Data as a CSV file is collected from Kaggle. Data is pre-processed using,

stop_words = stopwords.words('english')
lemma = WordNetLemmatizer()

def clean_text(text):
    text = text.lower() # lowering
    text = text.encode("ascii", "ignore").decode() # non ascii chars
    text = re.sub(r'\n',' ', text) # remove new-line characters
    text = re.sub(r'\W', ' ', text) # special chars
    text = re.sub(r'\s+[a-zA-Z]\s+', ' ', text) # single characters
    text = re.sub(r'\^[a-zA-Z]\s+', ' ', text) # single char at first
    text = re.sub(r'[0-9]', ' ', text) # digits
    text = re.sub(r'\s+', ' ', text, flags=re.I) # multiple spaces
    return ' '.join([lemma.lemmatize(word) for word in word_tokenize(text) if word not in stop_words])

Whole data contains a total of 20,800 records, of which 15,600(75%) are used for training and 5200(25%) are for tesing purposes. Count of 1(unreliable) label is 10,413(50.06%) and count of 0(reliable) label is 10,387(49.94%).

Three approaches are taken,

1. Using only text and label columns to train and test.
2. Using title, author and label columns for training and testing.
3. Using title, author, text and label columns for training and testing.

In each approach, nan values are filled with titleunknown, authorunknown and textunknown for title, author and text columns respectively (if required). During text-to-feature conversion period TfidfVectorizer is fitted on training data with ngram_range=(1, 2). Then PassiveAggressiveClassifier is trained on training data and used to predict the testing data.

MIT License