ARTIFICIAL INTELLIGENCE
to work with. It is often the case that, during the design stage of the model, we realize that we don’t have enough features in our data. In this case we look to capture additional data in the future. We saw this in Australia with the use of data to detect possible Money Laundering. In this case the introduction of mandatory carded play is being implemented to attach player features/attributes to gaming machine transactions. This increases the number data dimensions thus helping to improve the accuracy AI models. AI Models are generally created in software which needs to be trained and adjusted. Training occurs by feeding lots of data through the model. Algorithms work on a variety of reliable and well understood mathematical equations some dating back to the 1700’s based on probability statistics in which coefficients and parameters are created by executing multiple iterations until the optimal result is achieved. Once the AI model is developed, it may perform well on training data. However, it’s not known at this point how it will perform on unknown and new future data. In fact, if the model performs too well on training data, it may perform terribly on new data, an issue known as overfitting. To evaluate the effectiveness and efficacy of the model, the test data is used essentially as new and unseen data to the model and objective assessments are made using standard statistical formulas. Predictive models, as their name suggests, predict future behavior based on current known features. For instance, a model may identify players likely to display risky behaviors and to develop problem gambling issues in the future. The success of such a model very much depends on the volume and quality of the available data and the desired feature to be identified. The “labelled” data for each player which identifies them as likely to go on to exhibit risky behavior or not as it relates to gaming, must be defined for the AI model to learn. As mentioned earlier, the model can be tested and scored to determine accuracy, recall and precision. In addition to predicting future behavior, AI models can also provide valuable insights into the elements that make up the major factors in decision making when it comes to how players are managed. Whilst, for some, this will conjure up images of Minority Report, it is not as futuristic or alarming as it sounds. It is not unusual to find fewer than 10 factors can contribute to more than 60 percent of the decision. These may be things such as time on a device or the time of day. AI can support or challenge long held beliefs as to which of these factors are most
relevant. Taken together, the factors can provide an accurate basis for an initial classification of a minotiry of players as potentially problematic gamblers, but it does not give the full picture. Human intervention, comes in at this point but AI has done the job of processing the vast amounts of data needed to identify intervene with the minority. Classification models will become increasingly important in the gaming industry and AI models will categorize players based on their transactions and other features. Whilst this is certainly not new, Casinos, Airlines and other industries currently employ simple models such as loyalty tier structures which yield a much more one-dimensional picture . AI models will continue to evolve through more complex data structures associating players with transactions through digital wallets, player accounts and biometrics. It is not possible for humans to visualize or categorize a player based on any more than three or maybe four dimensions, AI can however do this on incredibly large scale with complex data structures and detect patterns that can simply not be observed by humans. Like the Prediction models previously discussed, it is possible to develop an AI Classification model that is well understood as to the features used in categorization and the AI model can be evaluated and scored. Classification models are extremely useful when it comes to identifying fraud and money laundering and in clustering customers for targeted marketing campaigns. Categorization models are used to recommend similar games within the gaming environment whether it be online, or land based in a similar way to AI algorithms are used by streaming services and on-line shopping. Tread with caution The process of building an AI Model provides valuable insight into the key data features that contribute to accurate and effective prediction or classification. For instance, through the training process, it may be determined that time on device or average bet amount may contain the greatest insight or Information Gain and thus have a high correlation with the predicted value, such as money laundering. This may be an error. It is an easy trap to fall into to conflate correlation with causality and although useful it is not enough to show a conclusive link between correlation and cause. A high correlation between bet size and money laundering does not support a policy of reducing bet sizes to reduce money laundering. A better way to understand this is the example of
PAGE 20
IMGL MAGAZINE | JANUARY 2024
Made with FlippingBook flipbook maker