SAP Predictive Analytics Industry Use-Cases



everybody welcome to that session just a few words again to myself my name is Andres Austin I'm based in Switzerland working for a CPS global Center of Excellence on oolitic in which I'm an expert for predictive analytics and in today's session I want to cover a number of things I want to very very briefly introduce the area of predictive analytics but really focusing on use cases and what kind of areas of predictive analytics can be used to implement various different use cases have brought along a couple of slides on customers that have that solution in production in action and also I would like to show a number of demos using the tool implementing or showing how to create models that are very typically used by our customers and everything that I will be showing today life in our tool you would actually be able to carry out yourself later after the webinar where you may know we have a patron version of our tool that you can use yourself you can download it free of charge and the data that I'll be using we are offering it as a download and we have tutorials that guide you through the demonstration exactly the way that I'm doing it here and now basically introducing the area of predictive analytics very very briefly divided a seamless like a file is helpful so that's like them you might have seen it before we see how the maturity of an analytics environment is growing providing more and more value over time and we all know we we start with raw data we have to put the data together we have to clean it we have to take care of the data quality we have to use the data and we have to put it into processes just to make it available for analysis often formatted reporting sometimes advanced reporting but no matter how agile or interactive your environment is generalize the data there will always be a point some kind of limit as to what you can find yourself as a business user in that data of course it can get extremely complex and that is exactly the point where predictive analytics kicks in where we can connect to existing environment could be in data warehouses that we are connecting to leveraging the data you already have so all the effort you have already put in we can leverage in the area of predictive analytics where now we can use them let's say clever mathematical or statistics to find specific patterns in your historic data and find these patterns or trends and try to continue that trend into the future essentially predicting or estimating what is likely to happen in the future it could be one predictive models that could be angled it could be thousands of predictive models giving you a lot more insight into your business into your customers into your products and now with an insight giving you the chance to take action without additional information if you can really leverage the historic data to to shape the future I like to call it so that is predictive analytics and to make it a little more more realistic or cosmetics and I think this could be a helpful second introduction where very often predictive analytics is used to optimize the resource which is not available and without end so most of the resources were working with unlimited it could be your time it could be your finances if it could be anything where predictive analytics can help you make the best of the resources that you have and often that starts with a name that could be a strategic goal of your company what is your company working toward and what are actions that help support that aim that your company is aiming to achieve and reactions that you're undertaking typically will have some bottlenecks and this is where predictive comes in handy expands to make best use of it and to have it as a real direct in case let's say you want to optimize marketing campaign and hear the end you might be handing or many of our customers have is to improve market share in that case by growing revenue all of us would like to do that and then the action that will drive to increase the market share of the revenue could be to increase the number of successful cross held existing customers they'll give them ideas about further products that they are likely to light and their that bottleneck very often can be the marketing pressure that a customer is likely to accept so you might define that for typical customer you only want to get in touch twice a month maybe by email maybe by telephone maybe other channels but very often if you send too much information people get overloaded maybe noid you don't have enough information so people think you don't care so in this case imagine you don't want to have two contacts a month which is that bottleneck so once you have a contact with the customer once you send an email which message do you want to transfer which product for example would you like to promote and here that prediction that we produce it could be done for one product it could be done for a hundred products then it sets into the marketing process where now of course the campaign will be optimized so you may want to recommend the product where the customer is most likely interested or where estimated profit for a product might be highest to the customers agreed to it and you can build and in in such a process typically we have different personas on one side we have the data analyst who is creating these predictive models and we have the market here who's benefiting from the inside so it doesn't have to be the same person in the business department creating this predictive model very often it can be data analyst it could be the subject matter expert somebody was go to good Flair for data but very important it does not have to be a data scientist so data science skills are very rare and we have an environment that will show you where we can create these models highly automated like in that case a data analyst can really focus on the kind of business question you would like to optimize our second example here from manufacturing imagining um want to increase your customer satisfaction and one way to achieve that could be like when an order comes in that you are ready to deliver but when the certain product is ordered of a certain quantity that ideally you would like to have the product in stock you can deliver immediately and that bottleneck could be the this place in the warehouse the cost to produce product of a capital that is tied up in the warehouse of products that are not selling or not selling as quickly as you would like and with that predictive analytics could estimate likely