yeah I will talk about FinTech use-cases uses of AI for that I actually added a PhD have skin cancer recognition but I'm leading the company called culture where some of the customers are vintage customers and obviously I cannot say everything right it's a presentation based on from our experience working with FinTech companies but I try to say as much as possible not breaking the contracts at the same time when it comes to when it comes to FinTech nai it can be combined and used for in many in many cases here's the list of the cases obviously security there are some classic issue classic cases like the fault detection for example it's obviously used for the analysis stock market prediction it's it can be used also for for blockchain and obviously risk management insured prediction it's kind of ours as well at least these are the topics that we have experience with probably all of you know cattle right have you checked how many competitions are actually related to FinTech there are 20 currently that are
finished some of the some of the competitions here some of the companies that actually gave started the competition are also our customers don't worry I will share the presentation later you don't need take a photo don't worry it will be on our blog and there are a few that are really very interesting I mean all of them are interesting but this is something that are specially interesting for me and one of that is a sentiment analysis for stocking prediction I will talk about later in more details but as you see also in short our potential AXA for example and even more there are some hedge funds here on the list that are also the as you can see here on the list it's not always about about classification a lot about developing a model that will at the end give accuracy like you know all 90 percent 95 percent or whatever sometimes the challenges are more about for example denoising the data reducing the dimensionality so it's not always the case that we need to
build a classifier and here's again another another the first one from Santander customer satisfaction competition it's also about sentiment analysis so the first topic is the security that I mentioned in the first slide you know what it was that right I don't need to I don't need to explain as a convolution Network used very widely in many many many cases so now it's I'm actually traveling many many many conferences some of them are actually related to FinTech and it looks at there are more and more companies actually using actually face detection face recognition for security for banking obviously banking cases a bit more complex because they are highly regulated so the the bigger issue of that but I think at some point it will it will be applied widely sentiment analysis it's about getting know what's the emotion behind the text for example it's all the voice as well and we have we did we have build a library that is open sold you can actually find it here here's the link and the link
for github is on the next slide where he we have build a library that we compare with core NLP that is a commonly used library and it is well it gives our library gives better better results for most of the datasets that we have tested so this is one thing that we have develop also for our customers obviously for each customer it's a specific case this one is our library that we have shared on open source I don't have too much time actually to explain all of the topics related to sentiment analysis so if you're interested you can also go to our live training on the already platform where I'm giving also as training in Python how to actually build our sentiment analysis method from scratch more or less how does it work how does sentiment analysis method work or our method how does it work it has its it's combined about one two three four five six steps all as the most important thing and it's not only related to sentiment analysis it's it's
not really the random flowers or any other classification method it's about the data so when we are training showing how to build a sentiment analyze the library we are actually mostly focusing on cleaning the data to get good data that we can actually analyze and obviously now splitting the data using using some NLP methods prediction reducing the dimensionality using some classification method like in this case random for us and get new jobs if you step by our proof you can see this library used on twists live so it is working and here's here's the code here on the on the bottom you can just go to download and test it the next topic is actually blockchain and IOT and combined together because we are doing both but because of the contract means we can I can't say too much what I can say about not chain is that if you go and google it blockchain AI in most cases you will find that some people will say that they are trying combine AI with blockchain
in using AI in a more distributed way we don't have such an experienced attorney what we do we are using more like a aiya like a part of the process before the blockchain itself is used so it's a bit different for now I cannot say exactly which customer is what exactly it is it's combined with obviously before the coins and probably next thing else I will be able to make say more about that and when it comes to IOT this is a known topic IOT ensure tech cars the devices will actually measure how you're driving and based on that predicting the enterprise insurance price the other big topic in FinTech it's about chatbots it's replacing the customer service but it is one of the example or measuring the sentiment of the social medias it's also an example we have three types of regeneration of chat pods it's a rule based retrieval paste and generative paste in the first case there were before some other species and only dedicated to chat bots but just to go
fast through route paste is when you have to list question answer getting the question replying with the answer to what easy notnot really don't way I know machine learning used well female pairs when you actually learn you have a narrow network behind usually and we have today's that and we are generating the answer based on the data set and generative pace when we use for example autoencoders of deep learning metals organs where we are narrating the answer that was not known before it's totally generated with full well it's properly from the grammar point of view and the language point of view but it's totally new and again here is a workshop on github it's a it's a Jupiter notebook you can download and figure and check it out and the first link obviously we are also having a training on Safari online and you can actually learn how to build each of kind of desportes I me I don't have enough time PhD - also known topic and when it comes to FinTech we have
