{"id":6848,"date":"2023-07-28T21:49:30","date_gmt":"2023-07-28T19:49:30","guid":{"rendered":"https:\/\/eurocc.nscc.sk\/?p=6848"},"modified":"2023-08-02T14:28:49","modified_gmt":"2023-08-02T12:28:49","slug":"detekcia-anomalii-v-casovych-radoch-prevencia-gamblingu-pomocou-hlbokeho-ucenia","status":"publish","type":"post","link":"https:\/\/eurocc.nscc.sk\/en\/detekcia-anomalii-v-casovych-radoch-prevencia-gamblingu-pomocou-hlbokeho-ucenia\/","title":{"rendered":"Anomaly Detection in Time Series Data:\nGambling prevention using Deep Learning"},"content":{"rendered":"
\n
\n
\n
\n

Anomaly Detection in Time Series Data:\nGambling prevention using Deep Learning<\/strong><\/p>\n\n\n\n

<\/p>\n<\/div><\/div>\n\n\n\n

Gambling prevention of online casino players is a challenging ambition with positive impacts both on player\u2019s\nwell-being, and for casino providers aiming for responsible gambling. To facilitate this, we propose an\nunsupervised deep learning method with an objective to identify players showing signs of problem gambling\nbased on available data in a form of time series. We compare the transformer-based autoencoder architecture for\nanomaly detection proposed by us with recurrent neural network and convolutional neural network autoencoder\narchitectures and highlight its advantages. Due to the fact that the players\u2019 clinical diagnosis was not part of\nthe data at hand, we evaluated the outcome of our study by analyzing correlation of anomaly scores obtained\nfrom the autoencoder and several proxy indicators associated with the problem gambling reported in the\nliterature.<\/p>\n\n\n\n

<\/p>\n<\/div>\n\n\n\n

\n
\"\"<\/a>
illustrative image<\/figcaption><\/figure>\n<\/div>\n<\/div>\n\n\n\n

<\/p>\n\n\n\n

Gambling prevention of players with problem\nor pathological gambling, currently conceptualized\nas a behavioural pattern where individuals\nstake an object of value (typically money) on the\nuncertain prospect of a larger reward [1], [2], is\nof high societal importance. Research over the\npast decade has revealed multiple similarities between\npathological gambling and the substance\nuse disorders [3]. With the high accessibility of\nthe Internet, the incidence of pathological gambling\nhas increased. This disorder can result\nin significant negative consequences for the affected\nindividual and his\/her family too. Therefore\ndetecting early warning signs of problem\ngambling is crucial for maintaining player\u2019s wellbeing. This work is a joint effort of Slovak National\nCompetence Center for High-performance\nComputing, DOXXbet, ltd. \u2013 sports betting\nand online casino, and Codium, ltd. \u2013 software\ndeveloper of the DOXXbet sports betting and\niGaming platform, with the goal to enhance\ncustomer service and players\u2019 engagement\nvia identification and prevention of gambling\nbehaviour. This proof of concept is a foundation for future tools, which will help casino mitigate\nnegative consequences for players, even for a\nprice of less provision for the provider, as in\nline with European trends in risk management\nrelated to problem gambling.<\/p>\n\n\n\n

In our study we propose a completely\nunsupervised deep learning approach using\ntransformer-based AE architecture to detect\nanomalies in the dataset - players with anomalous\nbehaviour. The dataset at hand does not\ncomprehend the clinical diagnosis, and amongst\nother proxy indicators mentioned before only\nfew are available - requests to increase spending\nlimits, chasing losses by gambling more (referred\nto as chasing episodes later in this article), usage\nof multiple payment methods, frequent withdrawals\nof small amount of money and other\nmentioned later in the text. Clearly, not all the\nanomalous users must necessarily have problem\ngambling, hence the proxy indicators are\nused in combination with AE results, namely the\nanomaly score. The foundation of our approach\nrests on the idea that a compulsive gambler is an\nanomaly within the active casino players, with\nthe literature mentioning their fraction amongst\nall players being between 0.5% to 5% for chancebased\ngames.<\/p>\n\n\n\n

<\/p>\n\n\n\n

Data<\/h2>\n\n\n\n

The data acquired for this research consist of sequences\nof data points collected over time, tracking\nmultiple aspects of player\u2019s behaviour such\nas frequency and timing of their gaming activities,\nfrequency and amount of cash deposits,\npayment methods used when depositing cash,\ninformation about the bets, wins, losses, withdrawals\nand requests for change of deposit limit.\nFeature engineering resulted in 19 features in\na form of time series (TS), so that each feature consists of multiple time stamps. These features\ncan be classified into three categories - \u201dtime\u201d,\n\u201dmoney\u201d and \u201ddespair\u201d, as inspired by Seth et\nal. [7]. Table 1 summarizes the full set of TS\nfeatures with a short explanation.\nEach feature is a sequence of N values, where\neach value stands for one out of N consecutive\ntime windows. This value was produced by\naggregating daily data in the respective time\nwindow, with the time window length being\nspecified in the Table 1 together with the information\nabout the time window being sliding or\nnot. Hence, for each sample we needed a history\nof N time windows. Feature engineering procedure\nis displayed in Figure 1 and the final data\nshape is depicted in Figure 2.<\/p>\n\n\n\n

