{"metadata":{"kernelspec":{"name":"python3","display_name":"Python 3","language":"python"},"language_info":{"name":"python","version":"3.10.14","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceId":50160,"databundleVersionId":7921029,"sourceType":"competition"}],"dockerImageVersionId":30787,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"## Objective  \nMy task was to design an algorithm to output the **credit risk score** for each bank customer based on the provided customer data. This involved tackling a dataset of significant size and complexity:  \n- **Dataset size**: Several million records for training dataset and unkonwn number of records for testing dataset\n- **Fields**: Over 400 attributes per customer. Attributes are anonymized for security reasons   \n\n---\n\n## Solution Approach  \n  \nI adopted **GBDT (Gradient Boosting Decision Trees)**, a robust boosting algorithm composed of multiple decision trees. It is suitable for solving such problems. \n\n### **Steps Taken**:  \n1. **Data transforming**:\n   - Handle fields of date, Converting to the number of days relative to a reference date.\n   - Aggragte records from each customer by functions of min mean max count.\n   - Filter features based on their data-missing ratio.\n   - Covert to Pandas type.\n\n2. **Data Partitioning**:  \n   - Divided the large dataset into 6 parts for cross validation.  \n\n3. **Model Training**:  \n   - Trained six GBDT models, each focusing on a specific part of the training dataset.  \n\n4. **Model Prediction**:  \n   - Combined the weighted outputs of all the trained GBDT models to produce the final credit risk score.\n\n---\n\n## Results  \nThe algorithm performed well, achieving a **top 30% ranking on the Kaggle leaderboard** for this competition in September 2024. This demonstrated the effectiveness of my solution in handling large-scale datasets and delivering accurate results.\n","metadata":{}},{"cell_type":"code","source":"import gc\nfrom glob import glob\nimport copy\nimport os\nimport numpy as np\nimport pandas as pd\nimport polars as pl\nimport warnings\nwarnings.filterwarnings('ignore')\nfrom pathlib import Path\nfrom sklearn.model_selection import StratifiedGroupKFold\nfrom sklearn.metrics import roc_auc_score\nimport lightgbm as lgb","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-23T05:24:58.115327Z","iopub.execute_input":"2025-01-23T05:24:58.116094Z","iopub.status.idle":"2025-01-23T05:24:58.120703Z","shell.execute_reply.started":"2025-01-23T05:24:58.116059Z","shell.execute_reply":"2025-01-23T05:24:58.119787Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Ensure the fields in the train and test datasets are consistent when applying feature filtering and type transformation\nclass train_test_allignment:\n    keep_columns = None\n    cat_cols = None\n    dtype_guidance = {}","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-23T05:24:58.122297Z","iopub.execute_input":"2025-01-23T05:24:58.122552Z","iopub.status.idle":"2025-01-23T05:24:58.133024Z","shell.execute_reply.started":"2025-01-23T05:24:58.122527Z","shell.execute_reply":"2025-01-23T05:24:58.132349Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class Data_transformer:\n\n    def read_file(path, depth=None):\n        \n        df = pl.read_parquet(path)\n        df = df.pipe(Data_transformer.set_dtypes)\n        if depth in [1,2]:\n            df = df.group_by(\"case_id\").agg(Data_aggregator.get_exprs(df))\n        print('complete {}'.format(path))\n        return df\n\n    def read_files(regex_path, depth=None):\n        \n        chunks = []\n        for path in glob(str(regex_path)):\n            chunks.append(Data_transformer.read_file(path, depth))\n        df = pl.concat(chunks, how=\"vertical_relaxed\")\n        df = df.unique(subset=[\"case_id\"])\n        return df\n\n    # This function sets the data types of DataFrame columns based on column names\n    # - Columns like \"case_id\", \"WEEK_NUM\", \"num_group1\", \"num_group2\" are cast to Int64.\n    # - Columns ending with 'P' or 'A' are cast to Float64.\n    # - Columns ending with 'M' are cast to String.