Reply is the place to meet an incredible variety of enthusiastic, passionate, ideas-driven people, who want to make a difference and an impact.Would you like to know more?
Consumer companies have been the first to respond to changing customer expectations -think Amazon, Netflix, Walmart and Spotify. In many ways, our customer experience with these brands is about far more than what we are buying from them, it is also about the experience of buying from them.
For instance, every Monday over 100 million Spotify users found a fresh new playlist waiting for them; it’s called Discover Weekly and it’s a custom mixtape of 40 songs they’ve never listened to before but will probably love.
This feature proves there is a high degree of Customer Engagement and at the same time shows how Recommendation Engines have become a crucial aspect of an online User Experience.
You can think of a recommendation engine as a black box that takes some data in input (related to the user-item interaction) and produces in output a list of items paired and ordered by a score, this score is the index of propensity that the user likes the item.
Recommendation engines harvest real-time and historical data from customers, time series events such as frequency of visits to a web site, pages they have visited, time spent on specific pages, products clicked on, and on purchases made, in order to find relationships between products and services based on the inherent complementary nature of items. All customers’ interactions across multiple channels need to be collected and analyzed to deliver an interactive, engaging, convenient and accessible customer experience. In markets such as retail, recommendations have shown to deliver a 3-5% revenue uplift and a 30-70% increase in order value per visitor on average.
Nowadays machine learning technologies are becoming mainstream, tools like Apache Spark or PredictionIO give a set of already implemented and well tested Recommendation algorithms that allows everyone to build their own recommendation system.
Reply has developed the framework Robotics for Customers, which is the synthesis of the different Reply Companies and allows customers to build a time-to-value Recommendation System that can be easily integrated into any existing platform.
The data-driven process starts from the early stage of the developing lifecycle, which is very important when you are building recommender systems. The simple process of coding/releasing without doing the A/B Test brings the project to a certain failure. The framework Robotics for Customers provides a series of small, agile steps in order to reduce/eliminate every possible point of failure:
• Validate: in this phase, we use the source data in order to work out how to define the success metrics of the project. It is important to define the success metrics at this time, in order to clarify the goals of the project and to prevent the result of the A/B Test giving the wrong impression, these metrics must be defined by taking into account the following three aspects: - Audience: how big is the audience you are reaching? - Depth: what is the penetration degree in the audience? - Retention: for the users you reach, how many do you retain? • Prototype: deliver a first small prototype; • A/B Test - Refinement: this can involve many iterations, using the A/B Test on progressively bigger groups will ensure constant feedback about the prototype; these feedbacks will be used in order to improve the original prototype; • Release: release the product into production.
The framework Robotics for Customers was successfully adopted in projects and initiatives in different domains, helping the customers to build time-to-value Recommendation Systems in a cost-effective way.
In order to choose the appropriate recommender system, some issues must be taken into consideration; the best algorithm is not always the one that gives the most correct overall recommendations. In a dataset containing data that belongs to many different groups (classes), such as item groups or different genres, the precision of recommendations can be different for different kinds of items. This means that some users will get accurate recommendations, while others, who prefer some other type of item which are hard to predict, will get significantly lower precision. We have identified four different factors as being major concerns when making a decision about the choice of algorithm, as follows:
• rating schemes, there are two kinds of schemes: - implicit: when the user-product interaction is a binary value (1 if an interaction exists between the user and the product, 0 otherwise); - explicit: when the user explicitly rates the product (most often using a scale from 1 to 5) • user-item interaction: the recommendation provided takes advantage of the analysis of just one behavioural event for the user-item interaction, or maybe of the combination of the co-occurrence of several events (user/buy/item, user/search-for/item, etc.); • computational time: most of recommender systems require a batch computation in order to provide prediction, so the computational time is an important parameter for evaluating the performance of a recommendation algorithm; • explaining ability: the model must be self-explainable; the provided prediction must have a sense.
Reply has acquired deep knowledge in solving this kind of problem in various market environments, allowing the company to identify the right algorithm or even to benchmark a group of them with ease.
Reply developed several PoCs by using state-of-the-art open source Big Data technologies, these are now available as packaged solutions that enable you:
• to quickly build and deploy a recommendation engine as a micro-service in production with customizable algorithm templates; • to respond to dynamic queries in real-time once deployed as a micro-service; • to evaluate and tune multiple algorithm variants systematically; • to unify data from multiple platforms in batch or in real-time for comprehensive predictive analytics; • to speed up machine learning modeling with systematic processes and pre-built evaluation measures; • to support popular machine learning and data processing libraries such as Spark MLLib and OpenNLP; • to implement your own machine learning models and seamlessly incorporate them into your engine; • to simplify data infrastructure management; this packaged solution can be easily deployed as a Docker container so it is compatible with On-Premises and Cloud-based installations.