Improve script

linear-regression.py

@@ -1,47 +1,78 @@
 import pandas as pd
+import plotly.express as px
 from pandas import DataFrame
 from sklearn import preprocessing
-from sklearn.linear_model import LinearRegression
-import plotly.express as px
-
+from sklearn.linear_model import LinearRegression, Ridge, RidgeCV
+from sklearn.metrics import mean_squared_error
+from sklearn.preprocessing import PolynomialFeatures
 
 col_names = ["price", "mileage", "year", "brand", "engine_type", "engine_cap"]
 col_names_in = ["mileage", "year", "brand", "engine_type", "engine_cap"]
-df_train = pd.read_csv("train/train.tsv", error_bad_lines=False, header=None, sep="\t", names=col_names)
-df = df_train.drop(df_train[df_train["price"] < 1000].index)
-dev0 = pd.read_csv("dev-0/in.tsv", error_bad_lines=False, header=None, sep="\t", names=col_names_in)
-testA = pd.read_csv("test-A/in.tsv", error_bad_lines=False, header=None, sep="\t", names=col_names_in)
+df_train = pd.read_csv(
+    "train/train.tsv", error_bad_lines=False, header=None, sep="\t", names=col_names
+)
+df = df_train
+dev0 = pd.read_csv(
+    "dev-0/in.tsv", error_bad_lines=False, header=None, sep="\t", names=col_names_in
+)
+testA = pd.read_csv(
+    "test-A/in.tsv", error_bad_lines=False, header=None, sep="\t", names=col_names_in
+)
+
+test = pd.read_csv("dev-0/expected.tsv", error_bad_lines=False, header=None, sep="\t")
+Y_a = test[0]
+
+brands = df.brand.value_counts()[:35].index.tolist()
 
 
-def preprocess_data(df: DataFrame) -> DataFrame:
+def preprocess_data(df: DataFrame, brands: list) -> DataFrame:
     """Prepare dataset to linear regression"""
-    df["brand"] = df["brand"].str.lower()
-    scaler = preprocessing.StandardScaler()
-    df[["mileage", "year", "engine_cap"]] = scaler.fit_transform(df[["mileage", "year", "engine_cap"]])
-    enc = preprocessing.LabelEncoder()
-    enc.fit(df[["brand"]])
-    df[["brand"]] = enc.transform(df[["brand"]])
 
-    enc.fit(df["engine_type"])
-    df[["engine_type"]] = enc.transform(df[["engine_type"]])
+    df.brand = df.brand.apply(lambda x: x if x in brands else "0")
+    df["year"] = df.year / 2000
+    df["mileage"] = df.mileage ** 0.3
+    df["engine_cap"] = df.engine_cap * 0.3
+    df["brand"] = df["brand"].str.lower()
+    scaler = preprocessing.RobustScaler()
+    df = pd.get_dummies(df, columns=["brand", "engine_type"])
+    # takes 1k rmse more ;(
+    df[["mileage", "year", "engine_cap", "year"]] = scaler.fit_transform(
+        df[["mileage", "year", "engine_cap", "year"]]
+    )
+    poly = PolynomialFeatures(2, interaction_only=True)
+    df = poly.fit_transform(df)
 
     return df
 
 
-df_train = preprocess_data(df_train)
-dev0 = preprocess_data(dev0)
-testA = preprocess_data(testA)
-
-fig = px.imshow(df_train.corr())
-fig.show()
+indexes = df_train[(df_train.price < 1000) & (df_train.price > 1)].index
+df_train.drop(indexes, inplace=True)
 
+index = df_train[(df_train.mileage > 900000)].index
+df_train.drop(index, inplace=True)
 Y_train = df_train["price"]
+df_train.drop("price", axis=1, inplace=True)
+# df_train = df_train[df_train.price not in range (2, 1000)]
+
+df_train = preprocess_data(df_train, brands)
+dev0 = preprocess_data(dev0, brands)
+testA = preprocess_data(testA, brands)
+
+# fig = px.imshow(df_train.corr())
+# fig.show()
+
 
 lm_model = LinearRegression()
-lm_model.fit(df_train[["mileage", "year", "brand",  "engine_cap"]], Y_train)
 
-dev0_predicted = lm_model.predict(dev0[["mileage", "year", "brand", "engine_cap"]])
-testA_predicted = lm_model.predict(testA[["mileage", "year", "brand", "engine_cap"]])
+# clf = RidgeCV(alphas=[0.1, 0.01, 0.001, 0.00001,  1e-1], cv=10, fit_intercept=True, normalize=True)
+# clf.fit(df_train, Y_train)
+lm_model.fit(df_train, Y_train)
+
+dev0_predicted = lm_model.predict(dev0)
+testA_predicted = lm_model.predict(testA)
+
+# dev0_predicted2 = clf.predict(dev0)
 
 pd.Series(dev0_predicted).to_csv("dev-0/out.tsv", sep="\t", index=False, header=False)
 pd.Series(testA_predicted).to_csv("test-A/out.tsv", sep="\t", index=False, header=False)
+print(mean_squared_error(Y_a, dev0_predicted, squared=False))