Best Project Center in Existing System
The problem addressed can be defined as follows: given a tweet t, classify whether it is a spam or not. In this section, we first discuss our proposed CNN architecture using various word embedding’s with different dimensions to detect the spam at tweet level. Next, we discuss the traditional feature-based model which uses user-based, content-based, and n-gram features for the same problem. Finally, we discuss our neural network-based ensemble architecture to detect whether the given tweet is spam tweet or non-spam tweet.
Convolutional Neural Networks
For better working of any neural network algorithm, there are two major decisions to take. One is feature representation, and another one is network architecture. Here, we describe about these two aspects in detail.
Get AB Technologies’s stories in your inbox
Join Medium for free to get updates from this writer.
Feature Representation
The main features of tweet come from the words contained in it. Each word in the corpus acts as a feature. There are various ways to represent the features such as one-hot vector representation and dense representation. In one-hot vector feature representation, all entries of the vector are 0s except for the entries in which feature is present as best project center in nagercoil.
The value of that entry is 1. For example, assume our vocabulary has six words: actor, actress, cricketer, politician, student, and teacher. The one-hot vector for the word “student” could be: 000010. This is a natural representation to start with, although a poor one. One major drawback with one-hot vector representation is that it is high dimensional.
The dimensionality of the vector depends on the size of the vocabulary. If the vocabulary size is of |V | then a window of k words correspond to an input vector of at least |V |.k units (vectors are concatenated). This feature representation makes no assumption about the word similarity. All words are equally different from each other, and it is difficult to capture the semantics. For example, “apple,” “mango,” and “king” are equally distant in the feature space, despite “apple” should be closer to “mango” than “king” in semantic view.
In dense representation, the features (words) are represented in low dimension. The main advantage of dense vector representation is its generalization power. Similar features can have similar vectors in dense representations. However, this behavior is not there in one-hot vector representation.
Another advantage is its computational speed because of low dimensions. In cases where there is less number of features and no correlation between the features then one-hot vector representation can be used otherwise dense representation is better.
