Squeezing Out the Last 5% of Conversion Rate with AI

Online sales and marketing generate enormous volumes of data that reveal a great deal about user behavior and preferences. Putting that data to work is Acceleraid's core mission. To do this, we use a range of machine learning algorithms, from Markov models to neural networks.Today, AI, big data and machine learning are used mainly by IT departments, business intelligence and data science teams. Our focus is on making complex technology accessible to everyone. Through AI-based data analysis and personalized customer engagement, we support your sales activities and help you increase your deal closures and revenue.

To do this, our software analyzes your customer and visitor data, segments target audiences, automates the engagement process, and optimizes your channels based on all click and transaction data using machine learning. Our algorithms are flexible and versatile. We help you tune them to your specific business case.

Contextual Bayesian Bandit

DATA-DRIVEN CONTENT PERSONALIZATION

Not every website design or email performs equally well for every customer. We optimize user engagement by showing each visitor a variant tailored specifically to them. Acceleraid offers several options for optimizing this delivery: optimization can be based on a neural network or on a parameter-based Bayesian bandit.

In typical A/B testing scenarios used by similar tools on the market, different variants are tested against each other, and a preferred variant is chosen once statistically significant differences emerge. Our machine learning algorithms work iteratively and converge much faster toward the best possible traffic distribution.

HOW DOES IT WORK? (USING THE BAYESIAN BANDIT AS AN EXAMPLE)

The underlying mathematical problem is known as a "multi-armed bandit" problem. Here, there are multiple triggers (~variants) with different, unknown win probabilities (~conversion rates), and the goal is to pull them in a way that maximizes total payoff.

One of our solutions to this problem is an advanced version of the contextual Bayesian bandit. It calculates the probability densities of the real conversion rates for different variants based on data collected so far. For every incoming user, a conversion rate is drawn from these probability densities, and the best-performing variant is served.

WHAT'S THE BENEFIT?

This algorithm converges much faster toward an optimal traffic distribution and also makes it possible to optimize for user context — meaning individual probability densities are estimated for every context parameter. The algorithm then automatically and independently recognizes, for example, that smartphone users from northern Germany prefer a different website variant on Mondays than desktop users do on Fridays, and serves each visitor the design, product, or messaging optimized for them accordingly.

Trends and shifts in user preferences over time can also be detected through dynamic weighting of the data within these densities and used for optimization. Our contextual Bayesian bandit is already deployed at many companies, delivering optimization gains of between 10% and 30%.

Multi-armed Bayesian bandits estimate user behavior through probability distributions

Requires no training data or statistical significance — learns from day one

Processes all parameters to estimate conversion

Uses all data for learning and optimization, and detects trends

By comparison, SCORING optimizes for the best-performing variant per individual user within a text

Requires training data to calculate scores

Focuses on the parameters that are especially relevant

Uses training data from a predefined time window

With NEURAL NETWORKS, learning happens by training an abstract network

Requires training data to optimize the network

Processes all parameters to train the network

Uses training data from a predefined time window

Clustering with HDBSCAN

To make sure your messages reach the right audience, we apply various automated clustering methods to your customer data. Using a density-based method built on HDBSCAN and fuzzy clustering, we generate meaningful, usable customer segments from your customer data.

All of these customer segments can be made directly usable for personalization through our system, delivering insights into your customer base.

Process Automation

Intelligent process optimization for large data volumes saves time and money. Automated optimization of business processes, image and text recognition, churn detection, and sentiment analysis — there are countless applications for machine learning.Toward hyperautomation — we know how to train artificial intelligence: with Acceleraid's software, you can launch rule-based automated email campaigns or have the Bayesian bandit optimize them automatically. The Bayesian bandit continuously and independently learns which email variant and content achieves the highest conversion rate for which customers.