis (PCA) to capture latent patterns in the data:\n```python\npca = PCA(n_components=10)\npca_features = pca.fit_transform(X_numeric_transformed)\nfor i in range(pca_features.shape[1]):\n X[f'pca_feature_{i}'] = pca_features[:, i]\n```\n\n### 5. Early Stopping\nTo prevent overfitting, we used early stopping based on validation loss during training.\n\n\n## How We Evaluated It\n\nOur primary evaluation metric was the **weighted F1 score** to account for imbalanced class distributions. We also:\n- Reviewed the classification report to analyze predictions for each category.\n- Used cross-validation to ensure stability across data splits.\n```python\nweighted_f1 = f1_score(y_val, y_pred, average='weighted')\n```\n\n\n## Biggest Challenges\n\n1. **Date Handling:** Engineering meaningful features from date columns required careful thought.\n2. **Categorical Variables:** Balancing effective encoding with model performance.\n3. **Feature Selection:** Ensuring that all new features contributed positively to predictions.\n\n\n## What We Learned\n\n1. **Feature Engineering:** Time spent on creating meaningful features, especially around dates and loyalty programs, greatly improved model performance.\n2. **Importance of Pricing Features:** Discount-related features were highly predictive.\n3. **Cross-Validation:** Ensured that our model was generalizable.\n4. **PCA Features:** Helped uncover patterns missed by manual feature creation.\n\n\n\n", "votes": 2 } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T13:43:46Z", "competition_id": 90922, "competition_slug": "Predict-the-Customer-Satisfaction-CSE-22", "team_id": 13197363, "solution_url": "/competitions/Predict-the-Customer-Satisfaction-CSE-22/writeups/sd-solution-overview" }