 So thank you so much for coming today. I know that this is the siesta time, so I'm really sorry but you can take a coffee and Take here So we are talking about recommender systems, but with an special Issue that is the conjoined analysis So let's me introduce him my colleague Jose Maria He is one of the best experts in market research at CDO Telefonica She has a PhD in physics and he is associate professor at Universidad Autónoma de Madrid He also Has been working at Telefonica 20 years more or less and he always working with issues related with conjoined analysis Customer uses behavior and so on so please and applause for him Thanks very much and Let me introduce you Paula López. Paula is one of the best experts in Is one of the best data scientists in the chief data office unit in Telefonica He is working with machine learning algorithms and obtaining very interesting insights from Customer data. She's graduated in statistics. He also has a master degree in statistics And he's preparing his PhD and he sees also collaborating with the university and several institutions as lecturers and what is the pressure working with her Let's see this nice painting this Today we are not working We are not speaking about art Maybe we are speaking about philosophy here in this painting of Raphael You can see the most relevant ancient Greek philosophers discussing About how they see and how they understand the world and like in this picture in the chief data office unit Paula, me and our colleagues. We also discuss about how to meet our customer needs And in fact, if you look in the picture, probably you can find us Paula and me Yes, we are there just in the middle here. There are two of the most relevant Philosophers of the ancient Greek in your left side. You can see plateau Plateau represents the world of ideas what customers think what customers like and what customers believe And as you can suppose That a scientist from nowadays are like Aristotle The philosophy of Aristotle was that the facts and observation was the cornerstone of the universal reason And this is the same philosophy that we apply today when I when we make decisions Basin based on data As you probably know Telefonica has over 300 million customers over the world with the presence in 18 countries Each of them with different decision-making processes with different Cultures with different relationships with different ways of relationship with the company and with different things They want or different things they like at the end what they what we have are three hundred millions or different customers and there are a lot of recommender based on big data, but one key point is that People don't take the decision of what they do people take the decision of what they believe or what they remember and probably You you think in situation like this when I was when I when I am in a traffic yam I always think that I am in the slowest lane in that in the in that traffic yam Or when you are in the supermarket you always think that you are in the slowest queue of the supermarket You look for information about Free online courses you obtain a statistics as say that only 20 percent of customers finish free online courses However, when you ask the students the students declare that 80 percent of them finish free online courses 20 percent or 80 percent for the teacher Courses finish when the course is finished the content is clear But when you ask the customer or the student in this case For the student the course is finished when the student has reached the goal the goals they had from the beginning So this is the origin of the difference of the data statistics of 20 and 80 percent Reinforcing that message Winter is coming sorry Christmas is coming If you Christmas is coming and probably you are preparing presents for some special person If you only take a decision about the present taking into account the big data You are going to buy something based on previous Purchasings of the customer or purchases or other or other people similar to that person However, maybe it's also interesting to ask that person. What do you like for Christmas? But if you only asked to the customers Maybe he only say to us Little things like I want a board game or I want sportswear but big data can help us to learn the idea so What is the best board games for families or maybe what is the best sneakers for the best athletes? So I think that there are no opposed concepts, but complimentary And in this sense, I'm going to explain the our study case the smartphone world We know that choose a good smartphone is an easy task for a lot of us Why because there are a lot of devices and a lot of features to analyze and Please Jose Maria if I tell you do you want a good battery or The latest fashion recognition for your smartphone. What do you what do you prefer? Oh, I always have my charge at me. So I prefer the facial recognition Alright, so this was easy, but if I tell you you want the biggest screen or The latest fashion recognition or a low weight or maybe so on there are a lot of features to take into account but not but only But you have a lot of possibilities to combine all of them So you need a tool like big data to sell it that and I forget the price. That is so important so Let's me start introducing big data. We know the value of data We know that as a company we need big data to take data driving decisions To know what customers will be do in the future and to know what the scenarios we are going to Do to encounter in the future. So Let's me start introducing our problem We have a world of non-historical information we have to We have to use a A Different algorithms to work without historical information. We have to we don't have it and We have to split our devices in futures because we don't learn about Something like devices that are really changing in the real world So we have to do all that in real time and from the source From the surf information available So let me introduce with the brain of our product our product is composed by three parts The first one is the world's algorithm this algorithm is based on a reinforcement learning