Website builders – online hosting services that provide tools for non-technical users to build their own websites  – are frequently exploited by spammers looking for a convenient launching pad. Checking thousands, or sometimes millions, of web pages manually to look for evidence of a spammer is both tedious and inefficient.

In this article, I show how to build a suitable spam-searching neural network with help from Google's TensorFlow machine learning library [2] [3] and TFLearn [4], a library with a high-level API for TensorFlow. Even if you don't spend your days searching for spammers, the techniques described in this article will give you some insights on how to harness the power of neural networks for other complex problems.

Training Day

The neural network needs both positive and negative samples in order to learn. This solution starts with a manually compiled list of sample users divided into spammers and legitimate users, taking care to distribute both types in equal numbers. Alongside this classification (spammer or not spammer), the data set contained the user's name or the website that belongs to the user, the IP address with which the site is registered, and the language version associated with the site.

As a result of the solution described here, the neural network now automatically recognizes new spammers as they register. The next step is to combine this automatic check with a manual check. A Python script automatically blocks sites that the network classifies with a very high probability of being spam, and an employee manually checks the sites that are deemed high probability.

Sound Network

Neural networks are mathematical models that can approximate any function. A neural network is guided by networked neurons similar to those in the human brain, such as in the visual cortex. What makes these networks special is that you do not have to model their behavior explicitly; instead, you train the network using sample data.

Neural networks help out when it is difficult to model functions manually, and they are often used in image and speech recognition. You need to provide the neural network with training data that has already been classified, and it will then attempt to classify new data in a similar way.

A single artificial neuron comprises several weighted inputs and an activation function, which is usually non-linear and helps to determine the output value of the neuron. There is also a threshold value or bias, which complements the weighted inputs, thus influencing the activation function. The mathematical formula behind this concept is as follows:

The formula uses the w vector to weight the input vector x and calculate the sum of both. It then adds the bias b, using the activation function phi. When developers skillfully combine several neurons, they can compute more complex functions (see the box titled "Solving Problems with Neural Networks").

Figure 1: A binary XOR is not linearly separable per se, although it can be split into linearly separable sub-problems.

Networked Learning

Several layers of interconnected neurons form a neural network (Figure 2). These layers consist of at least an input layer, which receives the input values, and an output layer, on which the data arrives after passing through several hidden layers. All the neurons on a particular layer generally use the same activation function.

Figure 2: A neural network with two hidden layers.

Neural networks learn through an optimization process that determines the parameters of the network, the weightings of the connections, and the bias of all the neurons, then refines these values step by step. The process that determines parameters is one of the optimization problems. This process involves the use of many traditional numerical analysis methods (e.g., the gradient method  [5]).

The script first initializes the network using random parameters. Next, the script applies the training data set to the neural network and determines the difference between the network's results and the correct results from the training data. The gap between these results is the loss, which the script attempts to minimize in the course of the optimization process.