Conference Agenda

Relevance & Research Question

As a way to reduce survey length, or even to replace surveys, we have been involved in passive data (web navigation) collection for a couple of years. Passive data is relevant to the description of online behaviour, but declarative data is still needed to interpret and explain behavior; as one could hope to directly infer individual attitudes from internet behavior. Due to the recent advances in natural language, such automated analysis of contents and attitudes is no longer an elusive dream. As an experiment, we have thus used BERT (Google's deep learning based language model) and further proprietary deep learning techniques in order to try and analyze opinions, based on online media consumption.

More precisely, is it possible to instantly, without asking anything, combine passive data and BERT and get a deep understanding of the audience of any website? For example, is it possible to get the specific attitude of visitors to Audi.com towards ecology? Our talk will be based on the presentation of a prototype of an online tool/dashboard. Its objective will be to share the promises and the challenges of this usage of AI.

Methods and Data

The data we use is based on 7000 respondents (from France, Germany, UK) who agreed to install a tracking software. For 347 days on average, we continuously collected the navigation data (urls visited and / or apps used) for each of them. Data is analyzed via BERT properly trained. Realtime vizualisation of the results powered by Tableau.

Results

The quality of results relies on the ability of our neutral network to accurately categorize words in a consistent semantic field. Some results are pretty impressive: without having been trained on these particular fields, “Ronaldo” is associated to football and “parenting” is related to family life. But some are disappointing: “psoriasis” is associated to medicine in general (not even to dermatology only). We will discuss these results and will try to explain them.

Added Value

It is a work in progress, at this stage, the main benefit is to present and discuss concrete applications of AI in market research.

We were wondering: If we find out, how many people have been looking for a specific thing (or basically just words) on the internet in the past, would we be able to calculate their interest in the future, too?

The use of Artificial Intelligence in combination with existing technologies has been repeatedly discussed lately. We were interested in combining AI with traditional trend research and developed Pythia, a tool which forecasts culture and consumer trends. How? By examining Google search data and other sources on new trends, evaluating and structuring them individually. The neural networks analyze huge amounts of data and are trained on search data from the past decade. The trend research tool was created to obtain insights that could be used to find new products, improve them and present them more effectively to have a positive impact on product development by using trend forecasts.

We know what you will be needing to sell and how to interact with your customer in the future.

As of today, Pythia can forecast the latest culture and consumer trends of the next 18 months with 95 percent probability in over 50 countries.

The results and experiences with cooperating companies have supported our initial goal to successfully AI with traditional trend research. In an early cooperation with our Co-Founder Rossmann, Pythia suggested “CBD”, "Ingwer Shots" and many more trending topics in Germany. CBD products have been strong performers in their online shop ever since. The tool also proved to be useful for enhancing polls: 10 days prior to the election of the SPD federal chairman, Pythia predicted the correct result.

Want to know what's going to happen within your business? Ask Pythia.

When we talk about predictive analytics, we should not leave aside a method that has been around for what feels like ages (i.e. at times when the term predictive analytics was not even born yet...), but whose predictive power is still one of the best that the market research toolbox has to offer: conjoint analysis. Its value can be seen simply from the fact that it is still one of the most relevant methods of price and product research and is used globally.

In contrast to what is commonly known as predictive analytics, however, conjoint is not based on existing data, but on data collected in decision-making experiments within the framework of surveys.

As an indirect method, it is free of inflation of pretensions and scale effects, and as a reflection of a real decision situation, it is also able to cover behavioural economics effects.

If we assume that there are essentially three variants of predictive analytics (predictive models, descriptive models and decision models), conjoint analysis even includes all three.

But Conjoint not only helps us to develop better products, but can also help to determine the pricing strategy and improve communication and marketing.