but some time ago before actually the PSU topic appeared to be a topic we have developed our library you can download here which is used to predict the users behavior it means that on the server side you are predicting what will be the next step of the user using an API and we are sending the data predicting what's actually the next step in one in one request one respond instead of many so we are reducing the number of requests which is at the end faster for the user and here's also a document how how I actually is working another topic that is quite important there are some legacy models we have we spoke with some of our customers especially the big companies they have a problem of do that problem we can say issue it's a better word here that they have built some models three years ago five years ago and now they're using a totally different framework library let's imagine we are going to years back if I ask you what would be
what is the what library for machine learning are you using probably the answer will be Tiano right now if I ask the same question you will say probably kerosene or PyTorch right what will be dance in two years fortunate at the same that today you'll be and there will be new frame new libraries that's for sure you know that it will happen right and if we go back five years probably some of the companies used I don't know C++ or maybe are to build the models and that's a problem and in the meaning that if you have some legacy models that you you still can use because it's working but would like to actually move it to a new library or new language it's kind of a tricky because you need to spend some time just to rewrite it to the new language new framework and we have actually a solution we have a lot solution that actually is solving this problem it's a converted between the libraries and frameworks if I couldn't just ask
you to show the demo it's just really short demo here as you see it as a converter I know it's it doesn't have any UI because it's our back-end converter I mean it's it's a back-end so you can sit here on the on the telus here on the side and as you see we are now converting from tens of thought to Chara's from cara's to tens of law the next will be from tensorflow to Kara's we have also a converter from our from some libraries from r22 to Kara's and so on so it's quite a hard thing because we have around as you see it's it's dock arised now it's converting you see the code here that is generating that now when you go through and do an analysis how many libraries we have in machine learning there are around 100 about three of them are used in production and what we have built we have build a converter between all of these but not all about most popular libraries and frameworks you have here
the next one if you if you are if you can see that actually the results are exactly the same for different for different data sets so it's converting and giving the same results which is good if you know Apple they have caramel right they're importing from different libraries they have this such a solution Microsoft now also has started our open source project where actually they are building something similar and we we do the same I mean not the same we are doing something similar we are building our own converter for other for everyone actually because finally we like publish this as open source as well fancy here's here's the model build based on the previous one and next we will have the results here one after one of our customer that is actually in the automotive is something wrong is an automotive has the problem some other of our customers has the same issue but it's also the only reality big companies that actually are working on machine learning form for many years you are
the results and you will have four different different different models for different attitudes and it's the same okay I have so few more minutes so and there is also one thing one more thing that I don't have on slides we see also one thing when it comes to FinTech especially obviously this is this this problem is also related to the FinTech it's all to actually to all all different branches the more and more companies are actually established in cooperation with universities is something really especially the FinTech one of our customer did our research that 2 of 10 research papers from the best journals in data science and motivation cells just can be commercialized so if you are student I would like to do some side projects in machine learning just let me know and find that book number three maybe we can help you to get more experience in this area that's it for now 15 minutes perfect if you have any questions feel free to ask microphone if you like so where was it
I like questions so on one on one of your slides you showed term frequency inverse document frequency I forgot what that was classification or something and you had also pca forgot which don't mean a sentiment Allah is exalted yes so my question was is how important is the PCA part in that algorithm it depends on the problem and on datasets obviously it's good to have it because you know if you have if you can reduce the dimensionality with having the same result and it's good but it's you are right it's not you it's not it's not a we need to always use it I'm asking because you're using random forests after that so I assume that random forests could just deal with how many how many features you had in that model but if you had even like a thousand he could deal that's true it's wonderful one of the example obviously we can put a different different classifier here there's a different metal t as well so that's why we have to hold the
whole approaches but obviously we can replace the classifier with a different one one more question maybe following up on that question in terms of the classifiers you show the bag of words approach with the tf-idf and then the PCA and what's your experience based on that using for example support vector machines random forests or a nice base comparing to what's your recommendation like over four four four sentiment analysis we use for show you some probability method as well as VM I think it's good for pattern recognition maybe not really for for for this case but it depends on the problem and then I think the classifier obviously you can get better or worse results but it depends on the problem I would say it's hard to say some of the examples actually you can say Oh SVM is the better one here in this case it's random forest we put random forest because we get better results with this classifier compared to coronal P but you know for a different example you probably will give
better get better results for most different kind of classifier that's it okay so if you have any questions any more questions I'm somewhere around here trying to answer all of your questions thank you