demand for your controller for the next month the month after but the next year whatever time intervals make sense for your industry and again the same principle to data analysts who might be the same analyst that is also helping along on the marketing side can create these insights and deliver them to the processes where they are needed in that case it could be the production manager finding that further information that's already to different use cases but these cases can be extremely gathered and a belief in every industry in every department predictive analytics can be used to help the company her to profit and extremely often companies are starting their journey with predictive analytics in sales and marketing such as what other products that a customer is likely to be interested in and but also think of the opposite way what is the risk of accusing an existing customer where and if you know that information in advance that as likely to move to the competition if you know that early enough you can try to take action and try to try to then actually consider next best actions what can you do to strengthen that relationship before the customers actually asks you another area is can be clustering or customer segmentation where the customers that you have now all of them are of course different but others are more similar to each other than others and segmentation can help to find certain groups of customers that are fairly similar without you necessarily knowing upfront what is it that makes them similar the segmentation can identify these groups which you could then handle specifically or individually maybe you run different marketing campaigns for each segment that was found or maybe different sales approaches different LC may want to handle these different customers because at different requirements then moving to operations where here then we could think of predictive maintenance so if you're manufacturing what is the risk of a machine and having a failure where he showing unwanted behavior which could be vibration or other things you'd like to avoid the earlier you know that the better of course you can intervene something else is very common is in demand optimization we knowing in advance which products are likely to SEM in which quantities I showed it earlier in the example on optimizing scarce resources and most companies don't just have one or two products it could be dozens it could be tens of thousands so we have the ability to must produce models from Agee's we don't have to do everything manually but it scales with the demand that you might be having all we think of them quality management which could grow in the area or them while you have produced an item or might be semi produced it is not completed yet here you could use them predictive analytics to estimate how likely is it that if I complete this product will it be of good or bad quality if you think it's likely to be a best quality you can decide what to do maybe you want to discount the product but invest any more money into that product just to throw it away later but at the same time predictive analytics can also tell you the reason or drivers patterns that lead to a certain behavior so in that case predictive analytics could give you indication what kind of details lead to that product quality maybe there's a certain sensor in the machine maybe to certain raw material from a specific supplier that if you know that information you can investigate and find out what you can do of course try to avoid these problematic areas to actually increase product quality or now moving to fraud and risk and I'm just picking out a few where in fraud clearly we're thinking very often in insurance cases where fraudulent claims can be made where with predictive analytics the fraud well the insurance case the claim can come in it can be analyzed for any suspicious behavior and it was you can reflect for internal auditing to have a closer look if needed you can also go into the area of compliance issues in the thing of your own internal IT systems or security access where it can happen the people by mistake are given too many access rights towards some certain activities in their own system they might know it they might not know it but based on cases that we have seen in the past you can find a predictive model trying to identify further employees who might have been given too many access rights and again you can look into their access settings and adjust if needed then we look into and hey John very very commonly the cashflow forecasting where it is all about trying to optimize your finances so the better you know how much finances are likely to come in tomorrow the next week the next month the more accurate your insight to that the better you can adjust your financial planning we're using financing costs I'm having to take smaller loans potentially or take advantage of different payment terms of your own suppliant single better how much in cash you're likely to ask them in the future or also a budgeting where it is about them have really being more accurate in your planning application we're not all planning applications take predictive some insight into account when trying to estimate them the next planning cycle it could be internal numbers you know because it could be cost but again it could be revenue but we're here predictive analytics could produce more accurate plan values which can still be fed into your existing planning application when please can still look into the numbers that can adjust if needed but hopefully the required adjustments should be smaller or then the last area we have the other sectors where literally across any kind of industry as we call it though there are so many use cases if we think off of healthcare just a big one example it was involved in where it could go into the area of hospitals where of course there's a very high need or desire to prevent patients from getting new infections while in hospital you could use predictive analytics for new patients trying to assess the risk of these people being infected by certain CDs if you know that in advance maybe on certain patients you have a clothesline or be more careful then you would be anyway but also here you can identify the patterns that can lead to infections also trying to avoid these under in public sector could be security-related or big data and IOT could be another insurance case where telematic data can be leveraged so think of the little boxes that are now part of some insurance contracts which are embedded or placed into your car