Deep learning at work

Intelligence is core to SmarterMe’s vision in the future of work. Whether it is timely customer follow-up, next best actions suggestions, chatbot integration, or ranking for geo-located accounts, deep learning is the foundational technology that is used to power such IQs.

What is Deep Learning?

To answer this question, we ought to first understand supervised machine learning. In supervised machine learning, a prediction model is built based upon a set of input data ( features ) together with the corresponding output ( label). The set of input features is known as the training set; the model is a trained function such that it will give the best estimate or prediction for a new input. For the predicted output, it can either be a class that falls under a finite set of values ( e.g. male, female ) or it can be a real value ( e.g. housing prices ). To illustrate, one can build a model to predict the housing prices in the San Francisco Bay Area. The set of input features can be area footage, number of bedrooms, or the presence of a swimming pool. The training data will consist of the values of these features and the actual selling prices of the houses. Another example will be a credit loan risk model, with features such as age, income, loan size and existing debt, to predict the default rate.

Building an effective predictive model requires selecting a set of relevant features, an exercise that warrants extensive data inspection and experimentation. What if the features are unknown or there are too many to choose from? This is where deep learning’s network architecture would become very useful.

In a deep learning neural network, there are nodes ( the neurons ) that are connected with each other. Additionally, a neural network contains multiple layers of neurons. The first layer is known as the the input layer composed of a set of input features. The output layer consists of neurons representing the different labelled classes. Multiple layers, known as the hidden layers, can be found between the input and output layers. The hidden layers aim to generate intermediate outputs, which in turn will serve as inputs for subsequent layers. Each hidden layer represents a model that needs to be trained. With this multi-layer approach, a deep learning network would find the best overall estimate function for a set of input features.

wavenet

Consider a handwritten digit recognition problem where by scanning images of digits, a model can be trained to determine the exact numerical digit. In this case, the pixel values of a 120x120 image can be used as the input features for the neural network. There will be a total of 14400 features as the input. Hidden layers can be set up to determine characteristics such as edges and lightness of the images. This model would not have been possible for some of the traditional supervised learning algorithms such as Support Vector Machines or logistic regression which can only support a limited number of input features.

Deep learning for Follow-Up

In my previous blog post, we have identified that there is a correlation between the amount of email exchanges in various stages of a sales cycle and the closing of the deal. Based on this schedule, SmarterMe would remind the sales user to follow-up with the customers. Upon seeing the notification, the user can take the action to email, call or text the customer right at that time. However, by dismissing the follow-up reminder, the user can also provide the feedback that it is not the right time for follow-up.

A deep learning model is created in which there are 120 input features, which are divided into 4 groups under the following category

1. the email interaction history in the last 30 days
2. the text interaction history in the last 30 days
3. the call interaction history in the last 30 days
4. the opportunity update interaction history in the last 30 days.

Each group has 30 features representing the number of interactions under the action type in each day of the last 30-day period.

In this same model, there will be 3 potential outputs ( “Dismiss Follow-up”, “Seen Follow-Up”, “Take action”). They represent potential responses from the user. Notice that for the second option, it is equivalent to an no-op where the user has seen the notification but chooses not to respond.

There will be one hidden layer with 4 nodes, each representing the corresponding action type that includes email exchange, text exchange, call exchange and opportunity update.

To seed the model, we should place stronger weight for those days corresponding to the inflection points at which a pivotal result can be yielded. These nodes should be set up to point to the “Take Action” node.

As the user starts providing feedbacks ( or not ) upon seeing the follow-up reminders, these responses would be recorded and serve as the new training data to provide a better result.

To augment the training set, a group of selected customers are shown to the user as potential targets for follow-up. The specific actions taken by the user would be recorded, which would then be used to train and enhance the model. Over time, as more training data is collected, a function will be trained to predict whether a follow-up should be displayed to the user. Specifically, when the corresponding values for the interaction history for a contact for the last 30 days are used as the input for the neural network model, a recommended action for the follow-up will be shown to the user.

The Future is so Bright

Several key strategies are used for this particular neural network model.

1. No extensive historical data is needed for training the model
2. By using the number of interactions for the different communication type as the input features, we provide a very flexible feature sets which can drive the result of the follow-up
3. Feedback from the user will be used to train the model

With this foundational architecture, other enhancements can be easily added. For example, we can expand the number of days beyond 30 days. Or instead of just recommending the user to follow up, a more specific suggestion such as asking the user to make the phone call can be made.

With such Deep Learning at work, the future of work is indeed very bright!