Taking the Fear out of Innovating With Data

A large logistics firm that we have worked with has a unique proposition for online retailers that focuses on solving the challenge of managing returns at volume. As online retailing has grown, so have returns, which have grown exponentially.

The company has many systems and processes all generating vast amounts of data. Uncovering useful management information from this data however, is a major challenge. Additionally, there was uncertainty as to whether the data could provide any meaningful insights, or if the systems and infrastructure in place, would be able to provide what was needed.

They faced the common question that many large businesses have; is there any value in the data they hold and if so, is it worth investing in unlocking its potential?

As there were numerous questions that needed answering, a data feasibility study was conducted focusing on one customer, where the technology was at a suitable level of maturity and advancement.

To truly understand and appreciate the context and scale of the operations, a site visit was conducted. The site visit gave a deeper understanding of the processes and the environment the people were working in. The most impressionable aspect of the site visit, was appreciating the vast scale of the operations and how even a marginal improvement in day to day activities, would translate into a tangible gain for the business.

Now there was an appreciation of the context, we started with the questions; what is it that the business would benefit from having answers to. In this feasibility study, it was all focused on returns data and what value could be provided back to the online retailers.

Once the questions had been defined, a sample set of raw data was extracted from the warehouse management systems (WMS), so that the meaning of the data could be understood and analysed. The purpose, was to establish if the data set could answer the questions posed and also if any patterns could be identified that would provide even more value to the business.

Scripts were created that processed the data in HDInsights, Microsoft's big data platform, to produce charts and data extracts that could be plotted and visualised. Once it was identified that patterns were emerging that could answer the posed questions, a full data extract was taken and processed. Using HDInsights within Azure, the processing power could be increased for a short period, to process the tens of millions of rows of data, from the full extract.

Although the data from the WMS captured a lot of data, it was missing some vital detailed information that would unlock huge potential. The data could only categorise returned items at a high level such as womenswear or electrical. There is no easy way to retrieve this missing data, which is also compounded by the fact that the well known retailer rotates SKU's, making it even harder to find.

The research was conducted on the sample data that produced software that used a combination of strategies and matching algorithms to capture this additional metadata from across the web. Blending this additional data set with the data from the WMS, provided insight that wasn’t previously possible, opening up the opportunity for greater insight and predictability.

The data feasibility study provided the business with the confidence that the data would be able to provide the insight they needed, removing the risk around investing heavily into a data platform that may not be able to provide a return.

In addition to the reduction in risk, the project identified aspects of the operation, where machine learning could drastically improve the efficiencies of the business. This was primarily focused on the first touch bench where all returns are handled, processed and decisions are made on what happens next to the item. This step in the process is the most crucial element and is also the most inefficient stage that requires the most amount of training. It is one of the most expensive steps in the process.

It was identified that machine learning could be placed on the first touch bench to predict the outcome of the inspection. This prediction would be used in conjunction with the human interaction to determine the next step more efficiently. Using historical information, along with computer vision, images could be taken of the item as it is inspected and the computer would aid in the decision making.

The machine learning would flatten out the cognitive influences, depletion and discrepancies between personal opinions, speed up the decision making and also provide real-time training.