Welcome back. It's my privilege today to bring you an interview with professor Francesco Ricci. A professor or computer science from the free university of Bozen-Bolzano in Italy. And a leader in the field of recommender systems for a long time. He's done work in many areas of recommender systems. From work related to recommenders for groups to recommender fundamentals. And today we're going to talk specifically about context aware recommendation. He's also come to the field with a history of work in everything from case based reasoning. And other areas that bring AI into the field of reccomender systems. Including extensive work in user modeling. He was the program co-chair of the Second Recommender Systems Conference. And he's really one of the people who's been a formative leader in nurturing the field. Including many of the graduate students who've gone on to success in the field. So I'm delighted to have you. Welcome to our class Francesco. >> Welcome, thank you for the introduction. >> Well I am delighted to talk today, because our topic is context aware recommenders. And perhaps we should start with this question. Obviously, it's a hot topic. There are a lot of papers that are out about context aware recommenders. The last several years, every conference seems to have a workshop on incorporating context into these recommendation systems. Why don't you start by telling us, what do we mean by a context aware recommender system? >> Okay, let's start with some examples, so just to clarify the idea. So imagine that you want to recommend a vacation to someone. Obviously you have to take into account whether it's winter or summer vacation. Or the system has to recommend some purchase. And the system has to know whether it's a gift or it's something for you. Well, these two examples, as you point out, the importance of context. The first case is the period of time and the second case is the motivation for the recommendation. So essentially in general, this idea of context. Where a recommender system is coming through this. Understanding that context is in general any kind of information or condition. That can influence the perception of the usefulness of an item for the user. And for that reason, so since context is any kind of information this type. Recommender systems must take into account this information. In order to provide more precise and more, let's say, useful recommendations. >> So you've given examples like time or whether something is a gift. Are there some other examples of the kind of contexts that you think have been particularly important in recommenders recently? >> In general, there are several kinds of factors. I think that maybe we can show some examples, though. And looking at these examples, you may understand better what kind of applications have been developed using this idea of context. So let me show some slides. >> Okay, we see your slide here on a mobile application. >> Okay, so this is an example of a recommender system for places of interest. And we have here three of these snapshots of the user interface. And on the left, you can see that the user can select what kind of factors the system has to take into account. One producing recommendation. So for instance the distance from the points of interest, and the temperature, or the season, or even the companion of the user. And a user can select what factors have to be taken into account by this system. Then in the middle part of this slide you see the recommendations. The recommendations are customized based on the context. And, in fact, some of them are, let's say, considered by this system particularly [INAUDIBLE] for this specific context. So situation made with this icons where there is this [INAUDIBLE] green arrow. This means that this particular point of interest is the [INAUDIBLE] for the particular context of the user. And on the right you see also an explanation for that. So for instance this place is good to visit with a family. [INAUDIBLE] better motivate the user to take this recommendation and go for that point of interest. >> So this user might not get the recommendation of the cable car at all if they said, I'm coming alone. I'm not here with my family today. >> Yeah that's the point. This is particularly [INAUDIBLE] and could even not be shown to the user if in a different context. And this screen you see the possibility of the user to select contextual preferences. So to set some contextual conditions that are used by the recommender system to customize the recommendation. So in this case, the user can say whether it's calm or active. Or the system can collect some context from data like the weather, or the traffic conditions, or the type of the road. Based on these contextual conditions, the music which is recommended is selected. In a catalog of music tracks that are made available by the content provider. Okay, in these last snapshots from the user interface. You see on the left that the user can select [INAUDIBLE] the other users [INAUDIBLE] can say to the system. Please consider that there are these other passengers. And the system will recommend the music by balancing the preferences of the different users. By the way, you can even, let's say, preset. You can store some of these context recommendations or preferences, and you can create some customized music channels. And so, when you click on these music channels that are shown on the right. The system automatically will reload these context recommendations. And will produce music which is adapted to these contextual conditions. >> So, tell me, when you talk about these different contextual conditions. It seems like some of them are about an individual's preferences changing. Some of them seem to be about not necessarily preferences but what's appropriate for a situation. And some of them are external factors like groups. We talked with Tony Jamison about recommending for groups. So that one I think we understand. Do you treat things like am I happy or sad the same way