Leveraging Deep Learning and Tensorflow.js to Learn Workflows
This example uses Tensorflow.js to create and train a deep neural network on the client side to detect workflow based on navigation through a site. In general, for a web application, there will a few unique "pathways" through the application that a user will repeatedly take. The customized neural network will learn the pathways that an individual takes through an application and start to predict your next move. This is displayed as "quick actions", and also a graph showing certainty about each next action.
To use, choose a few patterns and click around following those patterns for about 100 clicks. Then click the "Learn" button. This will build and train the model based on the first usage. A current limitation of the system is that the model must manually be created before it will start predictions. After that point, it will periodically re-train to improve predictions and learn new patterns.
https://edamtoft.github.io/WorkflowDetection/index.html
This sample uses a simple deep neural network and trains the model as a state machine based off the last 5 actions. This keeps the model simple, but also sophisticated enough to do a reasonably good job at detecting simple patterns.
The network takes a 5x8 input (5 most recent pages * 8 unique pages represented as one-hot tensors) and generates a flat 8 wide output which represents probabilities of each individual page. There are two hidden layers with 10 and 15 neurons respectively. This was a mostly arbitrary choice that seemed to provide relatively good results. All layers use ReLU activation except for the output layer which uses Softmax.
A recurrent neural network using LSTM nodes may also do a good job and be better fitted to this problem, but during testing, a basic deep neural network performed as well if not better and was much faster to train.