LLM 및 데이터 분석을 사용한 금융 추천 엔진

### LLM을 사용하여 금융 추천 엔진 구축하기
* 👋 사용자 행동을 분석하여 상품이나 서비스를 제안하는 추천 엔진에 대해 알아보는 금융 분야 **Rag Plus LLM 사용 사례의 일곱 번째 파트**에 오신 것을 환영합니다.
* 📊 이러한 엔진은 **사용자의 급여, 소득, 부채**를 기반으로 신용카드나 대출과 같은 금융 상품을 추천하여 개인화된 경험을 통해 고객 참여를 높일 수 있습니다.
* 🔍 추천 엔진에는 유사한 사용자의 행동을 기반으로 제품을 추천하는 **협업 필터링**과 개별 사용자의 선호도에 따라 유사한 제품을 추천하는 **콘텐츠 기반 필터링**의 두 가지 주요 유형이 있습니다.
* 🛠 이러한 엔진을 개선하기 위해 **최인접 이웃 또는 신경망과 같은 알고리즘**을 활용할 수 있으며, **임베딩 모델을 사용하여 향상된 하이브리드 추천 시스템**을 만들 수 있습니다.
* 💻 이 프로젝트는 인구 통계 및 국 데이터를 사용하여 은행 상품을 추천하는 **J Financial 추천 엔진**을 구축하는 것으로, **Python Lang 체인과 대규모 언어 모델**을 사용합니다.
* 📈 이 방법론에는 **오픈 소스 데이터 전처리, 벡터 데이터베이스에 저장, 사용자 데이터를 기반으로 추천을 생성하기 위해 RAG QA 체인**을 사용하는 것이 포함됩니다.
* 🧑 다양한 고객 속성에 대한 합성 데이터를 생성한 후 상품 적격성을 판단하고 **소득과 부채 수준에 따라 개인 대출이나 모기지를 제안**하는 등 맞춤형 추천을 제공할 수 있습니다.
* 🔮 다음 동영상에서는 머신러닝 애플리케이션에 유용한 특정 패턴을 생성하기 위한 **합성 데이터 생성**에 대해 알아보고, 추가 리소스를 통해 시청자가 AI와 데이터 과학에 대해 더 자세히 알아볼 수 있도록 독려할 것입니다.

J Financial 추천 엔진은 인구 통계 및 금융 데이터를 사용하여 은행 상품 추천을 개인화합니다.

hi guys, welcome to Free Birds Q, and welcome to Rag Plus LLM use case in finance domain playlist part 7. So let's get started. So if you don't yet watch my 75 hard gen challenge playlist, please watch and learn gen for free. I will put the link in the description. So here I show you the recommendation engines that are used to recommend products or services to users based on their behaviors and attributes. So this whole system will analyze your personal behavior on applications like Netflix, Amazon, or Flipkart and can recommend you the product or services based on your behavior that you might buy. Okay, and in the finance domain, we can use those recommendation engines in many ways, like recommending the best credit card to the user based on their salary, income, and debts as well. Or we can recommend them the banking products or the lending products as well based on their behavior and the bureau data. So you can personalize user experience, segment them, that can increase the customer engagement as well. Okay, and the recommendation engine is of two types.

one is collaborative filtering, and the other is content-based filtering. The collaborative filtering means you are comparing with the behavior of other people and recommending products combinations they buy. So, let's just say if you buy milk, then they also recommend you the bread as well because that is the behavior that they get from another person who might buy bread and milk together. Okay, the second is content-based filtering. You will get the recommendation based on the similar kind of products. So, if you buy milk, then you can also recommend another milk-based product like butter or cheese as well. So, in that way, the content-based filtering works. So, J can enhance this recommendation engine algorithms that are K nearest neighbor or neural network that are used in the industry right now. But with the help of Gen, we can use the embedding models that can generate the embeddings of the customer attributes and compare them with the products and other users and build the hybrid or enhanced recommendation engines. Okay, so as I just promised, this is the next project from this list, and

I'm building this with the help of a mistal fine-tuned large language model that is called Zyer 7B llm. So let's get started to build our project and build our J Financial recommendation engine. Let's get started. So guys, here is our project, the J Financial recommendation engine. So we have a problem statement that we need to build a recommendation engine that can recommend bank products to the user based on their demographic information, along with personal information and bureau information, using the Python Lang chain and large language models. So this is the complete methodology that we have to follow. We need to get the open-source data of a company product that includes demographic and bureau data as well. Then we use Python to load that data and preprocess that data as well. And then we load that preprocessed data into the vector DB by using an embedding model from Hugging Face, because Hugging Face embedding model is completely free. And in this whole video or in this whole playlist, I'm using only

the uh op Source free models only okay, and then we build the rag QA chain that will help you to uh build a complete rag llm based system that can take the data from the vector DB and can uh build your whole uh project as that can recommend the product to the user as well okay. So let's get started with it. So you can also watch my other playlist from this link, and you can also get the SE SE uh 75 hard gen challenge playlist from this link as well okay. So first, I need to install the faker library because I need to generate the synthetic data here okay. So I just generate all the things that I just need: the age, gender, medal status, income level, as along with the education, occupation, idential status, dependent uh debt to income ratio, or the credit buau score as well. So these are the features that I just need because these are the features that you can easily get if you have the financial data with you okay. Then I just uh store all this data, and then I also need to generate the

inquiries data as that what kind of inquiries that person did in the last 3 months or in the last 6 months as well. Okay, so that I get to know what kind of product that customer is looking out for. Okay, so here I just complete the whole script here and build my whole kind of data here. And in that data, you will see we have all this information. We have customer ID too, and then we have the last 3 months personal loan inquiry, last 3 months credit card inquiry, or the mortgage inquiry, last six-month variables as well. So, these kinds of variables will help you to understand what kind of products customers are looking for, and based on their other information, are they eligible for that product, so that we can recommend that product to them. Okay, so we need to also check for the eligibility here based on this kind of information that we have. Okay, so first thing is we need to build a Content column here that takes all this data and stores it into the.