that you treat things like, am I sleepy or awake? Or are some of those hard-coded by your system to say. If we know the person's sleepy we're only going to recommend loud music? >> They are treated in the same way. And so they are all context conditions that are treated the same way. What is not treated in the same way is this, let's say the group composition. This requires a completely different modeling approach. And probably Anthony told you about how to do that. In that case, for instance, in our application we did the kind of aggregation of recommendation. So we generate the recommendations for the different users. And then we try to find out the best aggregation of these recommendations. So this is also showing that the context recommender system called for a complete range of techniques. So in general, we particularly say that the model is a multidimensional model. Because the classical recommender system obviously is by dimension. And so the [INAUDIBLE] of users in the items, and add in the contextual data that I mentioned we have in the three dimensional and multi-dimensional model. And this is something that makes the the design all recommendations outlets are quite difficult to do. >> This multi-dimensional model seems to add a lot of challenge. It's already hard to do a good job recommending with the sparsity, in just two dimensions. What techniques do you use, to be able to handle the complexity of these contexts or recommendations in this more complicated multi-dimensional model. >> Yeah, in fact that one technique is very, very basic and try to go back to the big dimensional models. So they need us to, it's called this reduction method or for filtering method that, so the idea is that when we are under contested conditions, let's consider the ratings that have been acquired the exactly in the contested condition. Then you have a slice in this third dimension, and we just consider the way things are acquired in that specific context or condition, and we use a classical [INAUDIBLE] model. This is one approach, it's called pre-filtering approach. >> So with this pre-filtering approach, the major challenge is you just have less data for any given condition. But if you have enough data, you can use the same techniques that we've already been using for classical recommender systems. >> You are right. And in fact now, coming on some hybrids methods, instead of using exact pre-filtering try to say attitude their way things collect in their particular target contested condition. Some other way things are that have been collected in contested condition that run somewhat similar to the target blob. And then this idea of similarity can be tuned or can be identified according to different criteria. So for instance, you can say that to contest a condition is similar if it's influencing in a similar way the rating behavior of the users. This is something for instance that we have now presented at the last time at the conference enroll. >> So when you use your music example, might you say similar would be, I had many different dimensions I'm I happy or sad, I'm I driving fast or slow, if the data hits five out of eight of those is that enough to be similar that we might include those ratings, or do you only look at similar within each dimension? >> You will look at the similar in the effect of the conditions, so what I'm saying is that, for instance when you are sad you will tend to rate a certain kind of item lower but maybe you will have the same behavior when the weather is not good. And then we'll say that these conditions that are totally, I'd say not similar, they are not even compatible. They are similar from this point of view, in the sense that they influence in a similar way your waking behavior. >> So that requires a fairly rich model of user behavior. Yeah, yeah. All you have to let's say, analyze the data trying to find out this kind of similarity, which is typically difficult to do because again, your very fast about the rating theory of use. >> So are there other techniques that you commonly use for incorporating context aware recommendations? >> Yeah, and let me show slides where there is a summary of these, let's say, approaches. >> Okay, so on the left there is the diagram representing this contextual pre-filtering. So the idea is that data, reading data, that are let's say, multidimensional data, whether it's the user, the item, the contents and the rating. Using the context are let's say, you have this bi-dimensional model and so you have just the user, the item, and the waking dimension. And you generate the two dimensional recommender system that you could use recommendations. In the center part of this slide, there is this contextual post-filtering. This is even easier in some sense. The idea is that you generate recommendations by using all of the data that you have, even this contextual conditions. But after the recommendations have been generated using the context, you filter out those recommendations that are not applicable for the specific contextual condition. For instance, if you have a mobile application, you can generate recommendations for points of interested in general that could be interesting for the users. And then looking at the particular location of the user, you may detect that, the points of interest are too far from the current position of the user. I'm not good for the user. So this is post-filtering, after the recommendations have been generated. This is easy. The more complex and probably the most interesting modeling approach is depicted on the right and this is called context for modeling. Here in this case, let say the context is used inside their prediction model. So you need really to operate on this multidimensional data. And in order to do that particular approach is to use Danson's factorization instead of Martin's factorization which has become a very popular approach for evaluation recommender systems. And in this case you composed the multidimensional, data and you