\n    # - Columns ending with 'D' are cast to Date.\n    def set_dtypes(df):\n\n        for col in df.columns:\n            if col in train_test_allignment.dtype_guidance.keys():\n                df = df.with_columns(pl.col(col).cast(train_test_allignment.dtype_guidance[col]))\n            else:\n                if col in [\"case_id\", \"WEEK_NUM\", \"num_group1\", \"num_group2\"]:\n                    df = df.with_columns(pl.col(col).cast(pl.Int64))\n                elif col in [\"date_decision\"]:\n                    df = df.with_columns(pl.col(col).cast(pl.Date))\n                elif col[-1] in (\"P\", \"A\"):\n                    df = df.with_columns(pl.col(col).cast(pl.Float64))\n                elif col[-1] in (\"M\",):\n                    df = df.with_columns(pl.col(col).cast(pl.String))\n                elif col[-1] in (\"D\",) or 'year' in col:\n                    df = df.with_columns(pl.col(col).cast(pl.Date))\n                train_test_allignment.dtype_guidance[col] = df.schema[col]\n        return df\n\n    # Transform all dates into number of days by calculating the difference from 'date_decision' \n    def handle_dates(df):\n\n        for col in df.columns:\n            if col[-1] in (\"D\",):\n                df = df.with_columns(pl.col(col) - pl.col(\"date_decision\"))\n                df = df.with_columns(pl.col(col).dt.total_days())\n        df = df.drop(\"date_decision\", \"MONTH\")\n        return df\n\n    # Filter out fields with many missing value or many options\n    def filter_cols(df, isnull_per = 0.7, freq_num = 200):\n\n        for col in df.columns:\n            if col not in [\"target\", \"case_id\", \"WEEK_NUM\"]:\n                isnull = df[col].is_null().mean()\n                if isnull > isnull_per:\n                    df = df.drop(col)\n                    \n        for col in df.columns:\n            if (col not in [\"target\", \"case_id\", \"WEEK_NUM\"]) & (df[col].dtype == pl.String):\n                freq = df[col].n_unique()\n                if (freq == 1) | (freq > freq_num):\n                    df = df.drop(col)\n        return df\n\n    # Combine all tables into one tale\n    def join_tables(df_base, other_tables: list):\n        \n        for i, df in enumerate(other_tables):\n            df_base = df_base.join(df, how=\"left\", on=\"case_id\", suffix=f\"_{i}\")\n        return df_base\n    \n    def to_pandas(df_data, cat_cols=None):\n\n        df_data = df_data.to_pandas()\n        if cat_cols is None:\n            cat_cols = list(df_data.select_dtypes(\"object\").columns)\n        df_data[cat_cols] = df_data[cat_cols].astype(\"category\")\n        return df_data, cat_cols","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-23T05:24:58.134148Z","iopub.execute_input":"2025-01-23T05:24:58.134398Z","iopub.status.idle":"2025-01-23T05:24:58.150936Z","shell.execute_reply.started":"2025-01-23T05:24:58.134367Z","shell.execute_reply":"2025-01-23T05:24:58.150285Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class Data_aggregator:\n    \n    def num_expr(df):\n\n        cols = [col for col, dtype in df.schema.items() if col[-1] in (\"P\", \"A\",) or dtype == pl.Float64]\n        expr = [pl.max(col).alias(f\"max_{col}\") for col in cols] \\\n            + [pl.mean(col).alias(f\"avg_{col}\") for col in cols]\n        return expr\n    \n    def date_expr(df):\n\n        cols = [col for col in df.columns if col[-1] in (\"D\")]\n        expr = [pl.max(col).alias(f\"max_{col}\") for col in cols] \\\n            + [pl.min(col).alias(f\"min_{col}\") for col in cols]  \n        return expr\n    \n    def str_expr(df):\n\n        cols = [col for col, dtype in df.schema.items() if col[-1] in (\"M\",) or dtype == pl.String]\n        expr = [pl.n_unique(col).alias(f\"count_{col}\") for col in cols] \\\n            + [pl.col(col).mode().first().alias(f\"mode_{col}\") for col in cols]\n        return expr\n    \n    def count_expr(df):\n\n        cols = [col for col in df.columns if \"num_group\" in col] # max & replace col name\n        expr = [pl.max(col).alias(f\"max_{col}\") for