switch to machine learning

main.py

@@ -1,13 +1,14 @@
 import pandas, seaborn
 from requests import get
 from sys import argv, exit
+import sklearn
+import sklearn.linear_model
+import sklearn.metrics
+import sklearn.model_selection
 
 # chack args
-cmds = ["and", "max"]
+if "help" in argv:
-if len(argv) < 2 or argv[1] not in cmds:
+    print(f"""USAGE:\npython3 main.py""")
-    print(f"""USAGE:\npython3 main.py <{'|'.join(cmds)}>
-max = guess fraudulent if the heuristic for the best column finds it
-and = guess fraudulent if ALL the heuristics determined to be good find it""")
     exit(-1)
 
 # download dataset if it isn't already present (too large for github)
@@ -25,54 +26,23 @@ data['mean'] = data.mean(axis=1)
 fraud_set = data.loc[data["Class"] == 1]
 legit_set = data.loc[data["Class"] == 0]
 
-#find the best columns for determining fraud
+#get variables ready for training
-good_heuristics = []
+x = data.loc[:, "V1":"V28"]
-for col_name in fraud_set.columns:
+y = data["Class"]
-    fm = fraud_set[col_name].mean()
-    lm = legit_set[col_name].mean()
-    corr = 0
-    incorr = 0
-    for r in data.iterrows():
-        if abs(r[1][col_name] - fm) < abs(r[1][col_name] - lm):
-            if r[1]["Class"] == 1: corr += 1
-            else: incorr += 1
-        elif abs(r[1][col_name] - fm) > abs(r[1][col_name] - lm):
-            if r[1]["Class"] == 0: corr += 1
-            else: incorr += 1
 
-    print(col_name)
+xtrain, xtest, ytrain, ytest = sklearn.model_selection.train_test_split(x, y, random_state = 0)
-    print(fm)
-    print(lm)
-    accuracy = corr/(corr+incorr)
-    print(accuracy)
-    if (accuracy > .95) and col_name != "Class":
-        print("good heuristic!")
-        good_heuristics.append({"name": col_name, "fraud_mean": fm, "legit_mean": lm, "accuracy": accuracy})
-    print("")
-print(good_heuristics)
 
-# create new dataframe with guesses based on found heuristics and chosen type (max = best column, and = all good columns must match)
+#train
-guessed_class = []
+classifier = sklearn.linear_model.SGDClassifier()
-best_heuristic = None
+classifier.fit(xtrain, ytrain)
-best_acc = 0
-for h in good_heuristics:
-    if h["accuracy"] > best_acc:
-        best_acc = h["accuracy"]
-        best_heuristic = h
 
-print(f"using heuristic: {best_heuristic['name']}")
+#predict
-for r in data.iterrows():
+predictions = classifier.predict(xtest)
-    if argv[1] == "and":
+
-        bools = []
+#save
-        for h in good_heuristics:
+accuracy = sklearn.metrics.accuracy_score(ytest, predictions)
-            bools.append(abs(r[1][h["name"]] - h["fraud_mean"]) < abs(r[1][h["name"]] - h["legit_mean"]))
+save = input(f"accuracy: {accuracy}, save data to .csv? [y/n]")
-        fraud = all(bools)
+data["guess"] = classifier.predict(x).flatten()
-        guessed_class.append(1 if fraud else 0)
+print(data)
-    elif argv[1] == "max":
+if save.lower() == "y":
-        good = True
+    data.to_csv("guesses.csv")
-        if abs(r[1][best_heuristic["name"]] - best_heuristic["fraud_mean"]) < abs(r[1][best_heuristic["name"]] - best_heuristic["legit_mean"]):
-            good = False
-        guessed_class.append(0 if good else 1)
-data["guess"] = guessed_class
-print(data.head(10))
-data.to_csv("woo.csv")

main_old.py

@@ -0,0 +1,78 @@
+import pandas, seaborn
+from requests import get
+from sys import argv, exit
+
+# chack args
+cmds = ["and", "max"]
+if len(argv) < 2 or argv[1] not in cmds:
+    print(f"""USAGE:\npython3 main.py <{'|'.join(cmds)}>
+max = guess fraudulent if the heuristic for the best column finds it
+and = guess fraudulent if ALL the heuristics determined to be good find it""")
+    exit(-1)
+
+# download dataset if it isn't already present (too large for github)
+try: open("creditcard.csv").read()
+except FileNotFoundError:
+    print("downloading dataset...")
+    txt = get("https://the.silly.computer/creditcard.csv").text
+    open("creditcard.csv", "w").write(txt)
+
+# read dataset into dataframe
+data = pandas.read_csv("creditcard.csv")
+data['mean'] = data.mean(axis=1)
+
+# isolate fraud & legitimate sets
+fraud_set = data.loc[data["Class"] == 1]
+legit_set = data.loc[data["Class"] == 0]
+
+#find the best columns for determining fraud
+good_heuristics = []
+for col_name in fraud_set.columns:
+    fm = fraud_set[col_name].mean()
+    lm = legit_set[col_name].mean()
+    corr = 0
+    incorr = 0
+    for r in data.iterrows():
+        if abs(r[1][col_name] - fm) < abs(r[1][col_name] - lm):
+            if r[1]["Class"] == 1: corr += 1
+            else: incorr += 1
+        elif abs(r[1][col_name] - fm) > abs(r[1][col_name] - lm):
+            if r[1]["Class"] == 0: corr += 1
+            else: incorr += 1
+            
+    print(col_name)
+    print(fm)
+    print(lm)
+    accuracy = corr/(corr+incorr)
+    print(accuracy)
+    if (accuracy > .95) and col_name != "Class":
+        print("good heuristic!")
+        good_heuristics.append({"name": col_name, "fraud_mean": fm, "legit_mean": lm, "accuracy": accuracy})
+    print("")
+print(good_heuristics)
+
+# create new dataframe with guesses based on found heuristics and chosen type (max = best column, and = all good columns must match)
+guessed_class = []
+best_heuristic = None
+best_acc = 0
+for h in good_heuristics:
+    if h["accuracy"] > best_acc:
+        best_acc = h["accuracy"]
+        best_heuristic = h
+
+print(f"using heuristic: {best_heuristic['name']}")
+for r in data.iterrows():
+    if argv[1] == "and":
+        bools = []
+        for h in good_heuristics:
+            bools.append(abs(r[1][h["name"]] - h["fraud_mean"]) < abs(r[1][h["name"]] - h["legit_mean"]))
+        fraud = all(bools)
+        guessed_class.append(1 if fraud else 0)
+    elif argv[1] == "max":
+        good = True
+        if abs(r[1][best_heuristic["name"]] - best_heuristic["fraud_mean"]) < abs(r[1][best_heuristic["name"]] - best_heuristic["legit_mean"]):
+            good = False
+        guessed_class.append(0 if good else 1)
+data["guess"] = guessed_class
+print(data.head(10))
+data.to_csv("woo.csv")

readme.txt

@@ -1,8 +1,2 @@
 USAGE:
-python3 main.py <{'|'.join(cmds)}>
+python3 main.py
-max = guess fraudulent if the heuristic for the best column finds it
-and = guess fraudulent if ALL the heuristics determined to be good find it
-
-the goodness of a heuristic is determined by the proportion of rows it can correctly determine the class of
-
-i.e. if the accuracy is .90, guess = Class 90% of the time