that measures your driving pattern and this data could be used to assess or let's say your driving style or the risk of having an accident in future which then the insurance company will be very keen on I'm using four connections on for example the risk of you having an accident so many many different use cases it is really hard to focus down on a few but them what we do here with our tool we call it sub Business Objects predictive analytics we believe it's very very special where we can create these predictive models without you having to be that data scientist where we have a whole comprehensive framework but there's not any assumptions about the data that you're using but where now you can focus on the business question you want to solve leading to results in a very short timeframe so typically it can be days not weeks it can be even a lot faster depending on what you're working with but also to set the expectation right and we do need some kind of data source to connect to so if today you would like to leverage predictive analytics on five six different sources still you know some Essen needs to be put in to connect these different pieces whether physical or semantic so at once we have the data of course then we work that automated approach and this automated approach also means that we can produce these thousands of models it is not not just one model that an expert produces and of course also has to maintain so we think of the custom analytics with affinity of a person to be like interested in the product so the model that is created today is accurate to gain by customers behavior will change over time and if the behavior of your customers is changing if it does then the predictive model isn't as good anymore as it was on the first day because it is it describes a situation which doesn't exist anymore as such which means model need to be monitored need to be retrained if needed so that we have a component that is doing that automatically I want to go into ATM today to details to mention the name we call it the factory which takes these models monitors them at a scale and of course just having this insight having these predicted values somewhere sort in a Seidel doesn't really improve your business it only changes the way you work if you really see that information back into a process you bring it where people need it where people interact there could be for example that marketing process will now read that information from the inside you know which product you recommend and these use cases we've been speaking about and we cover with a number of core data mining function we have embedded in our tool and the naming a belief can be almost dental exchange it could be a machine learning artificial intelligence is very close to it we very often call it predictive analytics but they are extremely similar areas so I believe the terms most cases really can be interchanged and with the core functionalities here like the most likely you will need your fist city and I'll go through them one by one before drilling deeper into a few starting with a classification where here it is all about trying to predict them a yes/no event when something happen or will it not happen so that could be trying to predict if your customer likely to turn or not or is an invoice likely to be fraudulent or not it could be preventive maintenance is the Machine likely to show strange behavior or not so any kind of decision you can put into a yes/no classification goes into that classification then the secondary here is a regression where we estimate a numerical value it's not a percentage it can be let's say any number very often it is like it can be sales quantities as you see on the slide here how many products is the customer likely to buy next quarter giving you an idea about the strength or relationship you have with a customer but it could be totally different information it could be for a retailer to make sure you have them you plan to open a new store you have two or three locations available so where would you like to open your store then a regression can be used to estimate how much revenue or profit you are likely to make in each of these locations based on different criteria where is it how close is you to train line how well is it connected to the public transport how's parking how far is next competition so that potential store that regression will give you an estimate of how much revenue you're likely to make all the regression could even be using the area of data quality imagine your business to consumer company you store detail about your end-users so some people you might know the celery father's you might not hear regression can be used to estimate that the celery or the household income for people we don't yet have that information the third area here is an augmentation and clustering where we try to identify these groups of similar things whether customers whether it's products doesn't matter we find item that somehow similar that you can treat them in a certain common way or if you look at forecasting we're here we look into the area of them let's say numerical values that appear over time in regular intervals so it could be daily resin use it could be monthly revenues it could be how many people are calling a call center day by day so that is forecasting and then recommendations is very much where you have let's say hundreds thousands of products that we try to use his health history to come up with product recommendations like you know it many many websites you know clamoring you look at the product he will show you recommendations for other products you might also be interested in and now going deeper into the classification where I show grief example here coming back to the case where every optimize the telephone campaign where you have many many contacts that you could possibly call but of course time is limited we have a restriction your bottleneck so who should you call cells who should you invest your time on the principle ISM you look at who did already buy that product in the past these are the ones that are marked in green of course that gives you two clear groups people dips and you didn't buy and now predictive analytics comes in trying to separate these two sabisu with a classification rule once the rule is found you can apply on the people that you have come and here looking at history we take an all the information to counter is available for these people h gender the more you have the better buying behavior and it was separated to as good as it can and looking at the past