col in cols]\n        return expr\n    \n    def get_exprs(df):\n\n        exprs = Data_aggregator.num_expr(df)   + \\\n                Data_aggregator.date_expr(df)  + \\\n                Data_aggregator.str_expr(df)   + \\\n                Data_aggregator.count_expr(df)\n        return exprs","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-23T05:24:58.151709Z","iopub.execute_input":"2025-01-23T05:24:58.151913Z","iopub.status.idle":"2025-01-23T05:24:58.164729Z","shell.execute_reply.started":"2025-01-23T05:24:58.151891Z","shell.execute_reply":"2025-01-23T05:24:58.163933Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Read row train dataset\nROOT            = Path(\"/kaggle/input/home-credit-credit-risk-model-stability\")\nTRAIN_DIR       = ROOT / \"parquet_files\" / \"train\"\ntrain_data_row = {\n    \"df_base\": Data_transformer.read_file(TRAIN_DIR / \"train_base.parquet\"),\n    \"depth_0\": [\n        Data_transformer.read_file(TRAIN_DIR / \"train_static_cb_0.parquet\"),\n        Data_transformer.read_files(TRAIN_DIR / \"train_static_0_*.parquet\"),\n    ],\n    \"depth_1\": [\n        Data_transformer.read_files(TRAIN_DIR / \"train_applprev_1_*.parquet\", 1),\n        Data_transformer.read_file(TRAIN_DIR / \"train_tax_registry_a_1.parquet\", 1),\n        Data_transformer.read_file(TRAIN_DIR / \"train_tax_registry_b_1.parquet\", 1),\n        Data_transformer.read_file(TRAIN_DIR / \"train_tax_registry_c_1.parquet\", 1),\n        Data_transformer.read_files(TRAIN_DIR / \"train_credit_bureau_a_1_*.parquet\", 1),\n        Data_transformer.read_file(TRAIN_DIR / \"train_credit_bureau_b_1.parquet\", 1),\n        Data_transformer.read_file(TRAIN_DIR / \"train_other_1.parquet\", 1),\n        Data_transformer.read_file(TRAIN_DIR / \"train_person_1.parquet\", 1),\n        Data_transformer.read_file(TRAIN_DIR / \"train_deposit_1.parquet\", 1),\n        Data_transformer.read_file(TRAIN_DIR / \"train_debitcard_1.parquet\", 1),\n    ],\n    \"depth_2\": [\n        Data_transformer.read_file(TRAIN_DIR / \"train_credit_bureau_b_2.parquet\", 2),\n        Data_transformer.read_files(TRAIN_DIR / \"train_credit_bureau_a_2_*.parquet\", 2),\n        Data_transformer.read_file(TRAIN_DIR / \"train_applprev_2.parquet\", 2),\n        Data_transformer.read_file(TRAIN_DIR / \"train_person_2.parquet\", 2)\n    ]\n}","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-23T05:24:58.165970Z","iopub.execute_input":"2025-01-23T05:24:58.166241Z","iopub.status.idle":"2025-01-23T05:30:11.666688Z","shell.execute_reply.started":"2025-01-23T05:24:58.166218Z","shell.execute_reply":"2025-01-23T05:30:11.665806Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# add meaningful columns\ntrain_data_row['df_base'] = (\n    train_data_row['df_base']\n    .with_columns(\n        month_decision = pl.col(\"date_decision\").dt.month(),\n        weekday_decision = pl.col(\"date_decision\").dt.weekday(),\n    )\n)\n\n# combine tables by join operation\ndf_train = Data_transformer.join_tables(train_data_row['df_base'], \\\n    train_data_row['depth_0'] + train_data_row['depth_1'] + train_data_row['depth_2'])\n\n# transform date into number of days\ndf_train = df_train.pipe(Data_transformer.handle_dates)\nprint(\"train data shape:\\t\", df_train.shape)\n\n# Drop the useless features\ndf_train = Data_transformer.filter_cols(df_train, 0.7, 200)\nprint(\"train data shape:\\t\", df_train.shape)\n\n# Convert Polars data to Pandas DataFram\ndf_train, cat_cols = Data_transformer.to_pandas(df_train)\ntrain_test_allignment.cat_cols = copy.deepcopy(cat_cols)\n\n# clear row data to save Ram space\ndel train_data_row\ngc.collect()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-23T05:30:11.668603Z","iopub.execute_input":"2025-01-23T05:30:11.669403Z","iopub.status.idle":"2025-01-23T05:30:42.357052Z","shell.execute_reply.started":"2025-01-23T05:30:11.669357Z","shell.execute_reply":"2025-01-23T05:30:42.356176Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Create training dataset and validating dataset\nX = df_train.drop(columns=[\"target\", \"case_id\", \"WEEK_NUM\"])\ny = df_train[\"target\"]\nweeks = df_train[\"WEEK_NUM\"]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-23T05:30:42.358066Z","iopub.execute_input":"2025-01-23T05:30:42.358303Z","iopub.status.idle":"2025-01-23T05:30:44.328909Z","shell.execute_reply.started":"2025-01-23T05:30:42.358279Z","shell.execute_reply":"2025-01-23T05:30:44.327864Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# create GBDT model\nparams = {\n    \"boosting_type\": \"gbdt\",\n    \"objective\": \"binary\",\n    \"metric\": \"auc\",\n    \"max_depth\": 12,  \n    \"learning_rate\": 0.05,\n    \"n_estimators\": 2000,  \n    \"colsample_bytree\": 0.8,\n    \"colsample_bynode\": 0.8,\n    \"random_state\": 42,\n    \"reg_alpha\": 0.1,\n    \"reg_lambda\": 10,\n    \"extra_trees\":True,\n    'num_leaves':64,\n    \"device\": \"gpu\",\n    'gpu_device_id':0,\n    'gpu_platform_id':1,\n    \"verbose\": -1,\n}\nmodel = lgb.LGBMClassifier(**params)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-23T05:30:44.331584Z","iopub.execute_input":"2025-01-23T05:30:44.332085Z","iopub.status.idle":"2025-01-23T05:30:44.337913Z","shell.execute_reply.started":"2025-01-23T05:30:44.332032Z","shell.execute_reply":"2025-01-23T05:30:44.336712Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# train modles with using a K-fold cross-validated dataset\ncv = StratifiedGroupKFold(n_splits=5, shuffle=False)\nfitted_models = []\ncv_scores = []\nfor idx_train, idx_valid in cv.split(X, y, groups=weeks):\n    X_train, y_train = X.iloc[idx_train], y.iloc[idx_train]\n    X_valid, y_valid = X.iloc[idx_valid], y.iloc[idx_valid]\n    \n    model.fit(\n        X_train, y_train,\n        eval_set = [(X_valid, y_valid)],\n        callbacks = [lgb.log_evaluation(200), lgb.early_stopping(60)] )\n    fitted_models.append(model)\n    \n    y_pred_valid = model.predict_proba(X_valid)[:,1]\n    auc_score = roc_auc_score(y_valid, y_pred_valid)\n    cv_scores.append(auc_score)\n\n    model.booster_.save_model('./lgb_model_{}.txt'.format(auc_score))\n\ncv_scores = np.array(cv_scores)\nprint(\"CV AUC scores: \", cv_scores)\nprint(\"Maximum CV AUC score: \", max(cv_scores))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-23T05:35:45.656203Z","iopub.execute_input":"2025-01-23T05:35:45.656804Z","iopub.status.idle":"2025-01-23T05:57:07.062345Z","shell.execute_reply.started":"2025-01-23T05:35:45.656769Z","shell.execute_reply":"2025-01-23T05:57:07.061425Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Prior to this, all training has been completed. Next comes the testing phase.","metadata":{}},{"cell_type":"code","source":"# load trained models\nfitted_models = []\ncv_scores = []\nmodel_files = [f for f in os.listdir('./') if \"lgb_model\" in f]\nfor model_file in model_files:\n    fitted_models.append(lgb.Booster(model_file=os.path.join('./', model_file)))\n    cv_scores.append(float(model_file.replace('lgb_model_', '').replace('.txt','')))\ncv_scores = np.array(cv_scores)\nprint(cv_scores)\n\ntrain_test_allignment.keep_columns = copy.deepcopy(df_train.columns)\ndel df_train\ngc.collect()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-23T05:59:52.266305Z","iopub.execute_input":"2025-01-23T05:59:52.267105Z","iopub.status.idle":"2025-01-23T05:59:52.578531Z","shell.execute_reply.started":"2025-01-23T05:59:52.267069Z","shell.execute_reply":"2025-01-23T05:59:52.577683Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# read row test data\nTEST_DIR = ROOT / \"parquet_files\" / \"test\"\ntest_data_row = {\n    \"df_base\": Data_transformer.read_file(TEST_DIR / \"test_base.parquet\"),\n    \"depth_0\": [\n        Data_transformer.read_file(TEST_DIR / \"test_static_cb_0.parquet\"),\n        Data_transformer.read_files(TEST_DIR / \"test_static_0_*.parquet\"),\n    ],\n    \"depth_1\": [\n        Data_transformer.read_files(TEST_DIR / \"test_applprev_1_*.parquet\", 1),\n        Data_transformer.read_file(TEST_DIR / \"test_tax_registry_a_1.parquet\", 1),\n        Data_transformer.read_file(TEST_DIR / \"test_tax_registry_b_1.parquet\", 1),\n        Data_transformer.read_file(TEST_DIR / \"test_tax_registry_c_1.parquet\", 