Machine Learning

As customers increasingly shop in different ways (online,in-store, reserve and collect, etc.), Dixons Carphone identified a need to provide an intuitive product suggestion journey on their websites, similar to what you will find in its retail stores, where skilled sales professionals assist customers with product suggestions that might be helpful (e.g., a laptop case for a laptop, or an HDMI cable for a large-screen TV).

The vision was to be able to offer bundles to online customers that were personalized - based on what was known about the individual customer - and offer them in real-time - so as to drive incremental sales and margin to the business. The objective then was to apply mathematical modeling, data science, and machine learning techniques to increase cross-sold products on the Currys and PC World websites.

A data science team was created within the Dixons Carphone eCommerce department and we set out to build a bundling recommendation engine.

overall solution

Syntasa was engaged to plug natively into their  and within their GCP environment to synthesize behavioral data so that it could be available for analysis rightaway. It’s part of what’s called their Supercharge Attach project. When a customer arrives at the Currys PC World websites and opens a product detail page, a query is made to the recommendation API, containing the customer’s visitor ID and product ID. This recommendation API then returns personalized or non-personalized bundle recommendations, based on the customer’s history on the Currys PC World websites. Personalized recommendations are prioritized when sufficient browsing history is available – and when this is not available, natural (non personalized) recommendations are served to the customer.

To produce personalized recommendations, Syntasa built a Nearest Neighbor model to generate a neighborhood of similar customers, based on browsing behavior and products purchased together by similar customers. Additionally, the Dixons Carphone team built an algorithm they call Natural Attach to produce non-personalized recommendations. The model was then productionized with Syntasa’s Composer module.

Challenge no.1

Increasing conversion rates for bundles and basket attachments.

Solution - Personalised items

Gain the flexibility to apply ML, value, and bespoke merchandising models for product recommendations, without the constraints of marketing tech stacks.

Challenge no.2

Saas DMP solutions were expensive to maintain, underutilized and underperforming.

Solution - Audience

Drastically cut your advertising spend by only targeting those that are likely to purchase or expand your audiences in the most cost-effective way using first-party data and machine learning.

Challenge no.3

Currys needed a robust platform to rapidly experiment its algorithmic audiences to drive personalization and marketing.

Solution - Audience

Build first-party Unified Identity Graph across systems and channels to analyze, optimize, and decision. Syntasa’s pre-built processes and automated auditing features significantly reduce operational costs. Activate directly into multiple marketing clouds and demand side platforms, such as Google, Facebook, Adobe, Oracle, Brightroll, Tradedesk, and more.

Challenge no.4

Adobe Analytics data was very complex and difficult to view at an individual level.

Solution - Adobe Analytics Adaptor

Unlock the true potential of digital clickstream data by transforming it into intelligent actions, utilizing Adobe adapters.

results

The group is now able to combine multiple business data sources into their big data environment to build machine learning modeling with the data, as well as to pass the modeling outcomes to their activation channels (i.e., websites, optimization tools,CRM, etc.). The team has seen the following results:

2x product coverage

Product coverage, which Dixons Carphone defines as the share of product views on the website where a recommended bundle was displayed, has increased from 32% coverage with the manual bundling solution to 72% with the automated and personalized data-driven bundles.

3x Add-to-basket rate

When provided with personalized, AI-driven product recommendations, customers are 3x more likely to add the product to their basket when shopping online. For customers who hadn’t generated enough data, Dixons Carphone’s natural attach recommendations outperformed manually-created ones by 1.3x.