content uh uh column in the prompt for format okay so that prompt format will look like this okay and then we have to uh install our Lang chain and Lang chain community libraries and so we need to convert this kind of a dictionary or this kind of a list into the documents uh format and we will store all that data into the doc M ments list okay then we need to install the sentence Transformers because we need to load our embedding models as well we need to install the chroma DB and along with the bits and bite and accelerate to load the large language models that we have is the Zyer 7 llm okay so here I first load the hugging phas embeddings here and then I just put all that document into my Vector DB and under the collection name of recommendation engine by using my hugging face embeding okay so now your all the data is stored in this Vector DP now you need to load the large language model so I just load the L language models in the quantized format because quantization is necessary.

here because if I don't quantize the model, it will take too much time to load, and it also consumes so much space and memory as well. Okay, so that's why I use a quantization here. Okay, so once I quantized it, and then I can load the model easily. So here, you see my whole model is loaded successfully, and then I build a query pipeline here because I need to generate the responses in the text generation format. So I pass my model, my tokenizer, max length, and everything into this query pipeline so that I can use it to generate the responses. And then, with the help of Hugging Face pipeline, I build a single function out there that can take the prompt and generate the response as well. Okay, and this is a kind of a simple function that can just format my output or input text as well. Okay, so here, I just check it with the small question here that is, What is recommendation Hine, and how it is used in the finance domain? As you see with

the help of this function. It can easily colorize the question and the answer and format all the responses that it gets as okay. Okay, so now it's time to check it on our ual data. So, we just build a rack QA chain, and we build a template here so that it can provide the best recommendation for the customer as well. So, we just build a retriever function with the help of our chroma DB, and it can get the best result out of it. And based on that best result, it can build a QA. That kind of QA chain will take this kind of customer's data, and we pass this whole kind of customer's data into the string format into our QA chain, and it will generate the response. So, you see here, this is the response that we just got based on the data. The stable lending product for the first customer would be a personal loan with a lower interest rate due to their low income level and very high debt-to-income ratio. Okay, very low income level and very high debt.

to income ratio. The second customer with the high income level and the low debt income ratio may be eligible for the mortgage of the home loan as well. So, okay, so in that way, it just provided you the complete answer in this format. Okay, so I hope you guys completely understand how I built this single recommendation engine for the finance domain by using the rag and lar language models. So, if you want to know more, you can also build your own models with multiple kinds of other language models and experiment with them. Okay, and in our next video, we will talk about the synthetic data generator that will generate the synthetic data for you so that you can get your own certain patterns data that can be used for your own products or own machine learning models or deep learning models as well. Okay, and if you want to know more about PR engineering, generated AI, data science, and machine learning, you can watch my videos on YouTube and also read my blogs on Medium. We'll meet in our next video. Thank you, guys. Thank you so.

GenAI Financial Recommendation Engine | LLM + RAG ChromaDB [ Zypher 7B LLM ] Mistral LLM

FreeBirds Crew - Data Science and Generative AI

추천 엔진은 넷플릭스나 아마존과 같은 플랫폼에서 사용자 행동을 분석하여 상품이나 서비스를 추천합니다.

금융 분야에서는 사용자 행동과 데이터를 기반으로 신용카드, 뱅킹, 대출 상품을 추천할 수 있습니다.

추천 엔진에는 사용자 행동을 비교하는 협업 필터링과 유사한 상품을 제안하는 콘텐츠 기반 필터링의 두 가지 유형이 있습니다.

이 프로젝트는 인구 통계, 개인 정보, 금융 기관 정보를 사용해 은행 상품 추천 엔진을 만드는 것을 목표로 합니다.

방법론은 오픈 소스 데이터를 수집하고 Python으로 전처리한 다음 무료 Hugging Face 임베딩 모델을 사용하여 데이터를 벡터 데이터베이스에 로드하는 것입니다.

그런 다음 이 프로젝트는 벡터 DB의 데이터를 기반으로 사용자에게 제품을 추천하는 RAG QA 체인을 구축합니다.

이 스크립트는 개인 대출, 신용카드, 모기지 등 지난 3~6개월 동안의 고객 문의 데이터를 수집합니다.

이 정보는 고객이 관심 있는 상품을 파악하고 해당 상품에 대한 적격성을 평가하는 데 도움이 됩니다.

그런 다음 데이터는 문서 형식으로 저장되고 벡터 데이터베이스는 추천 엔진의 임베딩 모델과 대규모 언어 모델을 로드하는 데 사용됩니다.

모델 로딩 및 응답 생성을 위한 정량화 사용

정량화는 모델의 로딩 시간, 공간 및 메모리 소비를 줄이기 위해 사용됩니다.

쿼리 파이프라인은 모델과 토큰화기를 사용하여 텍스트 생성 형식으로 응답을 생성하도록 구축됩니다.

이 프로세스에는 최상의 고객 추천을 제공하기 위해 크로마 DB의 검색 기능이 포함되어 궁극적으로 고객의 데이터를 기반으로 맞춤형 응답을 생성합니다.

추천 엔진은 소득 수준과 소득 대비 부채 비율을 사용하여 고객의 모기지 자격을 결정합니다.

발표자는 맞춤형 모델을 구축하기 위해 다양한 언어 모델을 실험해 볼 것을 권장합니다.

다음 동영상에서는 머신러닝 모델용 패턴 데이터를 생성하기 위한 합성 데이터 생성기에 대해 집중적으로 다룹니다.