Andreas Hartmann, of deep-learning provider VAIX, outlines the steps providers should take to get more personal with their users, and the options for time-to-market.
Personalisation has been around since the ‘90s, and over the last 20 years, it has received much attention with sensationalist headlines in the press, with each new development declaring itself as the perfect solution for serving and targeting users. On a second look, however, personalisation wasn’t the only reason for the hype, but rather those technologies serving to enable personalisation and allow companies to get closer to their users. Of course, the hype surrounding new ideas ebbs and flows. In some cases, the promises prove too good to be true and the technology is ultimately unable to fulfil its promises. In other cases, the technology does deliver, and becomes part of the regular fabric of how internet businesses interact with their users, so naturally is not considered remarkable for very long. Even in gaming and sportsbetting, personalisation is not a new concept. Over the last decade several applications have launched with the intent to ‘get more personal’ with users. Some, like ‘recently played’, flourished and are now a default part of casino operators’ lobbies. Others have received more of a mixed reception, such as the Favourite feature; it relies on user cooperation which players may or may not give.
Since around 2016/2017, reasonably well-informed internet users and Netflix binge-watchers have encountered the next big thing, a new technology which seems to truly deliver on personalisation. From the less enjoyable retargeting of e-commerce sites serving you a banner from a product you looked at a few days ago, to the likes of Netflix’s recommendations and Spotify’s Weekly Discovery feature, suggesting new movies or songs which are amazingly in line with our overall tastes. The technology facilitating this? Deep Learning (DL). And it doesn’t only work in e-commerce and entertainment. Having applied DL to gaming for three years now, before anyone else, we have seen it significantly outperform other AI solutions, as well as work across many more use cases than other technologies did before. DL is responsible for helping to identify the next best bet for a user, no matter if on a busy Premier League Saturday or boring Tuesday, make market recommendations, together with the bonus or free bet the user will most likely react to, and highlight similar events to the one a user recently placed a bet upon. Sports betting, with its increasingly deep and constantly changing content offering, but absent of an effective method of content curation, has so far only been increasing the paradox of choice and making navigation difficult.
Being able to easily discover personal and deeply relevant markets to every user will reap huge benefits, for players and operators. The following is a quick ‘cookbook’ for sports operators, based on our experiences over the last few years. The web defines personalisation as “customising the content of an offering based on the user’s history, preference and explicit instructions”. A broad definition, touching the entire user experience, and requiring multiple data points fed to provide real personalisation of the following: the user’s betting history, past activities compared to other punters, the relevance of sports or leagues similar in nature, and more. These are all dimensions which no normal AI algorithm, and only DL, is able to process.
Step 1: a mundane, but important, preparation is to keep in mind the business challenges which are most important to tackle, and which Deep Learning should help solve: reduce churn on the first active user day, increase activity of regulars, or improve reactivation of dormant users. Your business challenges will help validate the chosen part of the user journey you want to improve using DL, and to define the model architecture and data output. Always start with the specific business problem, not with AI.
Step 2: describe the use case, defining it as narrowly as possible to focus solely on the main business challenge. This makes it measurable, enabling you to show the value of the solution quickly. Make it as simple as possible to build, deploy and measure. Example: to increase reactivation of dormant users, invite them to punt on the matches they really care about.
Step 3: to map it back to your business goals, define how to measure success. For activity and engagement goals, metrics like # wagers, repeat visit frequency or detection rate of new sports or markets are examples.
Step 4: guided by the above, define a clear objective or problem statement. For example, in alignment with the business goal to maximise business impact, the optimisation challenge for the AI could be ‘predict next event to bet on’. For a more sustainable goal, the objective could be ‘recommend bets to maximise lifetime or repeat visits’.
Step 5: define the input to the AI model. What is the data available to train and test the AI? Over the years, VAIX has learned to overcome the challenges of consuming and processing any type of raw or processed data to make is usable by the AI model, as well as find the right mix to not overfit.
Step 6: define the output. What should be the response of the AI? What exactly is needed? A list of markets by confidence for a user ID? A list of the next relevant bets to follow after a specified event? Predicted player wagering for an event? The model output should instantly be usable for a defined follow-on action, like sending an email, push notification or display in the UI.
Step 7: what is the output format to fit into your operations. A recommended bet ID, directly written to your CRM system, allows instant use by the promo tool or marketing automation system, connecting trigger, promised action and business value. Having a partner handle all those integrations and integration formats, from push to pull, saves you time and budget.
Finally, there comes the decision to build or buy. Both options have their pros and cons. The biggest advantage of the ‘build’ option is control and creating expertise in-house. However, this happens at a cost. Aside from resource and infrastructural cost outweighing an external solution, the biggest factors are opportunity costs: failing to deliver other company initiatives around a new market or regulation, lost time to market caused by ramping up the team internally (from those busy BI and backend dev teams) and managing constant distractions with other priorities. And finally, delay due to learning and experimenting to get to a first model. All compelling reasons to first give an external solution a chance. The biggest advantage, however, is to approach AI as a pure business challenge, leaving tech challenges and heavy lifting to the partner. From data pipeline over extensive model iteration, to performance tuning to take the traffic.
With years of experience, VAIX’s DL-powered personalisation platform Rightbet, outperforms both rules-based systems and conventional AI recommenders by more than two times, showing how effective DL can be. It already supports a dozen use cases, from deeply personalised recommendations for events & markets, pre-match and in-game, favourites & similar events, teams and more, both personal and location-based. Our strong belief is that in two to three years, companies will offer AI and personalisation services like Mailchimp or Drift are offering email services today. Operators will choose whichever service offers best quality and value. This is already happening in other industries, and is on the way to online gaming. Getting ready with a simple checklist and trying your first specific use cases with a trusted partner will be a good start.