Introduction

Starcount has been developing automated methods to understand the community structure of online society and the influencers within these communities. This helps marketers go beyond individually targeted adverts, identifying better ways for them to engage with relevant communities and allowing them to make positive contributions and offer services and products of real value. With billions of people using social media across the world, pinpointing the right communities to target is a challenging problem. We use Python to develop software that is able to search for influencers, the communities who connect with them, and the passions they share. Ultimately we hope to make social media spam a thing of the past, replacing it with useful information, and positive contributions to engaged communities.

In this talk we will describe an end-to-end Python stack that goes from data acquisition through API calls to social networks all the way to community detection and labelling and show how several major Python libraries are meshed together to achieve these ends.

We break the pipeline down into 4 major components: data acquisition with Twisted, data formatting and network construction with Pandas, Data compression with Numpy and Cython and Data Enrichment with Scikit-learn.

Data Extraction

We have developed a client/server program that enables us to download user profile, connections and user post data from various social networks, including: Twitter, Facebook and Instagram. The server is started with a list of network IDs to perform an operation against and access tokens to validate the operations. Once started it manages access rate limits and the distribution of work across a network of client applications. The clients perform the actual requests to the external APIs using asynchronous HTTPS requests. The Twisted event-driven network engine is used to provide features such as an event loop and asynchronous network calls as well as a simple custom client server capability.

Our core hypothesis is that social networks can be understood, at least for commercial purposes, in terms of the interactions surrounding the major influencers. Due to API limits, restricting the amount of data that can be gathered, managing the size of your network is an important consideration. The final product of our data extraction process is a directory populated with the following file data:

• Follower Metadata
• Influencer Metadata
• Follower-influencer Relationships

Network Construction

We believe that the best way to summarise the interests of social media users is to understand what they follow and so we store links between influencers and their followers. Our graph model differs from the standard network graph as it has two distinct types of nodes, influencers and followers, which are treated differently. The relationships between followers and influencers are key to understanding a network's user base.

Several of our processes require a full linear scan of the follower-influencer connection database. This can be a time-consuming task. We store each of the follower-influencer files in a numpy binary format to enable very fast reads over the data. We have used Cython to speed up this core processing significantly from our starting point of a pure Python implementation. We store each of these files in 1% percent samples, each of which contains roughly 7 million rows and parallelise execution using the multiprocessing and joblib libraries.

Compression

We wish to detect communities of influencers and so we need a way of determining relationship strengths between them. Our method is based on the number of shared followers. The networks are too large and dynamic for it to be practical to store all pairs of similarities and so instead we compress each influencer in a form that allows similarities to be quickly calculated when needed. This compression takes the form of minhash signatures. While using the signatures is very quick, generating them is an expensive operation that must iterate through billions of follower-influencer connections, incrementally updating the signatures. Our original implementation of this algorithm took six days to run.

We were able to make some improvements by profiling the code using cProfiler and line_profiler to remove bottlenecks. Really significant improvements were achieved by pre-processing our input data into a more suitable format (binary numpy memmap) and using a Cython main loop. In this specific case we managed to improve on the speed of numpy matrix operations (with broadcasting) by writing c-code which created less intermediate variables and hence made less calls to create new objects on the heap. The cython annotation tool proved useful in identifying when cPython was actually creating Python objects, especially when these were non-obvious like views on a numpy array. In total these optimisations reduce runtime from six days to three hours.

Enrichment

The final stage of our process uses machine learning algorithms to infer user attributes. These may include demographics data such as country, age and gender; or more advanced features such as robot detection - is the user a person, Twitter bot or company.

We used the scikit-learn package which contains tools to develop machine learning projects quickly. We typically progress by splitting the data into different sets for cross-validation before training a classification model such as a support vector machine or random forest. The initial model would be tuned using scikit-learn’s grid search and finally evaluated using validation and learning curves. Like many other features, scikit-learn’s support vector machine is implemented using fast and highly optimised C libraries.

Conclusion

In this talk we have described how a stack compose entirely from Python components can take raw data directly from social network APIs and manipulate it into a form that allows brand managers to interactively understand the communities and influencers that exist around their products and marketplaces.