The Proof of the Pudding …

Everyone knows that the real value of something can only be judged from practical experience and results; not from appearance or theory. And that’s exactly how one should measure any AI implementation – from the outcome of its application.

This is essential to clearing the hype around AI and focusing on methodologies that drive pure value. We believe that AI is the way forward and to substantiate, we’ve put together a case study to showcase how AI can bring true value to your business processes.

The Client: A global retail chain which sells products, both online and offline.

The Implementation: A behavior scoring model to generate machine driven recommendations, in order to resell/upsell products and drive revenue.

The Test: After six months of implementation, we ran up a test to understand how the logic fared. We filtered out data of all machine based recommendations that were triggered by specific customers, and see if they purchased the same product as recommended within a timespan of 0-5 days.

The below chart shows the outcome of the recommender logic.

 

The Results:
A total of 184,712 users responded to the recommendation, which is broken down by the number of days.
33,358 users purchased in less than 24 hours.
32,080 users purchased within 24-48 hrs of the recommendation being sent.

The machine had recommended about 2,000+ unique products to 20,000+ unique customers, based on their behavior. It generated revenues in excess of $200,000, every day!

This is proof that AI based Recommender Systems can make a big difference in the way businesses channelize their revenues and contextually engage their customers.

About Plumb5 Recommendations
The recommendation module connects to the unified data model on Plumb5, where data is primarily classified by users.  On every interaction, the algorithm creates and maintains an associated product list against each product, which gets updated after each relevant interaction, all in real-time. Read more about Plumb5 Recommender here

How does it work?

For a machine to personalize and drive the user to the desired goal path, it would need to match and deduce the next best event and target message that will drive him to the goal path. This computation needs to be done in real-time in order to personalize instantly when the user is active.

This real-time computation would require (a) appending current session behavior pattern to past behavioral pattern of the current user and (b) the deviation of the current user path to the most optimal goal paths in relation to a specific product.

With these parameters, the algorithm calculates the difference to arrive at the next closest pattern leading to the goal. As the user interacts, the algorithm appends the new action and re-adjusts to output the closest probable.

purchasey

The working of the algorithm can be visualized as per the diagram above, the pattern of user interactions is mapped from the initial event to the desired goal path. After analyzing a particular user’s interaction flow for a particular product, it can arrive at the most effective conversion pattern for a single product.

The area highlighted shows the most probable purchase paths for the given product, which is used to verify against with the current browsing pattern of the user. Now whenever the user’s pattern heads in a certain direction, the algorithm for matching patterns can quickly locate area coordinates and return the nearest event node to deduce possible pattern extensions.

These patterns are stored as strings and on every interaction, the string is updated which contains a list of associated products along with the association score, which gets updated after each relevant interaction, all in real-time.

This scoring is done after every interaction so that the updated score is available for recommendation at all times. The recommender outputs a string ([X101, P001{(P003, 0.966), (P004, 0.877), (P007, 0.746), (P002, 0.335), (P009, 0.200)}, where X101 is the user id, P001 is the product (product ID) and the product ids within the bracket are associated products with corresponding weights.

To ensure finer supervision, nested rules can be configured. For example, you can add a rule to exclude a product from the recommendation list if the product is already purchased by the user or exclude a product from the recommendation list if the user has an unresolved issue. This would allow marketers to fine tune their recommendation intelligence and make machine driven personalization more human. 

Advertisements

Customer Engagement Using AI

It could be said that in the near future, customer engagement which encompasses product marketing, query handling, customer support can be handled by machines without much human intervention.

Customer Marketing, Sales, and Support, which is mostly built on communication, is always a dialogue between a particular customer and the brand. This human dialogue (interactions) is mostly about seeking information about your brand, or negotiations before purchase or after sales queries or issues, managed by support teams. Continue reading →