At Rise of AI 2018 in Berlin, Karol Przystalski, CTO of Codete, surveys use cases for AI in fintech drawn from his company's client work, while noting contractual limits on what he can disclose about specific customers. He covers fraud detection, stock market prediction, blockchain, insurance risk prediction and security applications such as face recognition, though he notes banking's heavy regulation slows adoption there. He describes Codete's open-source sentiment-analysis library, built and benchmarked against the commonly used CoreNLP library, with better results on most tested datasets, and outlines its pipeline of data cleaning, NLP-based splitting, dimensionality reduction and classification, in this case with a random forest model. He covers three types of chatbot generation (rule-based, retrieval-based and generative), an IoT insurance use case predicting driving behavior for premiums, and a library for predicting user behavior server-side to cut the number of API requests. He closes on the 'legacy model' problem: companies whose models were built years ago in now-outdated frameworks, for which Codete built an open-source converter between machine learning libraries such as TensorFlow and Keras.
Key takeaways
- Przystalski lists fraud detection, stock prediction, blockchain, insurance risk and security as recurring fintech AI use cases from Codete's client work.
- Codete's open-source sentiment-analysis library outperforms CoreNLP on most tested datasets using a random forest classifier.
- He distinguishes rule-based, retrieval-based and generative chatbots, the last using autoencoders or GANs to produce new responses.
- Codete built a backend converter between machine learning frameworks such as TensorFlow, Keras and R to address the 'legacy model' problem.
- He cites a customer finding that only 2 of 10 top data-science research papers can be commercialized.
Detailed notes
Fintech AI use cases and Kaggle competitions
Przystalski lists recurring fintech AI applications he has seen through Codete's client work: security (including face detection and recognition, though banking's regulation slows this), fraud detection, stock market prediction, blockchain, and insurance risk prediction. He notes that of 20 finished Kaggle competitions related to fintech, some launched by companies that are also Codete customers, not all involve classification; some, like a Santander customer-satisfaction competition, focus on sentiment analysis, denoising data or reducing dimensionality.
Sentiment analysis library
He describes an open-source sentiment-analysis library built by Codete that outperforms the commonly used CoreNLP library on most tested datasets. The method's pipeline covers data cleaning, splitting the data using NLP methods, dimensionality reduction and classification, in this example using a random forest classifier; he notes the approach has been validated live on Twitter data and points to a GitHub repository and an online Python training course.
Blockchain, IoT, chatbots and behavior prediction
Constrained by client contracts, Przystalski says Codete's blockchain work uses AI as part of the process before blockchain itself, rather than AI distributed via blockchain. For IoT insurance ('insurtech'), devices measure driving behavior to predict insurance pricing. He distinguishes three types of chatbot generation, rule-based (fixed question-answer lists), retrieval-based (a neural network selecting from known answer pairs) and generative (using models such as autoencoders or GANs to produce fully new, grammatically correct responses), pointing to a GitHub workshop notebook. He also describes a library for predicting a user's next action server-side to reduce the number of API requests needed.
The legacy model problem and a library converter
Przystalski describes a recurring issue among larger fintech clients: models built years earlier in now-outdated frameworks, comparing the shift from Theano to Keras or PyTorch. Codete built a backend converter between machine learning libraries and frameworks, demonstrated converting between TensorFlow and Keras and from R to Keras, producing identical results across converted models; he compares it to Apple's Core ML and notes Microsoft has started a similar open-source project, and says Codete intends to publish its converter as open source.
Q&A highlights
Asked how important PCA is in the sentiment-analysis pipeline, Przystalski says it depends on the problem and dataset; reducing dimensionality while preserving results is beneficial but not always necessary, especially since random forest can handle many features directly. Asked to compare classifiers such as SVM, random forest and naive Bayes for sentiment analysis, he says results depend on the specific problem and that Codete chose random forest for this case because it outperformed CoreNLP, without claiming one classifier is universally better.
Questions answered in this talk
What are common uses of AI in fintech?
According to Karol Przystalski at Rise of AI 2018, common fintech AI use cases include fraud detection, stock market prediction, blockchain-related processing, insurance risk prediction, and security applications like face detection and recognition.
What is the 'legacy model' problem in machine learning?
Karol Przystalski explains at Rise of AI 2018 that many established companies have models built years earlier in frameworks that have since become outdated (for example Theano versus Keras or PyTorch), making them costly to rewrite; Codete built a converter to translate models between frameworks automatically.
What types of chatbots does Karol Przystalski describe?
At Rise of AI 2018, Karol Przystalski describes three types: rule-based chatbots that match fixed question-answer lists, retrieval-based chatbots that use a neural network to select from known answer pairs, and generative chatbots that produce entirely new, grammatically correct responses using models like autoencoders or GANs.