Explaining why you should switch to LARUS’ Graph Data Science Platform, Galileo.XAI. The transparent way to represent and exploit your data.
Abstract
Recommendation systems are algorithms aimed at suggesting relevant items to users for their everyday life, such as products to buy, movies to watch, text to read, services to use. Even though Recommendations are primarily used in commercial applications, they can also be applied for example to the Public Administration, when we want to suggest services to citizens, and to many other different sectors and use cases such as fraud detection with credit cards, transfer of money between accounts, anti money laundry, tax evasion, etc. Recommendation systems are really critical as they can be a way to stand out significantly from competitors and help generate a huge amount of income.
In this article, we will describe different methods of recommendation, from the traditional ones to the next-generation, which consists in the application of Deep learning to the most traditional approaches.
Ultimately, Graph Structure plus Graph AI are the most suited for the task because the relations are explicit and therefore scientists can visualise the graph and see that those recommendations are actually eligible. We at Larus understood that not only representing data is important, it is also important to show the connections between data and to have transparency and explainability of the results. Our efforts have therefore converged together with Fujitsu on the development of Galileo.XAI, our Graph Data Science Platform for Explainable AI, where we combine Graph Data Science (algorithms on graphs), Graph Machine Learning and Graph AI plus Explainability.
Recommendation systems
“People don’t know what they want until you show them”.
Steve Jobs
As Steve Jobs used to say, “People don’t know what they want until you show them”. Luckily, businesses are now supported by Recommendation systems, which seek to predict preferences and to suggest the right content to the right user.
But what are Recommendation systems?
Generally speaking, Recommendation systems are algorithms aimed at suggesting relevant items to users for their everyday life, such as products to buy, movies to watch, text to read, services to use or anything else.
In some industries they are really critical as they can generate a huge amount of income when they are efficient or also be a way to stand out significantly from competitors. Many prominent companies’ names can be mentioned in order to show how relevant recommender systems have become in the last decades. Ikea, Netflix and Amazon are just a few of those that grew their business thanks to them. We can also mention that, a few years ago, Netflix organised the “Netflix prize”, a public competition the aim of which was to produce a Recommendation system with a better performance than its own algorithm. $1M prize was offered to the winner.
Indeed, some of the benefits that could derive from efficient recommendations include:
- Decreased customer churn rate, since they are more engaged and loyal
- Improvement in Click-Through Rates
- Surge in average order value
In this article, we will go through different methods of recommendation from the traditional ones to the next-generation, describing how they work and discussing their strengths and weaknesses.
Traditional Recommendation systems
Before approaching the Next generation Recommendation system, let’s see what methods have been deployed until now.
- Knowledge-based method. It is based on explicit specification from the customer. This solution involves the construction of rules and patterns based on things that are known to be true. The problem with this approach is that complex rules and patterns are sometimes hard to describe.
- Content-based method. It is based on users’ ratings and reviews on which a feedback matrix can be built to calculate similarities between customers and items and give recommendations accordingly. This approach can give good performance but doesn’t reach the state of the art.
- Collaborative method. It is a technique that can filter out items that a user might like on the basis of past reactions by similar users, exploiting general behaviour and community ratings. It searches among a large group of people a smaller set of users with tastes similar to a particular user. It then analyses the items they like and combines them to create ad-hoc suggestions. It can be described as based on some kind of neighbourhood approach, like KNN. User-item interactions are thus sufficient to detect similar users and/or similar items and make basic predictions. However, for the system to be accurate, the recorded interactions have to be many and constantly up-to-date.
- Hybrid approach. This method involves a mixture of all the above.
The Next generation Recommendation systems
The Next generation of Recommendation systems consists in the application of Deep learning combined with the methods we previously saw.
Deep learning can be applied to both content and collaborative filtering to obtain better recommendations since it can exploit the implicit feedback from user interactions, which is abundant and can really boost the performance.
There are a lot of Deep learning models that can be used to boost recommendation such as Cooperative Neural Networks, Graph Neural Networks, etc.
We at Larus think that Graph Structure plus Graph AI are the most suited for the task and represents the best structure to deal with relational data because in the graphs the relations are explicit.
Graph AI exploits connections in order to recognize eligible recommendations. We can do so because looking, visualising the graph we can see that those recommendations are actually eligible. Indeed, the structure of the graph itself can be exploited to explain Graph AI decisions.
This is why we know that Graph AIs are really high-performing algorithms for recommendations and achieve the state of the art.
Our journey towards Galileo.XAI, Larus’ Graph Data Science Platform
Larus’ journey with graphs started in 2011 with the first spikes for retails and proceeded with us becoming the 1st Neo4J partner in Europe and the 3rd partner globally.
During this long journey of ours we understood that not only representing data is important, it is also important to show the connections, the links between data. That is the reason why in 2019 we started to partner with Linkurious.
In the last few years we also understood that using graphs with machine learning and AI in general could be a boost for machine learning models and for AI algorithms: we therefore started a partnership with Fujitsu in the field of graphs based on artificial intelligence with explainability. That was because for us it was also important to have transparency and explainability of the results when using, for example, deep learning.
Our efforts have converged in the development of Galileo.XAI, our Graph Data Science Platform for Explainable AI, where we combine Graph Data Science (algorithms on graphs), Graph Machine Learning and Graph AI plus Explainability.
Galileo.XAI is a platform that enables you to import data and create connections between them. One of the most important parts is to use these connections to improve the outcomes, to have new features to enhance the results of AI algorithms.
When we start putting things together and have these explicit connections, then we can get insights. When we get insights we can ask more questions. In this way we create knowledge on our data and we can start to predict things using graphs.
But of course the last step is to have transparency.
We know that sometimes it is impossible to understand how deep learning algorithms work and learn things and that using them entails the presence of some sort of black boxes. Generally speaking, graph architectures are not so simple and graphs are a complex representation of data, therefore we need transparency. This is what our platform can provide: transparency on the outcomes.
Moreover, Galileo.XAI Platform is modular by design. Its foundation can be extended with different models in line with specific needs or verticals. Micro applications tailored with ad-hoc functionalities can then been combined to cover and solve different problems such as recommendation.
Even though recommendations are primarily used in commercial and retail applications, they can also be applied to the Public Administration, when for example we want to suggest services to citizens, and to many other different sectors and use cases such as fraud detection with credit cards, transfer of money between accounts, anti money laundry, tax evasion, etc.
To find more on why we believe Galileo.XAI is the New Generation of Recommendation System that suits you, contact us.