will never do it perfectly if it we've done a perfect hundred percent accurate separation most likely something went wrong but their prediction so there's a risk of getting something wrong but once we have that rule we applied on the people that we haven't called yet and now third person we know their likelihood of being interested and very simple you sorted of course you know who to call first and that's the principle for classification no matter whether you do it for market analysis or for very different use cases could be churn analysis could be preventive maintenance and then I'd like to show you now in the tool I'm up in speaking about Business Objects predictive analytics think of me now as an analyst somebody who's got a flair for data was to create a new predictive model for the marketing department so working in the automated way if you're an expert we also has an expert environment that case today though I'm really taking to automate it you can see the core functionality is very much inside introduce them we're now can go into a classification and the first step is to connect to the history of the data that you want to learn from that could be as well it could be a database it could be Hana or very pragmatic it could be a fat file since I'm doing email we're in this data set that you can also download a mention all of this you'll be able to do yourself afterwards we're in this data set I see people that I have already called and it's actually real data set from the real world it's the Portuguese bank that was tight enough to share information not people ready so they did actually call ego is a person that was contacted we have some information about these people their age whether they have a loan for example or at the end did or did not buy the product that was being promoted which in this case water I get a savings product investment product so that's history and now all you need to do is almost tell the to what to describe you're not really exposed to two algorithms or anything else that many things you can fine-tune but the idea is to really focus on the business question you want to solve for classification you need a target variable to purchase is no decision that we have on the Left yes as much information as possible little just really short example could be hundreds thousands of columns the row ID I'm just excluding the bottom provide any value and with that I can simply hit a trigger is very much like an automatic camera you decide what kind of photo you like you press the button in all the settings of course it is done automatically as it was done here it is giving us some information about the quality of the predictive model that was being produced without going into too much detail now it's telling us that 14 of the columns were being used let them it has a good possibility to describe or to find the people who were buying it's got a value of 0.5 number one would be perfect zero would be a no added value whatsoever and it also found the model to be robust that's telling us a predictive predictive confidence we 0.99 and that could already be a model you could put into production but was likely especially on the first item so now you want to try to understand but the model you want to understand its its findings its value and here let's chance them that cane chart we can see the quality of that model on the x-axis we see all the people from the data set and on the y-axis we see only the people who did actually buy the product which of course are the ones that we want to identify and ideally would like to sort people e on the x-axis but at the very left I have the people that I believe are likely to buy these are the ones that I should be calling in to explain it I'm quite quickly if you were working randomly so no predicted inside whatsoever you mention you pick 10% of the people from history that means you will also find 10% of the people would buy the product that is a red line the more the better and nothing optimized if you are a wizard you would not exactly in advance who will or will not buy there would be the green line where every person you call or you identify if the purchaser is a buyer and you can just stop at some point you have achieved everything reality however is in between if you see in that blue line that where this blue line describes what our model is predicting and the it's called for every person the likelihood of being interested it sorted it thank you where I was explaining the value that our model is delivering which we see as a do curve here and imagine now we were just conducting 5% of the people that we have available and if I go up on the blue line then you can see could release five percent of the people we can achieve 22 percent of the people actually bought the product so we optimize our campaign by the factor of four or even the bus for in that case so that is the value provides a balla model and the question was now which attributes were actually selected that we can see in you you can look and drink you don't have to well now we see the 14 columns that were chosen and the job of course the most important one and anything that was irrelevant was it was removed it is also why we can work with tens of thousands of columns in we just remove the ones with the death no value but put it into the tool and maybe there's a chance it can help and year we can also drill into these columns to better understand how they assess the mode the column P outcome here stands for previous outcome tells me how did that person behave in the previous marketing campaign that I had already run you click into that column you see now the values available within that column and so the simply put the server you are on the left the more likely it is the person is interested so where the last campaign was are already a success chances are they'll be interested again before ever you know maybe we've never contacted them unknown as a chance they will not be interested and we can do that for all these columns I am even for numerical ones where we have in each column where we know the person's age and where the model was able to identify the ranges of ages that showed similar behavior people over 58 very high chance of a being interested in the product and then the second most interesting interested group that whether the younger customers aged between 18 and 30 so the other end of the age spectrum probably interested for very different reasons if you can speculate that can speculate people over 58 maybe they've paid of the mortgage they have some finances