Machine learning model management: integrated modeling

Translator | Cui Hao

Reviewer | Sun Shujuan

Opening

##Machine learning is applied by enterprises to different Business scenarios solve different business problems. With the widespread application of machine learning, organizations are overwhelmed when choosing learning methods.

Many organizations use advanced and classic learning methods in the application of machine learning. There is the familiar dichotomy of supervised and unsupervised learning, as well as emerging variants of machine learning such as contrastive learning, reinforcement learning, and self-supervised learning.

In addition, there are graph analysis, deep neural networks, segmentation, behavioral analysis and other techniques involved. When faced with large-scale, complex business problems—such as strengthening anti-money laundering measures to combat financial crime—how do organizations decide which machine learning methods to use?

Using integrated modeling makes this problem less important. This machine learning approach enables organizations to leverage various models and combine them with predictive accuracy to achieve optimal results.

This approach helps provide full context for high-dimensional data in financial services, fraud detection, and cybersecurity. Organizations that use integrated modeling say that "integrated modeling allows for more diversity in model building," and Martin Rehak, CEO of Resistant AI, admits, "We don't want a single model to stand out."

Diversity in using models Enables organizations to use different algorithms to evaluate various aspects of a business problem in order to adopt fully informed, consistent decision-making methods - which are explainable.

Consensus-based model decision-making

The principles of integrated modeling mentioned earlier are unquestionable. Data scientists do not need to spend a lot of time designing perfect models for business cases. They only need to combine those Imperfect models combine to produce predictive power. “When you look at machine learning in an ensemble approach, you make decisions from small algorithms,” Rehak noted. “And, in our case, these algorithms are dynamically combined for each trade in order to make the best decision.” What’s more, perhaps each of these models could specialize in a certain vertical, e.g. Identifying money laundering incidents.

For example, one model focuses solely on the size of the transaction. Another model focuses on the location of transactions. Different models can examine which specific participants were involved in the transaction. The goal is a situation where "there aren't any spikes," Rehak explains. "The distribution of the model is very flat, and the evidence corresponding to the model is relatively weak. By combining many weak evidence elements, you can make a stronger decision." Another benefit is that through classic machine learning and simpler models , less training data (and annotations) is required to put the model into production. Such models are easier to interpret than deep neural networks, which require large amounts of training data.

Contextual Modeling

It is important to distinguish the distribution-flat modeling approach described by Rehak from other ensemble modeling techniques. The most common examples of ensemble modeling involve bagging or boosting (the latter may require Xtreme Gradient Boosting). Random Forest is an example of boosting based on a combination of different decision trees. With this approach, "you build the collection one by one based on the previous versions in the collection," Rehak comments. Although it is a quick way to build models with high predictive accuracy, it runs the risk of overfitting (the model becomes less applicable to production data because the training data set is too small).

Rehak’s integrated approach is more suitable for AML use cases because it is based on the context that affects these events. "If you ask a money laundering expert whether a transaction was malicious, they first look at the history of the account and how the person has behaved in the past," Rehak said. Through his approach, factors related to geographical location, time of day, interested parties and financial institutions are examined using separate machine learning models. Only by combining the results of each of these models can the AI ​​system determine whether a criminal transaction has occurred, with significantly fewer false positives. “Machine learning can explain most outliers that would otherwise overwhelm anti-money laundering teams,” Rehak said.

Decision Boundary

When integrating use cases for modeling, it is common to use more than 60 models to model different aspects of the analysis transaction. The real-time results of the integrated approach are well suited for this application scenario. “One of these 60 algorithms can split everything into segments and then model the average transaction size per second,” Rehak reveals. "We can have thousands of clips that are dynamically updated at the same time."

With the large number of models combined into the ensemble, each of which evaluates different aspects of a transaction to uncover potential criminal behavior, a more comprehensive approach cannot be created. “We look at you from so many angles that it becomes very difficult to shape your behavior while simultaneously allowing you to avoid all of these criminal acts,” Rehak revealed. “Because, in order not to be identified, a ‘criminal’ needs to avoid more than one Decision boundaries, but a large number of dynamic decision boundaries. Each model in these algorithms is learned independently, and then we combine them together."

Explainable Artificial Intelligence

There are many aspects to how these sets enhance interpretability and what they correspond to. First, they do not rely too much on advanced machine learning and only include simpler, more interpretable algorithms (involving traditional machine learning). These models became the cornerstone of assessing transactional crime. “When we say something is important, we can tell you why,” Rehak said. "We can tell you which indicators indicate this. We can write a report for each finding indicating that there is a high risk of transactional crime due to these factors." Although each algorithm focuses on characteristics, not all algorithms all have the same weight in the model. Generally speaking, algorithms involving graph analysis (which are good at examining relationships) are given greater weight than other models.

The model can not only explain suspicious behavior, but also tell you why outliers occur. “Typically we have four or five dominant algorithms in an ensemble, meaning that when I believe it’s an outlier, others agree because of the algorithm behind it,” Rehak noted. “Also, we have four or five triggers, which guarantees that the results are somewhat biased towards anomalies.” Since individual models only evaluate one factor in a transaction, they provide interpretability and word-of-score interpretability. “Because we know the set, we know the micro-segmentation, and we know the volume, we can easily display that information with questions next to the score, and volume is very important to a company’s finance department,” Rehak added.

Integrated Pattern

Ultimately, integrated modeling is used more than any one application, although it can be a huge help for AML activities. If applied correctly, this technology can improve interpretability while reducing the amount of training data and annotations required to solve business-critical problems.

Ensemble modeling leverages various data science techniques to solve multiple business problems instead of limiting the problems to one or two. As a result, this integrated problem-solving approach may become the poster child for AI deployments.

Translator Introduction

Cui Hao, 51CTO community editor and senior architect, has 18 years of software development and architecture experience and 10 years of distributed architecture experience. Formerly a technical expert at HP. He is willing to share and has written many popular technical articles with more than 600,000 reads. Author of "Principles and Practice of Distributed Architecture".

Original title: ​​​Machine Learning Model Management: Ensemble Modeling​​​

The above is the detailed content of Machine learning model management: integrated modeling. For more information, please follow other related articles on the PHP Chinese website!