Machine learning is creeping into our everyday lives – maybe it recommended the last TV show you watched, or was responsible for routing your recent car journey. Now, machine learning might also curate your next museum visit, or help the institution’s executives understand how visitors behave to deliver better experiences for visitors while achieving business and social good outcomes.
Machine learning is a branch of artificial intelligence, enabled by data science. These three areas represent the intersection of advanced mathematics, computer engineering and business analysis to discover innovations for consumer and business alike. Machine learning is a set of methods which help computers understand a situation and discover answers to problems, without necessarily knowing a set formula.
These methods fall across a spectrum of intelligence: supervised, unsupervised, semi supervised (a hybrid of the earlier two types, often a reflection of cost and practicality) or reinforced learning. Deciding on a method depends on what data is available, how much the computer needs to be responsible for its own learning and what outcomes the model needs to support.
Supervised learning learns known patterns often modelled from historical data, with either continuous models (such as linear regression, random forest or gradient boosted trees), categorical models (including logistic regression or support vector machine) or a model adaptable to both (such as K nearest neighbors). Examples of this in the cultural sector include predicting attendance, classifying donors likely to renew and automating outreach for those at risk of lapse, or targeting marketing in conjunction with customer relationship management.
Unsupervised learning finds structure where there is none defined, with no known patterns – often finding similarities or differences, identifying outliers or self organizing. Again, these can be continuous (such as principle component analysis or hierarchical clustering), categorical (association rule analysis, such as Apirori or frequent pattern growth) or a model potentially bridging both (such as K means). Industry applications of this for museums include discovering visitor behavior insights from presence data or basket analysis for automating upsell suggestions and package product development.
Reinforcement learning generates learning patterns, often through playing out different scenarios. This could be a framework such as the Markov decision process and is often used in game theory. In the museum, reinforcement learning may start to appear in indoor trail route recommendations, art authentication, game theory for advanced digital experience or robotics for visitor services.
The future area of machine learning innovation that proposes to be revolutionary for the arts is in machine vision – technology inferring insight from visual input, not just being able to receive information visually, but being empowered to convert this visual into data for machine learning. As venture capitalist Benedict Evans puts it, “We’re going from computers with cameras that take photos, to computers with eyes that can see”. In the museum, machine vision could be used to identify subject matter, such as generating descriptions of art for collections management systems. In a similar vein, it could be used to recognize similarities and patterns among works in museum collections, to help with developing the interconnectedness of collections data (useful for the ‘you may also like…’ visitor recommendation) or comparing and connecting the collections of multiple museums (for example, achieving similarly personalized yet city wide cultural experiences). Machine vision could even be used for sentiment analysis, such as determining how visitors are responding to an exhibition from a video feed of the physical gallery space complete with facial recognition.
Implementing machine learning in museums
Despite the technical complexities, often the most difficult challenge to master when implementing machine learning is in the business application for this technology. Defining a detailed business challenge with clear outcomes and managing a consultative research discipline with a set hypothesis and experimental validation clears the path for data scientists to explore innovation opportunities.
On the technical side, the data may require new methods of instrumentation (determining visitor sentiment in the gallery would be dependent on camera infrastructure for example). Data may require integrity investigation, cleansing treatments and may invoke discussions around governance, such as privacy concerns. For automated solutions, the data sources will also require integration (and the resulting maintenance of this) with the remainder solution.
Once data is available and depending on the solution architecture, a technology will need to be selected. This will largely be dependent on whether the analysis will be conducted on occasion (at which point a language such as R could be employed), or productized into a touch free solution. Either way, the model needs to be:
- Researched and validated against a series of variables or features
- Trained against a sample data set
- Tested against an unseen data set as well
Importantly, models should be kept simple, rather than risk overfitting – other than adding unnecessary complexity under decreasing returns, this can ultimately impact accuracy.
As with many areas of data science, machine learning presents a huge opportunity for advancing visitor experience and museum operations. Whilst many of these applications are longer term aspirations for the museum field, big data analytics addresses an important foundation for this innovation that delivers immediate business benefits to revenue, efficiency and visitor engagement.