1),\n        Data_transformer.read_files(TEST_DIR / \"test_credit_bureau_a_1_*.parquet\", 1),\n        Data_transformer.read_file(TEST_DIR / \"test_credit_bureau_b_1.parquet\", 1),\n        Data_transformer.read_file(TEST_DIR / \"test_other_1.parquet\", 1),\n        Data_transformer.read_file(TEST_DIR / \"test_person_1.parquet\", 1),\n        Data_transformer.read_file(TEST_DIR / \"test_deposit_1.parquet\", 1),\n        Data_transformer.read_file(TEST_DIR / \"test_debitcard_1.parquet\", 1),\n    ],\n    \"depth_2\": [\n        Data_transformer.read_file(TEST_DIR / \"test_credit_bureau_b_2.parquet\", 2),\n        Data_transformer.read_files(TEST_DIR / \"test_credit_bureau_a_2_*.parquet\", 2),\n        Data_transformer.read_file(TEST_DIR / \"test_applprev_2.parquet\", 2),\n        Data_transformer.read_file(TEST_DIR / \"test_person_2.parquet\", 2)\n    ]\n}","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-23T06:00:00.991105Z","iopub.execute_input":"2025-01-23T06:00:00.991442Z","iopub.status.idle":"2025-01-23T06:00:01.351818Z","shell.execute_reply.started":"2025-01-23T06:00:00.991411Z","shell.execute_reply":"2025-01-23T06:00:01.350929Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# add meaningful columns\ntest_data_row['df_base'] = (\n    test_data_row['df_base']\n    .with_columns(\n        month_decision = pl.col(\"date_decision\").dt.month(),\n        weekday_decision = pl.col(\"date_decision\").dt.weekday(),\n    )\n)\n\n# combine tables by join operation\ndf_test = Data_transformer.join_tables(test_data_row['df_base'], \\\n    test_data_row['depth_0'] + test_data_row['depth_1'] + test_data_row['depth_2'])\n\n# transform date into number of days\ndf_test = df_test.pipe(Data_transformer.handle_dates)\nprint(\"train data shape:\\t\", df_test.shape)\n\n# Drop features like training dataset\ndf_test = df_test.select([col for col in train_test_allignment.keep_columns if col != \"target\"])\nprint(\"train data shape:\\t\", df_test.shape)\n\n# Convert Polars data to Pandas DataFram\ndf_test, _ = Data_transformer.to_pandas(df_test, train_test_allignment.cat_cols)\n\n# clear row data to save Ram space\ndel test_data_row\ngc.collect()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-23T06:00:05.460836Z","iopub.execute_input":"2025-01-23T06:00:05.461193Z","iopub.status.idle":"2025-01-23T06:00:05.684632Z","shell.execute_reply.started":"2025-01-23T06:00:05.461160Z","shell.execute_reply":"2025-01-23T06:00:05.683703Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Predict risk scores for each bank customer\ndef predict(models, X_test, models_weight):\n    y_preds = pd.DataFrame(index=X_test.index,columns = ['pred_'+str(i) for i in range(0,len(models))])\n    for i in range(0,len(models)):\n        try:\n            y_preds.loc[:, 'pred_'+str(i)] = np.array(models[i].predict_proba(X_test)[:,1]) * models_weight[i]\n        except:\n            y_preds.loc[:, 'pred_'+str(i)] = np.array(models[i].predict(X_test)) * models_weight[i]\n    return y_preds.sum(1)\n\nX_test = df_test.drop(columns=[\"WEEK_NUM\"])\nX_test = X_test.set_index(\"case_id\")\nmodels_weight = cv_scores / cv_scores.sum()\nlgb_pred = pd.Series(predict(fitted_models, X_test, models_weight), index=X_test.index)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-23T06:00:09.431321Z","iopub.execute_input":"2025-01-23T06:00:09.432271Z","iopub.status.idle":"2025-01-23T06:00:09.643529Z","shell.execute_reply.started":"2025-01-23T06:00:09.432229Z","shell.execute_reply":"2025-01-23T06:00:09.642705Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# save prediction results\ndf_subm = pd.read_csv(ROOT / \"sample_submission.csv\")\ndf_subm = df_subm.set_index(\"case_id\")\ndf_subm[\"score\"] = lgb_pred\ndf_subm.head()\ndf_subm.to_csv(\"submission.csv\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-01-23T06:00:13.254074Z","iopub.execute_input":"2025-01-23T06:00:13.254415Z","iopub.status.idle":"2025-01-23T06:00:13.270723Z","shell.execute_reply.started":"2025-01-23T06:00:13.254386Z","shell.execute_reply":"2025-01-23T06:00:13.270080Z"}},"outputs":[],"execution_count":null}]}