Building a Real-Time Player Bonus Reward System Using Neural Networks

Introduction

In the competitive world of online gambling, player retention and engagement are critical for business success. One effective strategy to achieve this is through personalised bonus rewards. Leveraging machine learning to create a real-time player bonus reward system can significantly enhance user experience and increase loyalty. This article provides a comprehensive guide on how to build such a system using TensorFlow, including data preparation, model training, and real-time deployment.

Compliance and Responsible Gambling

Before diving into the technical details of the problem, please understand that this article explains how to build a neural network to recommend rewards for players, but before to translate this into a real application, we must ensure that the bonus reward system complies with relevant gambling regulations, such as adhering to responsible gambling guidelines. We have to make sure that our dataset incorporates features that promote responsible gambling, so to prevent excessive gambling. Additionally, we have to ensure fairness and transparency in our model to avoid biased outcomes and prevent the creation of addictive behaviors. Last but not least, providing players with the option to opt-out of receiving real-time bonuses is also crucial.

Dataset and Synthetic Data Generation

Since there isn't a publicly available open-source dataset for this problem, we will build a data generator to simulate player behaviors. This generator will incorporate features related to player engagement:

• total_play_time: The total amount of time (in seconds) that a player has spent playing games on the platform. This indicates how engaged the player is with the platform.
• total_bets: The total number of bets placed by the player. This can reflect the player's activity level.
• avg_bet_size: The average size of the bets placed by the player (in currency). This shows the player's betting behavior.
• total_deposits: The total amount of money (in currency) deposited by the player into its account.
• withdrawal_frequency: The average time interval (in seconds) between the player's withdrawals. This can indicate how frequently the player takes money out, reflecting their cash-out behavior.
• game_genre: The primary genre of games the player engages with, encoded as integers (0: Slots, 1: Poker, 2: Blackjack, 3: Roulette). This shows the player's game preference.
• player_age: The age of the player. Age can influence the player's gaming and spending habits.
• gender: The gender of the player, encoded as integers (0: Male, 1: Female). This helps understand the demographic profile of the players.
• received_bonus: A binary indicator (0: No, 1: Yes) of whether the player has received a bonus. This is the target variable we are trying to predict, showing whether the player has been rewarded with a bonus.

To generate synthetic data that looks more realistic and can improve the performance of the model, we need to consider some realistic patterns and distributions in the data. For example:

• Total Play Time: This can follow a normal distribution with a mean around typical playing sessions.
• Total Bets: This can be higher for frequent players.
• Average Bet Size: This could follow a log-normal distribution.
• Total Deposits: This could follow a normal or log-normal distribution.
• Withdrawal Frequency: This could be higher for more frequent players.
• Game Genre: This can be distributed according to popular preferences.
• Player Age: This can follow a normal distribution around an average age.
• Gender: This can be around a typical distribution (e.g., 50/50 split).
• Received Bonus: We can directly control the class balance by setting a target proportion for class 1 and then generating data accordingly. In our case, we decided to assign class 1 a low distribution percentage of 10%. Even though it makes the dataset more imbalanced, this reflects a more real-world scenario In fact the event of receiving a bonus in gambling, occurs less frequently. A model trained on a dataset that reflects this natural imbalance may perform better because it learns to identify the more common patterns more effectively.

The image below shows the distribution of features in the dataset (60.000 rows) created with the data generator.

The pie chart below shows the final distribution of classes 0 and 1.

After generating the dataset, we can also visualise the patterns to understand the distribution and relationships of features for each class. This can help in validating that the patterns for class 1 are realistic and distinct.

The Neural Network

To train our model using a neural network built with TensorFlow, we split our dataset into training and testing data, with 80% allocated for training and 20% for testing.

Please note that before to define the architecture of the neural network described below, several experiments have been done, such as more complex and deep neural networks, fine-tuning optimizer hyperparameters, and evaluating various activation functions. When aiming to find the best model, experimenting is essential in figuring out the most effective neural network architecture.

Considering the imbalanced nature of our dataset, we used the compute_class_weight function from scikit-learn to calculate class weights. The resulting class_weights dictionary, which contains the calculated weights for each class, is then passed to TensorFlow during training. These weights help to mitigate the impact of class imbalance by giving more weight to minority classes, thereby improving the model's ability to learn from imbalanced data.

The resulting architecture of our neural network consists of one input layer with a number of neurons equal to the number of input variables, followed by five hidden layers with 4, 8, 16, 32, and 8 neurons respectively, and an output layer with one neuron. We utilised the Adam optimizer function with a learning rate set to 0.001. The model has been trained for 90 epochs and a batch size of 512.

Results

After training the model with the neural network described previously, we achieved a binary accuracy of 76% and a loss of 0.4314.

When evaluating the model using test data, the performance of the model was confirmed, indicating a binary accuracy of 0.7527 (75%) and a loss of 0.4475. Please note that test data is a set of data that the model did not encounter during training. A similar accuracy (76% and 75%) on both the training and test datasets indicates that the model is performing reasonably well and generalising appropriately.

In other experiments, we also achieved an accuracy of 90%, though with a more complex and deeper architecture for the neural network, which raised also concerns about potential overfitting.

Real-Time Implementation

For real-time implementation we can use a real-time data streaming platforms such as Amazon Kinesis, which offers a robust solution, in order to process player data in real-time. With Amazon Kinesis, we can ingest and process streaming data from from the online gambling platforms, ensuring continuous data flow for analysis. Once the model has been trained, it can be deployed using a scalable and reliable platform like Amazon Sagemaker which provides scalability and reliability to handle varying workloads, moreover allowing us to seamlessly integrate our predictive models into the streaming data pipeline.To optimise costs while maintaining responsiveness, we can adopt a near real-time approach by batching transactions per player and performing periodic inferences at set intervals rather than processing each transaction individually. This approach reduces the frequency of model inference calls, minimising operational costs. By balancing real-time responsiveness with cost efficiency, we can create a scalable and cost-effective solution for deploying predictive models in streaming environments.

Conclusion

Our exploration into using machine learning to predict player bonuses in online gambling platforms has provided valuable insights into personalised player reward systems.

By leveraging supervised learning techniques and neural network, we were able to develop predictive models capable of identifying the likelihood of players to receive bonuses based on their behavior on the gambling platform.

Overall, this article highlights the potential for machine learning to enhance player engagement and satisfaction in online gambling platforms by enabling the implementation of dynamic and personalised reward systems.

If you're interested in accessing the complete Jupyter Notebook, and to understand how to build a the entire architecture using Amazon Web Services within your own AWS environment, please feel free to reach out to me!

[email protected]