available people aged 18 to 30 maybe they haven't purchased the house yet maybe they entice them or set up some savings and now that you know it makes sense to invest good money so here this is the way where you can really understand what is driving that behavior which behavior would you like to enforce which you can move from behavior pattern would you like to avoid and now on the screen where I can sort of simulate through that model will now see the 14 columns that were selected the tool can pre-populate the most common value for the columns or finding like a super average customer jungle where for this person I have a likelihood of 5% of being interested and this is just a screen where you can try to better understand the working of the model is not how you would score your thousands of millions of customers in the production but you know we saw that if the previous campaign was successful and if the person was for example aged 60 the probability of success should be higher I will recalculate it and receive as a probability suddenly is 53% so compared to the very average person here I only need to pick up the phone twice with a 50% of chance instead of twenty times I've optimized my bottleneck and the model that we have here not it could mass-produce this course and write them into your CRM system or you can produce logic program code that you can use in real time the moment person calls a call center it could produce a scores for that person that have the call agent to structure or drive the conversation but of course I can also now score information in batch and yes you would now score people you can call just yet I just have the same data sets are applied on the same people again just a flat file you can say that you're interested in the probabilities for these people and I could write it in the system or again place into a flat file just call it problem and with that the logic that in had already produced is being applied its graphing that data where I needed it I can have a preview of that information so it shows by you for the first hundred rows from bottom output not the best 110 people so let's make it the bigger I can sort it on the column which contains the probability we're here we see people actually with a similar probability of 53% of chance of being successful so this could be the list for your marketing campaign these people you really want to contact and we're event I'd like to go back to the slides sort of leaving the classification um but briefly mentioning we have customers using that technology that try to keep it brief course because we have that delay but one customer of ours I can speak about this M Bank which is a banking group based in Poland and also actually call it first online bank so data for them is absolutely important because they don't really have that permanent face-to-face interaction with customers and here for instance they can use the card transactions let their customers have for marketing campaigns and three understanding the client and the user information for product recommendation but not just their own product but there's a network with partner organizations that provide totally different products there could be in consumer electronics it could be restaurant deals where they provide them this kind of product such as maybe a TV often bundled together with the standard banging product so here something that sets em bank totally parts from the competition the building up the network and providing information beyond pure banking really bringing information cost that is relevant to the people not another customers covenant transport there is just an use case for still a classification where in their industry it's the hottest company in the u.s. there's a challenge to to hire and keep their attract riders and now we use predictive analytics in that classification to predict what is the likelihood or risk of losing the driver but also trying to identify that the pattern or the reason why the drivers are quitting the job trying to work against that just to why how we had interpreted the model how we found certain information that was driving a certain behavior now covenant is a way of that they can try to learn to avoid these areas to keep people longer and employment and the return on investment you know that they achieved within the first half year so very very quickly paid for itself then going deeper into the forecasting so we have like time series and to also explain that quite quite quickly mention we have monthly sales numbers ourselves quantities demand with a certain history over time where in that case no we have a nice very steady behavior where we're selling more I'm over and over again dear friend so predicting is very very easy to do of course but reality is different is more complex or off there are seasonal patterns where here we can see spikes during the middle of the year which we can also identify fuel annoying at seven months and the historic demand we can predict what is likely to happen in the next year on a monthly basis but real reality still is even more complex where it's a lot more there's a lot more variants in these numbers no they're their ups and downs and not as regular as I shown so far but also these patterns we can predict especially if we can take condition information into account think of the weather information thing of promotions you are running so we have a huge impact on this number what I speak here now about Amanda it could be cash flow could be people call your call center it could be all sorts of things it is all the same content and also that I would like to show in the in a brief demo and yeah so unserious analysis and he I have another data set again from the real world pretty much demand forecast you could think of it as a is a demand forecast where many cities around the world now we have that concept where we have a concept you know where you can hire bicycles them for briefly for one or two hour you can cycle for kilometer or two you just chop it up elsewhere and let exists for example in in London and London is sharing the number of bicycles that will read out each day use these three for last couple of years where we have a day and we have the number that we interested in and only with that information we can ready do a first forecast where now this crease lightly different where does it well realize this time information there was only one column available with the highest and quality to do is again sort of them what okay tell it how many days to forecast let's say ten and now it