one This it could be useful to our customers to know what device I have to recommend to them For another hand we have to know where is the best channel to do it. So where and The third part is tell us why we're recommending this device. I'm not another one This is a part Formed by a natural language processing so I'm going to introduce the what and the why part, but I'm going to And not to To explain the where part because it's out of the scope of this talk So don't be crazy with this schedule I'm going to start explaining the algorithm that is inside the device recommender that That is called contextual multi-art banded algorithm This algorithm is based on a context We have a thousand of features about our customers and we have to sell it one device This device is selected in a base of a probability model in a linear various probability theory Select one device is called an action and as you can suppose in this kind of algorithms Multi-arm we have several actions. This is several arms So when you choose one arm, this is one option one action We have to show that to our customer and to tell him I recommended you this device and there are two options First one he likes and second one. He don't like it So in both cases, we are learning we are learning about he bought the device We are recommending but we are learning to if he don't buy the device that we are recommending And this way can we can learn this is easy in the first case we are Reward our algorithm and in the second case we are penalizing the probabilities of our model In this kind of problems of recommender systems is difficult to use collaborative filters in this case it's not possible because Collaborative filters are very useful if you have valuation or preference of our products of Or of our items, but in this case We don't have this kind of market because the smartphone world is so changing So it is not possible to do that For another hand, we have the white component This component is able to build a sentence that can you humanize the recommendation In this way, we create a sentence format by a customer feature and a top feature to recommend We are going to show you an example to understand better This is based on a cognitive process Concretely in our natural language processing This is an example For being a customer without limits when you're shooting impressive photos Maybe you are interested in this model X with an enduring battery and able to capture any perspective of your view This is an automatic sentence created by that but our algorithm As you can see The first part is composed by two customer features. These features are without limits and shooting impressive photos As you can see, we know this information about the customer and these two features are selected based on the probability model For another hand, we present the recommended device The recommended device is in this case a model X But we normally present the brand and the model not anonymous, obviously and in the third part We have the device recommended features that are in this case and are in battery a capture any perspective As you can see the first part and the third one are correlated We know that he don't have He consumes a lot. I'm sorry So we are going to recommend a good device with a good battery And I know that he has a good camera in her current device So I'm going to recommend another device with the best camera of the market So, Jose Maria, it's your turn No, I'm going to explain you what kind of information we obtained from market research to add to the big data available information The market research technique that we use is the conjoined analysis Conjoined analysis try to understand how customers take decisions based on the attributes of the product We are going to see that within an example Imagine a crucial decision that we all take every day Where to drink a coffee with your work colleagues? When you are going to drink a coffee, you have different options. Maybe you can go to the Office coffee machine Maybe you can go out of the office and go to the corner and drink a better coffee But a bit more expensive more time or maybe you can go to a top-rank coffee shop But maybe a bit far and a bit expensive What happens when you are taking these decisions you are Measuring in your head the trade-offs between the different attributes in this case We have three attributes. We are simplifying. We are going to consider only three attributes We are going to consider the price the quality and the time each person is going to take its own Decisions and probably the same person in different circumstances are taking different decisions So we are obtaining a probabilistic model You can think things like today is in a special day I want to go to the best coffee shop and then you are going in this case to a top-rank coffee shop Or maybe you are thinking I don't have too much time today and I need a good coffee Then you are going out of the office to go to the coffee shop at the corner Or maybe you think things like I have no time I need a very fast coffee and then you have to go to the office coffee machine Then conjoined analysis Analyze these decisions to see if for you is more important in each moment the price the quality of the or the time in the case of The smartphone recommender is time instead of price time and quality We are considered when considering a lot of attributes like price Camera selfie camera the processor memory and of course the brand there are much more attributes to take into consideration But take into account that for customers They only take decision based on probably in a small amount of these attributes mainly mainly brand Price and maybe camera or memory or things like that Each attribute is divided into levels so with Distribution of attributes and distribution of levels we can cover all the possible smartphones present or future smartphones, so we can arm a Recommendator for all the possible smartphone that could be in there in the market for example for the brand We have different levels that could be brand one brand two brand three We are going to speak about