Network Analytics and Neural Networks — The Twain Shall Meet

As Knowledge Graphs / Network Analytics gain increasing traction among analysts looking for meaningful connections and lynchpin actors/nodes, a few key features about networks need to be appreciated a lot better to analyze them effectively.

1. Large Data Volumes: Given that communication between nodes of a network have mutiple layers of metadata [eg: message body, sender/receiver details, message provenance (in case of secure networks), secondary interactions on each message (likes, shares etc)] even a medium sized network can easily generate vast amounts of data.
2. Unstructured Data: Most networks are inherently structured towards generating non-numeric data like text messages, demographic data of the nodes, transaction/message timestamp, video/audio content, transactional data from IoT devices etc. It is a generally accepted fact that 80% of data generated by the internet is Unstructured Data.
3. Nuanced Network Metrics: While there are many metrics for describing a social network, many of these metrics are fairly nuanced and interchangeable. A classic example is Centrality related metrics, where Betweenness and EigenVector Centralities can both be indicative of a node’s “influence”.
4. Massive number of input factors: For doing any kind of Machine Learning we need input factors that serve as independant variables to predict some output variable like, say, probability of someone sharing a post. As you might have noticed in points 1 and 3, every network has a large number of parameters / metrics. If we combine this with the fact that networks generates massive amounts of data, then ML techniques like logistic/linear regression become computationally inefficient. For example: If we are creating a Regression Model for a message corpus of 10,000 messages with 20 input variable, then the machine will have to solve 200,000 ‘equations’.

This is where Neural Networks (NN)come in. For the example mentioned above (i.e. 10,000 messages) a relatively simple NN of 2 hidden layers with 5 nodes each can create a good predictive model.

So, what are Neural Networks? A neural net consists of multiple simple processing nodes that are characterized by their connection weights. A neural net can have one or many layers. Each node in a layer gets inputs from the input nodes / previous layer and the output is computed based on the weights of incoming connections. The output is binary and is computed on the basis of a threshold value. The ‘training’ process of a neural net involves continuously adjusting the node threshold and connection weights to meet the ‘expected’ outputs provided in the training data.

Staying with the theme of Network Analytics, given below is a Neural Network based on a corpus of ~15,000 Twitter interactions. The model looks at 16 parameters of each interaction and computes the chances of an indivdual creating a new Tweet or how she will respond to an existing Tweet (i.e. Mention / Reply / Retweet etc.)

How can we leverage NNs for Network Analytics? Network Analytics (especially in areas like Social Network Analytics and Organizational Network Analytics) is a relatively new, but fairly complex field. There is significant scope for generating a lot of predictive insights around Network Analytics using Neural Networks. A few examples are:

1. Leveraging Natural Language Processing (NLP) to assess the community’s zeitgeist and predicting how future campaigns will be perceived.
2. Helping organizations run effective Influencer Programs and predicting how effective an Influencer will be.
3. Identifying latent Leadership Bandwidth within the organization and predicting who can be trained into effective leaders.

Two points to watch out for:

1. Don’t be overwhelmed by the NN hype: Neural Nets are required only in cases where we need to build a complex model with lots of data and large number of independent factors i.e. input variables. If that is not the case, a relatively simpler approach like Logistic Regression may still be able to create a reasonably accurate model. Using NNs where they are not required, can unnecessarily complicate the model and will also make it difficult to tune the hyper-parameters.
2. NNs as Black Boxes: While NNs are powerful constructs, they are still difficult to “tune” since it is difficult to understand why a prediction was made. Data Scientists are trying to solve this problem by using frameworks like Explainable AI, but it is still an evolving field.

To Conclude: Network Analytics is a promising field and Neural Networks are ideally positioned to help in generating these analytics and generate predictive insights. But for this to work, practitioners need to understand the nuances of Network Analytics and be able to tune Neural Networks which behave like Black Boxes.

A few useful links on Neural Networks and Network Analytics.

a) Andrew NG’s Course on Machine Learning (has a module on Neural Networks) — https://www.coursera.org/learn/machine-learning

b) A good Coursera course that gives a detailed understanding of Network Analytics — https://www.coursera.org/learn/social-economic-networks

c) A Deloitte paper on using Network Analytics for Organizational Redesign — https://www2.deloitte.com/us/en/insights/focus/technology-and-the-future-of-work/organizational-network-analysis-network-of-teams.html