tried to build recommendations directly on the multidimensional data. >> This shows that once we start getting into context, either we need to be in a situation where we have lots of data or we need to have some fairly interesting modelling techniques. Which is a good reason to explain why context to where recommendation is one of these advanced techniques that is at the state of the art today. It does seem like a lot of what you've brought up has to do with mobile. Is that an area where we're seeing more and more of this context to where recommendations that are location were in particular? >> Yeah, I think so. Since that mobile applications in some sense provides a natural scenario for this kind of context that we're recommended systems. Users are mobile and there's a context of conditions that are changing in a very quick way, so you need the system to adapt quickly to the different conditions. So that's the reason why essentially we have a lot of application in the mobile world. Actually, I think that mobile applications also bring some would say, different kinds of context, so they put into the, it's another different kind of content. Because we have this, for instance the location of the user, which is called the physical context. But we have also with this mobile application some other dimensions. So for instance, the social context, the people that are around the user while he is mobile called interaction context and media context. So we are into [INAUDIBLE], the user is interacted with the system that is mobile, with the obviously limitations. So this again, calls for particular techniques for generating a personalized recommendation and also for adapting the interaction to the media context. >> Wonderful, well thank you for sharing this with us. Are there any other last thoughts or any pointers to sources that you'd like to share with our class? Yeah, I would like to bring more general idea about what context is, because we said contexts are factors that are influencing the perception of usefulness of an item. Therefore they recommend the system ask that they take into account that but actually to answer the reason processes of that users are performing. And I want to show you again an example just to illustrate how the context may be useful to call ambiguous messages. And this is there is this figure that is quite interesting. So on the left you immediately clearly see this ABC. Now you read this ABC. On the right you immediately read the 12, 13, 14, no? But actually the B and the 13 are written exactly the same way. So the fact that on the left and on the right you read 13, it's based is essentially on the context of this two signs. So context is useful in order it's automatically used by our mind in order to, in some sense, decode this ambiguous messages. So this shows how important is context. And so essentially it's not possible to even interpret the meaning of certain items or recommendations without context. The context is always present. This makes sense from my point of view, we cannot build any more recommender systems without taking into account the context. It's a fundamental dimension we have that you can see there. >> Then tell us about that middle box when Ann approached the bank. >> Yeah, that's another example. I mentioned that when you read the sentence, Ann approached the bank, you think about a financial institution. I guess. But a bank has another interpretation. If you maybe had [INAUDIBLE] just a few seconds before, maybe you're not thinking about the bank but thinking about the weaver and the bank of the weaver. So the point is that what even you have thought, you have seen, let's say maybe a few seconds ago influences what the interpretation of the current messages. So this is for instance mainly important for music recommendations. The music that you have been listening to in the scene previously influences your preferences for the music that you will like to listen afterwards. And if you do not take into account this you will provide the very back recommendations, you maybe keep to recommend something as this recommendation will be coming boring for the using. We will not have enough diversity in the recommendation. Only B will create recommendation on top of the unrelated in the sequence. So this is again another example of the context condition which is not just the state, the location of the user but there's the items that have been recommended and are being used by the user before. >> Well, I think that's a wonderful set of examples. And it also points out that it's not just what happens before, but what happens afterwards. If you hear Ann approached the bank, the waters were near an all time high. You know it's a river bank. And if you play a particular song, and then follow it by a set of very fast songs. People may observed that the first song was too slow when they might not have felt that way if the song's progressed, more close in tempo. So that's a wonderful set of examples. >> Yeah, and I think this is also interesting when you take into account the point in time where the user is rating the items because normally the user is not rating the items immediately after let's say the consumption. But he's rating the items maybe some time ago. So even something that has been consumed after that it could have influence of their rating that the user provided for items that have been consumed before. That's also, I say shows how complex it is to generate meaningful recommendations by using this of ratings when you do not know anything about how these ratings are being collected. And when these ratings have been collected. >> Okay, well thank you for joining us and for educating us a bit about context aware recommendations, and we appreciate your willingness to join us in this experimental class. >> Thank you, Joe, for giving to me the opportunity to talk to your students, and all the people that are interested in context aware recommendation assistance. Thank you very much. [SOUND]
