HSBC UK Data Scientist Interview: Real-Time Fraud Detection System Design

This guide is for candidates preparing for HSBC UK Data Scientist Interview. The short answer is that the round usually tests three things at once: whether you understand the role, whether you can explain your thinking clearly, and whether your examples or solutions still hold up when the interviewer keeps digging.

这篇文章适合正在准备汇丰银行英国DS面试的同学。先说结论：这类面试通常不会只看你会不会答题，而是同时看岗位理解、表达结构，以及你的案例或解法在连续追问下能不能站得住。

If your time is limited, read the opening sections and the FAQ first. They will tell you what to revise first, which mistakes show up most often, and how to spend your next few hours on preparation more efficiently.

如果你时间有限，先看开头和文末 FAQ 就够了。读完这两部分，你基本就能判断自己该先补案例表达、框架思维，还是技术细节，不用一上来就把时间花在低优先级内容上。

My approach was a Hybrid Model combining supervised and unsupervised learning:

我的方案，是一个结合了有监督和无监督学习的混合模型：

Before the model intervenes, I would first use a rule engine to filter out the most obvious fraud patterns that don't need machine learning. These rules are typically formulated by experienced Fraud Analysts:

在模型介入之前，我会先用一个规则引擎，过滤掉那些最明显、最无需动用机器学习的欺诈模式。这些规则通常是由经验丰富的欺诈分析师制定的：

For transactions that pass through the rule engine, I would use a supervised learning model to predict fraud probability. I chose XGBoost because it typically performs well on tabular data.

对于通过了规则引擎的交易，我会用一个有监督学习模型来预测其欺诈概率。我选择了XGBoost模型，因为它在处理表格数据时通常有很好的效果。

I would construct two categories of features:

我会构建两类特征：

In fraud detection, fraud samples (positive class) are typically far fewer than normal samples (negative class). To address this, I recommend using the SMOTE (Synthetic Minority Over-sampling Technique) algorithm to oversample the minority class.

在欺诈侦测中，欺诈样本（正样本）通常远少于正常样本（负样本）。为了解决这个问题，我建议采用SMOTE算法来对少数类样本进行过采样。

Supervised learning models can only identify 'known' fraud patterns. For new fraud methods we've never seen, they're helpless. Therefore, I introduced an unsupervised learning model as a supplement.

有监督学习模型只能识别出那些「已知的」欺诈模式。对于那些新型的、我们从未见过的欺诈手段，它就无能为力了。因此，我引入了一个无监督学习模型作为补充。

The interviewer asked a question about deployment: 'Your three-layer model sounds complex. In a real production environment, how do you ensure it can complete a prediction within 100 milliseconds?'

面试官问了一个关于「模型部署」的问题：「你设计的这个三层模型，听起来很复杂。在真实的生产环境中，你如何保证它能在100毫秒内完成一次预测？」

My answer: I would package the rule engine and XGBoost model into one service, deployed alongside a low-latency in-memory database (like Redis) for quick access to user-level features. The more computationally expensive Isolation Forest model can run offline in batch mode—no need for real-time requirements.

我的回答是，我会把规则引擎和XGBoost模型打包成一个服务，并部署在低延迟的内存数据库（如Redis）旁边，以快速地获取用户级的特征。而计算成本更高的孤立森林模型，则可以进行离线的、批处理的计算，不需要满足实时的要求。