is looking at that historic data is looking for a trend is looking for different types of seasonality could be it could be weekly could be a monthly it could be a yearly could be annual and find lots of different statistical ways trying to describe that task and then write that into the future which will give me a forecast in a second where we can see the whole history but also the prediction that the tool has made such the time series is described with a different measure it's called make a median average percentage error so it has does on average what by how many percent will be wrong here we were wrong by about 19.7% and depending on the use case of situation this could be really really good or maybe there could be room for improvement and we'll try to improve it in a second but now to show the actual forecast was being produced when that screen here this one in a little we're in green we see the actual value and in blue we see the forecast that we produced and the model was also applied on the historic data that will of course note but then the ten days that are asked for we see on the right hand side here whether a shaded blue areas prediction interval and the smaller red area the more confident we are in the prediction and we can see that there's clearly like a weekly pattern for example every new bandage is weak but it sort of tries to follow the real values but it's not extremely close we have to say especially some numbers them that have blocked they're marked with a red square these are numbers where actual and predictions are far apart which we call them outliers so something special happened on these days and that something special could be all sorts of things in this case it could be the weather it could be a cube strike it could be the Olympic Games in town in London but it's included in the history and we can now try to describe the bigger context two additional columns which of course I have already done so I have a second data set it's called extended so we're here now I can still see my my day and the number of tiles with a head but there's a lot more information for example telling me now the weather data the mean temperature min or max humidity so how much did the temperature change during the day where the Olympics in town whatever you can think of you can try to put it in here and quick look at the numbers we can see now these additional columns with information under subscribed and let's a relative of the process in order to try to improve a prediction that you try to find additional data it has to it helps the forecast but it will help the model and here again I give all the information in thank I try to forecast ten days again so some warnings about the data structure which we can ignore and now it is it is looking at all these columns is identifying what column had an impact will improve the model how will it improve the model it is going through all of these combinations and this one now might take a second longer because there's more data to analyze that man is so now we see how the model is improved the make that dropped from I think would work 18 or 19 percent down to 13% which we should also see when we look at the actual forecast we're in here not what the color change rather parental is the same wear green is still the actual values that we had and now go forecast them year we have in Indian red where you can see how the model is a lot more close to the actual values in how it happened and this is not possible because these extra columns but extra information was available and we can also see and which columns were being selected here to be more accurate but the most important here was information where there was a working day or not a working day that's what's meant with holiday but also the temperature here the teaming mean temperature plays a very important part and with that back to the presentation and well customer the recommends and will integrate they use that to improve their the production planning to know better so how many quantities to produce not have too many or not much but not too few in stock but I'll one mention others or a Cisco for example Canaria meant to cover was recommendations I'll leave it out today but then it is about the area of if you are the company selling it say hundreds or thousands of products and how you can derive recommendations for these so not just one or two predictive models but out of thousands of products try to identify which products to individually recommend to your customers and if you like to see how that could be done it is part of the tutorials that we will be sharing and so you can follow all that yourself at the data set again from the real world where people listen to music it's about 700,000 combinations of people listening to a certain band you know how often listen to that band we could use that dataset to make your own little music recommendation engine or maybe you can find music such as a product that you yourself might be interested in with that oh I could give an introduction to the many many different use cases but there are for predictive analytics and how companies can benefit from that let we like the way we do it highly automated but it means you don't have to be that data scientist if you are better you can free up some of your time if your data scientist that we do that by by scaling we have a factory component that can maintain these models keep them up to date and bring them back into the processes where you need that insight to actually change the way and you're working to optimize your processes and hopefully this course you're interested in predictive analytics at the topic hopefully also as is the tool that we're providing here I mentioned initially that there is a trial we can download the software which gives you that installation exactly of the tool as I was using it you can use it for 30 days and to help you along and the first test there are tutorial you see the link on the screen but we wish have a site so you don't have to write it down where you can go through that purchasing affinity piece of customer interested in the product or not and we can do the demand forecast and that I showed but also in addition by product recommendation at in their time for now today that which in a way is link analysis so combination of products and customers so you can find more boy inside and with that thank you much

One Comment

  1. Aziz S said:

    Great insights!

    June 27, 2019
    Reply

Leave a Reply

Your email address will not be published. Required fields are marked *