<\/p>\n\n\n

\n
\"\"<\/a>
Figure 1: Visualization of the data aggregation from daily basis into time windows, and eventually to TS features.\nt1<\/sub>, …, t450<\/sub> represent time stamps for daily data x1, ..., x450. Daily data points from a time window are aggregated into a\nsingle value zi for all i \u2208 (1, . . . , 8).1<\/sub>, …, x450<\/sub>. Denn\u00e9 z\u00e1znamy z \u010dasov\u00e9ho okna s\u00fa agregovan\u00e9 do jednej hodnoty zi<\/sub> pre v\u0161etky i \u2208 (1, . . . , 8).<\/figcaption><\/figure><\/div>\n\n
\n
\"\"<\/a>
Figure 2: Final data shape obtained after feature engineering.\nEach sample is represented by 19 features consisting\nof 8 time windows.<\/figcaption><\/figure><\/div>\n\n\n

AE models comparison<\/strong> <\/h2>\n\n\n\n

Autoencoder is a \"self-supervised\" deep learning method suitable for anomaly detection in the Czech Republic. The idea behind using this type of neural network for anomaly detection is based on the model's reconstruction capability. AE learns to reconstruct the data in the training set and since the training set should ideally only contain \"normal\" observations, the model learns to reconstruct only such observations correctly. Therefore, when the input observation is anomalous, the trained AE model cannot reconstruct this input sufficiently correctly, resulting in a high reconstruction error. This reconstruction error can be used as an anomaly score for the given observation, where a higher score means a higher probability that the observation deviates from the general trend.<\/p>\n\n\n\n

In the study, we trained an AE model based on transformers, where both the encoder and decoder contain a layer called \"Multi Head Attention\" with four \"heads\" and 32-dimensional key and value vectors. This layer is followed by a classical neural network with so-called \"dropout\" layers and residual connections. The entire AE model has just over 100k trainable parameters.<\/p>\n\n\n\n

Reconstruction loss and Prediction ability<\/strong><\/h2>\n\n\n\n

We performed a 3-fold\ncross-validation by splitting the data into training,\nvalidation, and test sets, and trained the\nmodels for each split to assess their stability. Resulting\naverage loss values and their variances\nare displayed in the Table 3. The average reconstruction\nerror of Transformer model is significantly\nlower than all the other models. LSTM B\nmodel comes second in the reconstruction performance\nand CNN model seems to have the\nworst prediction performance. Generally, the\ntest loss is observed to be always higher than\ntrain and validation losses. The reason for this\nis that those 211 data points that were removed\nfrom the training set in the data cleaning process,\nwere moved to the test set. Without moving these samples, the test loss for transformer-based\nmodel would be as low as 0.012, for CNN model\n0.33, for LSTM A model 0.27, and for LSTM B\nmodel 0.13. More detailed overview of the models\u2019 performance\nis displayed on the Figure 6 as histograms\nof loss values of the test set. All histograms have\nheavy right tail, which is expected for datasets\ncontaining anomalies.<\/p>\n\n\n

\n
\"\"<\/a>
Figure\u00a03: Reconstruction error histograms of the transformer-based AE model for the test set. On the x-axis is the value of the anomaly score and on the y-axis is the frequency of the corresponding value.<\/figcaption><\/figure><\/div>\n\n\n

To demonstrate the quality of the CR reconstruction, the original (blue line) and predicted (red line) values for a randomly selected anomalous observation of one player are shown in Figure 4. The value of the anomaly score for the respective models is given in the caption of the graphs.<\/p>\n\n\n

\n
\"\"<\/a>
Figure\u00a04: Comparison of the predictive ability of AE models. All models reconstructed the same observation coming from the test set. Predictive ability: the blue line represents the input data, the red reconstruction obtained using the transformer-based AE model. The number shown in the graph header represents the anomaly score for that data sample.<\/figcaption><\/figure><\/div>\n\n\n

Results<\/strong><\/h2>\n\n\n\n

Since clinical diagnosis was not part of the data we had, we can only rely on auxiliary indicators to identify players with potentially problem gambling. We approached this task by detecting anomalies in the data, but we are aware that not all anomalies necessarily indicate a gambling problem. Therefore, we will correlate the results of the AE model with the following auxiliary indicators:<\/p>\n\n\n\n