Since its establishment, JFIN.US (JFIN.US) has always regarded technological innovation as an important engine for enterprise development, continuously exploring the application of big data and artificial intelligence technology in different business scenarios, and striving to bring better products and services to customers and partners. In order to further empower decision science and intelligent operation, Jiayin Technology recently launched a self-developed “Shilan” audio data mining algorithm and a “Ming Jing” text data mining algorithm to fully unleash the value of unstructured data, marking a new level of the company's scientific research capabilities and big data strength.

Based on years of business operation and data accumulation, Jiayin Technology has accumulated a rich “mine” of data. Due to the unstructured nature, how to extract valuable information from it and transform it into structured data to further improve the quality of decision-making, enhance customer experience, and ultimately drive business growth has become an important topic for enterprises to think about.

Through exploration and practice, Jiayin successfully developed its own “Shilan” algorithm for audio data and the “Ming Jing” algorithm for text data to efficiently extract valuable structured data from audio and text data according to different business scenarios to provide more decision support for downstream data analysis and modeling. These two recent achievements show Jiayin Technology's active layout in the field of technology and deep insight into future technological development. They mark a new chapter in the in-depth understanding and use of audio and text data driven by data.

The audio data mining algorithm, named “Shi Lan,” is inspired by sound fluctuations like water marks. The algorithm can identify subtle changes in the speaker's mood from the ripples in the sound. The Jiayin Decision Science Center uses digital signal processing (DSP) tools to convert audio files into signal sequences, and then uses Fourier changes to extract the speaker's acoustic characteristics, such as spectral center of mass, zero crossing rate, mean square root energy, etc. from the perspective of the time domain and frequency domain respectively. These characteristics aim to explore the information contained in speech, tone, and speed of speech. Since there are big differences in how people speak in different emotional states, such as speeding up speech, sharp pitch, and raising volume when emotional, etc., by analyzing this information, it is possible to more fully understand the speaker's state attributes at the time of the audio.

The “Ming Jing” is a type of ancient academic examination designed to test students' ability to understand and use Confucian classics. Jiayin named the text data mining algorithm in the hope that it can “learn and apply” to find useful information for business from massive amounts of text. Currently, the team at the Jiayin Decision Science Center has upgraded traditional machine learning text mining ideas, so that the model can automatically search for valuable keywords according to different business scenarios, and expand the keyword database from the perspective of synonyms and synonyms, which plays a role in multiple recalls. In addition to this, the company also uses the Big Language Model (LLM) to further understand the semantic information of the hit text to improve the accuracy of recognition. This method of combining traditional machine learning with big language models not only enables simultaneous improvement in semantic label recall rate and accuracy, but also helps reduce the company's business costs and increase efficiency.

Currently, these two data mining algorithms have been successfully applied to scenarios such as data modeling. Xia Chunqiu, a model development expert at Jiayin Technology, said, “Structured data and unstructured data complement each other very well. Currently, in many scenarios, the integration of acoustic and semantic information can improve the predictive performance and stability of the model. This fully demonstrates the business value of our exploration of unstructured data mining.”

The latest credit service model report shows that the variables developed by these two unstructured data mining algorithms account for 27% of the total model variables, which strongly supports model predictions. The prediction effect of the new model using text and audio variables has also been greatly improved compared to the old model. Under the final 20% quantile threshold, the new model can increase the capture rate by more than 60% and reduce the occurrence of risk events by 40%. The application of the new model not only reduced operating costs, but more importantly, significantly improved customer satisfaction.

In the future, Jiayin Technology will continue to adhere to the belief that “technology empowers service innovation”, continuously break through technical boundaries, optimize service processes, and work with customers to build a path of high-quality service quality. On this path, Jiayin will also always adhere to customer needs as the guide, technological innovation as the driving force, and continuously improve the level of professionalization and individualization of services.