Apple, Samsung, Huawei, CTA, Alcatel, Motorola, etc. etc. Adopted to every market of Telefonica and for example in the camera we can study the importance or the relevance of the Camera ranging from 2 megapixels to 20 or 25 megapixels and the same for all the attributes The price the selfie camera the processor the memory, etc. What customers see are random cards with random smartphones inside like the above in this case Customer are doing the survey and see three different smartphones Smartphone A say for example is brand one is a very good brand with a lot of very good features A camera screen memory, but a bit expensive $900 and there is a second option brand number two is a good brand also no so nice features because the Memory the camera and the screen are lower, but it's cheaper than option a and brand number Three in the option C has a nice features is the brand is not so good It's brand number three is not so good Not so good and one and two and it's cheaper only two hundred dollars And there is an option that is none of them fits my needs So I will choose none of them in your case Paula, which one will you choose? I found the brown one so I'm going to choose eight Okay, then every customer has to do between seven or ten of these choices Okay, so we have several choices per customer so in our research We can study the patterns of every customer. We can see if the customers always choose brand one Always choose the cheapest one always choose the one with the highest memory or maybe always choose the one With the best memory, but below three hundred Three hundred dollars We have at the end around two thousand or three hundred or three three thousand interviews of different customers Then we can introduce the utility the utility concept because from the point of view of the facts If you have 20 megapixels is is better than having 15 megapixels or better Had that than having ten megapixels or five megapixels But in the head of the customer that is that is not happening that way because maybe for a customer is enough to have Ten megapixels so moving to 15 megapixels or 20 megapixels If the customer doesn't need so good camera has no value added for him So a customer doesn't need 15 or 20 megapixels so has no extra utility what we are going to To try to obtain is the utility of every attribute and of every level every level of every attribute and we can do that Calculating the total utility of the smartphone that is the sum of the utilities of every cast of every component of the of the smartphone To do that we have a lot of equations per customer because we have explanatory variables that are the utility of every level of every attribute So we have a lot of explanatory Variables, but we have some dependent variables that are what customer has chosen in every car and we And we have a lot of information because for every car We have the answer of the customer saying if they choose or not choose that that option The results are results like this We have the utility for the camera and the utility for Brand and we have the utilities for all the rest of the attributes in this case for this customer That we have is the if we move from five megapixels in the camera to eight megapixels The utility is increasing a lot because we see that the curve increases However, if we move to 10 15 or 20 megapixels that we see is that The utility for that customer is increasing, but more slower So for this customer, maybe a smartphone with 10 megapixels is enough Maybe 10 megapixels and if we look at the brand and the brand figure that we see is that for these customers brand number one and two have a high utility However, brand number five or seven has lower utilities if we are going to recommend Smartphone for this customer using only these two Features we will recommend probably an smartphone of a brand one or two with a camera up to eight megapixels and and now the time where we place together the information from the Market research and we join that with the best Or the big data So thank you. So has Maria so can we merge the two words? That is the big data and the conjoin analysis We do that in a simple way. That is with like clustering over our customers base With that we can search and find different Behaviors of our customers and create different profiles This is an example of our base. So the clusters are Characterizing but some futures like the brand for example the main brand brands Versus other brands or the type of contract. It's not the same prepared or up space, etc and so on and Each cluster has a questionnaire associated and these results can be extrapolate to all the base You to this clustering So this is the equation The new score that we calculate is composed by the score of dinet by the contextual multi-art banded algorithm Multiplied but the exponential of the differential of the utilities This is not so complicated because the differential of the utilities are only the subtraction between The model utility of the model to recommend and the utility of reference. This is an example here in the table And with that we obtain a new score that is used to recommend We use that in a real case and this is we obtain By eight times we assist our mother versus a random recommendation. What is that? The standard models we use in the past were eight times worse in comparing with our recommendation system So the result is The results we obtain an incredible use case. I'm very funny with this This use case is very flexible. We can apply that to order in all the Countries or the present of Telefonica also very robust We can do that with the smartphone that we can do with other products Yes, and we can do that in a real time. So really fast obviously and Always with total transparency We only use the data for our customers if they want this is if they want to buy a new device They want to that and we can help us. So we are going to do it And also adaptability because we can do that in all the countries all the products with a lot of Attributes features, etc So thank you very much for coming. Hope you have enjoyed Any question you can There is any question Okay, okay, so thank you very much. Thank you. Let's