Understanding real-time personalization
As the world goes digital, we are seeing more and more digital implementations of customer decision making and personalization. Think about product/service recommendation, search personalization, digital marketing, problem identification and resolution, etc. These should now be real-time, based on client activity during the web/application session and short-term history and client attributes. This requires real-time data processing and in-session interventions alongside batch information.
To learn more about real-time personalization, Analytics India Magazine caught up with Aniket Koyande, Senior Data Scientist at Fractal.
AIM: How did you start your data science journey?
Aniket Koyande: I have worked in the data science industry for eight years; started my career at Fractal in 2014. However, I pursued my B.Tech.-M.Tech. in civil engineering. During my master’s, my thesis involved a lot of statistical modeling – because of that, I ended up taking courses in probability, statistics, and modeling and learned Python programming. This was my entry into the world of analytics and data science. Many exciting opportunities have presented themselves to me since joining Fractal. I started as an analyst at Fractal, and now I’m Lead Data Scientist here in the AIML practice. My job is to provide leadership from a data science perspective for different client engagements across multiple domains, such as Financial Services, GIC, Insurance, Healthcare, etc. Along with that, I’m a core team member of Customer Genomics, Fractal’s flagship solution for Customer Decision and Next Best Action.
AIM: What is real-time personalization?
Aniket Koyande: Personalization has been around for many years. It started with something as simple as mentioning people’s names in emails. Companies then categorize customers into rule-based segments and contextualize the offers they send to different segments. Then came machine learning, which gave marketers a whole new edge by enabling prediction at an individual customer level.
Over the past few years, the world has digitized at a rapid pace; this was further aggravated by the pandemic situation. For example, people are buying more things online instead of going to stores. Therefore, it becomes imperative for organizations to ensure that they deliver the same level of experience across online channels as they would in their physical stores or branches. This is what drives the need for real-time personalization. The basic idea is to pick up signals during an active customer session on a website or app and use them to deliver a personalized experience to the customer during the session.
AIM: How has personalization been spurred by the development and advancements in machine learning?
Aniket Koyande: From a machine learning perspective, we have models that can learn any nonlinear abstract model from the data, which traditional models may not be able to do. Modern deep learning models can identify structured and unstructured data patterns. These models are data-intensive and need a lot of data to learn these complex patterns, but they are able to do so with minimal intervention.
The second thing is that these architectures also allow for incremental learning. This means that you can put these systems almost on autopilot; they can learn incrementally over time and adapt to any changes that occur in the data. Apart from supervised models, there are other neural network models like auto-encoders, W2V etc. which make our life easier by providing a level of abstraction in case of high dimensionality issues like converting a large number of products or web URLs to n-dimensional vector embeddings.
Cloud computing was also a very important factor that enabled it; we now have an infrastructure that can be scaled up or down on demand at scale, and it can be done quickly. In the real-time context, computational resources are very important, considering the fact that a company has a window of only a few seconds to react within the client session; otherwise, they run the risk of losing a sale or not delivering the right customer experience. Since we talked about using a lot of data to train the models, technological advancements in terms of GPUs have been very important to save time, especially in the case of deep learning models.
OBJECTIVE: The amount of data companies collect can be a bit intrusive and could border on violating customer privacy. What is your opinion on this?
Aniket Koyande: There is no straight answer to this. Of course, there’s a fine line between being helpful and intrusive. And it’s a very subjective topic, which means different people have different attitudes about making their data available to businesses.
I think there are a few important things that organizations should be looking for. One is – compliance. We have laws like GDPR and businesses and individuals need to comply with them.
From a customer’s perspective, most people are not completely aware of how their data is used and if it is shared with third parties. Two things are very important here. One is awareness from the perspective of customers and transparency from the perspective of organizations (in terms of how data is used and how customers would benefit from it). And the second very important thing next to that is the customer’s consent, which gives them a sense of control over what information they want to share and what they don’t want to share. Companies should view customer data as an “asset”: if they share it with you, it’s only right that you give something back in terms of experience or monetary value.
The right mix of artificial intelligence and human intelligence is needed because humans understand what’s sensitive and what’s not and can add rules and exclusions to override AI recommendations. It’s also important to communicate with clients to make sure they feel comfortable and not insecure.
AIM: What’s the next best real-time action?
Aniket Koyande: Next Best Action (NBA) is the process of predicting what is the next ideal action or intervention a business should take at any point in a customer’s journey. Actions can relate to recommending a product, providing an offer, responding to risk, and more. Typical NBA models are run in batch mode at a fixed frequency, such as daily, weekly, monthly, or ad hoc. On the other hand, real-time NBA is all about identifying the ideal next action at any time during a customer’s active website/app session while the customer is browsing. Intervention can be based on activity during the session (as well as short-term history and client attributes). This is valid for authenticated and unauthenticated clients.
The most important thing in this context is to predict the intention. You can ask questions such as: what is the customer actually trying to do? Are they looking for a specific product? Do they need help? Predicting the customer’s intention or need becomes important in order to be able to take the right action. This requires real-time processing of clickstream data from the client’s session, converting it into signals for patterns, scoring the patterns, and returning the next best action to the web server for execution. All of this should be done within 10-20 seconds. Real-time NBA may include use cases such as recommending products/services, predicting intent and providing support, avoidance of dropouts, personalized offers and content, and personalization search (for easy navigation). Of course, there may be scenarios where you’re not sure what the customer is trying to do or if they don’t meet your criteria for any of these actions. Thus, no action may be the best possible action in such cases.
AIM: What is Fractal’s client genomics capability?
Aniket Koyande: Customer Genomics is Fractal’s flagship customer analytics and Next Best Action offering. Around 2014-15, we generated $600 million in new assets and a 26% increase in sales for one of the world’s largest asset managers through our early stage approach. We converted that into an ability over time. Additionally, we have built many modular accelerators to speed up the implementation of NBX systems for our customers.
The philosophy of the solution is to gather all possible signals available on a customer from different data sources and harmonize them. This brings an asset called Customer 360, which consolidates all customer information in one place, as well as statistical and domain-specific signals. Then we use AI models, which help us generate predictive results, leading to the next best action. Finally, the customer’s response to the action is captured in the system, which forms a feedback loop to gradually update the model with each new signal. We also have an explainability module to open black box deep learning models and understand the drivers behind decisions.
Customer Genomics has been very useful in marketing, sales and service – for example, you can create campaign lists for different actions in an automated way, which can be very easily integrated with Salesforce and all CRM tools that organizations can use. It also allows companies to get information on what to do with their customers the next time they interact with them. This has adapted very well to both the B2B and B2C context. As a result, we have been able to generate added value for our customers in many different business areas.
There’s a lot going on in terms of recent developments. We have just obtained a patent for our accepted solution. We also have our Real-time Next Best Action module in development, which should be ready by mid-April. We also recently partnered with Fractal’s Behavioral Sciences capability to create a viewpoint on integrating some of our proprietary Behavioral Science toolkits with Customer Genomics so that we can understand customers from a emotional and help them make better and easier decisions. .