Welcome back, in this video we will continue Dr. Mobasher's presentation on Context-Aware Recommendation. >> So here's what I was talking about earlier with representational model using like Tensor Factorization, right? The representational approaches, for contextual modeling, there are two general set of approaches. One is based the tensor factorization I was talking about earlier, which you can kind of see here as an example of its, where you can see the users and items being sort of one dimension in the tensor. But then for each one of the contextual variables you have another dimension in the tensor. And these tensors can of course get very complicated and computationally expensive to factorize. Especially if you have many different context variables, the other sort of type of contextual modeling approaches that people have looked at our for lack for a better term. We have termed them conditional or dependent conditional modeling, which focuses on interactions between users, items, and contexts, right? So trying to actually model one of those. So I mentioned in context that were matrix factorization. That's one of the original approaches based on that idea. More recently these kinds of approaches have been extended to look at context as sort of and what we called so look at deviation base context modeling. And the idea there is that the way you model context by looking at hard ratings deviates among context or to the same item by sets of users, right? So what's really you're trying to learn is to what degree a particular contextual condition is relevant based on different ratings. People's ratings are for the same item in that context compared to let's say, the general case where you're not considering context. So that deviation is actually what your modeling and that ends up being relevance or equivalence to many of the other major factorization schemes and you can model more things with it. Another sort of approaches are similarity based contextual modeling where the idea is to actually model the similarities between different contexts in terms of how users are rating items in those contexts. And so this is, you can think of it as the other side of the deviation based modeling. But these approaches are very useful in dealing with very sparse data sets, because then you can focus on more general similarities about contextual conditions and actually group together certain contacts. When there may not be enough information in one particular set of contextual conditions because there are other similar ones that you can sort of aggregate together. So on the interactional side I talked about latent variable context models. The idea there is you have serve generative approach to modelling users' context where the assumption is users interactions typically, involves some of small rating contextual states, that explain the users behavior. All right, so please get back to this idea of the context not being observable but sort of conflict you know giving rise to the activity and the activities from changing the context. There are lots of domains where these ideas have been used. But they're particularly useful when you're dealing with situations where users are involved in some kind of an ongoing activity. Performing informational or a functional tasks or maybe listening to music, for example, in a playlist or in a streaming fashion, things like that, right? The basic idea is you got items and you got users and you got some ratings or observations about interaction with these users. And from these, you want to learn some set of latent factors that sort of describe these interactions. And these latent factors, you cannot associate them necessarily with some specific set of contextual variables. But, they're sort of hidden factors that describe what's really happening. And once you have learned these factors, then you can use them to do certain kinds of reasoning, right with the in most approaches probabilistic type of research. So one set of, a group of papers that sort of people have worked on including my own group along these lines was the idea sort of looking at particularly in the constant music recommendations but, also web page navigation. Look at the sequential activities of users and try to infer context and changes in context by looking at the sequence of activities. In the music recommendation domain, for example, you can think about a user listening to songs, and then interacting with those songs either rating them or skipping them. Kind of like in a streaming service and then being able to determine or identify potentially potential changes that might mean the big user has gone from one context to another context. And typically, the way we done this is through as I said late [INAUDIBLE] modeling by looking at the typical situation users engage in and how they might be associated with different characters, things that were attributes of songs. In one particular paper, for example, we look at extracting feature information, tag information from songs associated with a particular, let's say a playlist, this is a sequence of songs. And then mapping those the learning models in this case, we used latent application to try to find topic models, basically, from these tags associated with songs. And then each one of those topics then would represent sort of a potential context. And then you can look at how users go from one topic to another topic potentially as part of a sequence or song that they listen to and observe changes in those contents. The last thing I want to mention in here is the context interactive, or interactional systems or interactional context, is the actual context adaptation is situations where you're dealing with an interactive system where you don't necessarily have a lot of information beforehand, right? So we don't have, let's say, a whole bunch of other user information where you can use to build topic models, or probabilistic rate and variable models and so on. So, in this kind of situation, then you're going to need a more active learning approach, or kind of a reinforcement learning approach, we're each interaction give some information to the user, or to the system which allows it to build a more accurate model as you go forward. So you can imagine in the context of let's say e-commerce, maybe you are trying to buy a camera right, so you look at a bunch of different cameras but at some point you're done with that particular task. Either you purchased it or maybe you were going to come back to that and you start looking at toys to buy for gifts, right? So, when your in that particular context for buying a camera, presenting to you with a recommendation or the next item or predictive item as they lens for the camera, that might make sense, right? So in that case that's a good recommendation, but once you've switched to your task then the new context really represent the new utility function, so in that context recommending the camera lens is not going to make sense. And the thing is we want to be able to identify the changing context automatically. And the way we will do that is by sorting frame that hidden context will change the context. So there are different approaches people have been looking at and this most of this is fairly recent work, but basically it comes down to this sort of general scheme where, you know kind of in the normal on the left, you have this sort of normal recommendation approach. Where the system observing the user feedback, on particular items and getting user recommendations either they stand it, learn model, or based on some other criteria and all concepts of items etcetera, so the users. Back here, we want to be able to estimate the utility of items based on the user profile but we also want to detect if there's a change in context and if there is, determine the effect of that change and then update the utility estimate. Right? So this adds this new sort of additional path in here that you can try learn that in different ways. The current approaches that are being use to do this is, as I mention earlier, is based on this idea of exploration, exploitation, right. So, the idea is you're trying to maximize utility, and their recommendation means sort of giving the highest expected estimated expected utility the item that has the highest expected utility. And also, there's a reward and the reward in the context of recommendation might be the rating from the user, for example, on that item or feedback, like thumbs up or thumbs down. So you can imagine a system that interacting with the user, is recommend something, that user gives feedback, and that feedback is used to update sort of the utility function, right? And this is kind of a exploration exploitation approach, because at the beginning, the system doesn't necessarily known enough and might need to give a variety of different items to the user. Some of u which may not get the highest rating for example from the user but then it learns from that that's sort of the exploration parts. And after a while then it uses whatever it knows to give as a recommendation the highest utility item and then there's a trade off between these. And you can do more exploration, more exploitation, and vary between these. This is sometimes the techniques that are used to do this are based on what's to be called the multi-arm bandit approaches where you sort of modeling It tracks the users and at the system. As kind of like pulling arms and arms of a slot machines. So we have a row of slot machines, you want to see in what order you want to pull the arms that give you the maximum reward. And when you ask to sort of explore for a while and see which machine is giving you a better odds And then sort of pool that more. And then every once in a while, you switch between this exploration, exploitation to exploration. So a similar model. The difference is that you will also now have to worry about the change that might occur in context. So I want to just say, one quick thing about the evaluation. I did just talk about that a little bit. But there are some challenges obviously and recommendation in evaluating context aware recommendation. When you're dealing with the representation of framework for instance, one big challenge is of course data sets, all right? For offline evaluation, because evaluation require users to have rated the same item in different context. And this is not very common, right? So if you look at let's say a typical data system we use in recommended system research, movie lens for example and so on, people don't generally go back and rate the same movie. And then record the context in which they rated it, right? So there are increasing number of data sets that are not available for this type of research, but a lot of them are based on user surveys. So they survey user arcade. Did rate this item in this context, did you rate it in that context? Did you eat this food in this context, did you eat this food in the other context? And then compile those kinds of data sets. And because they're based on user service, they're typically small in size and of course, they're also usually very, very sparse, right? And so that's part of a challenge of doing this work. As I said, there's increasing number of data sets now available since that situation was improving. In a context interactional models, the challenge for evaluating recommendation is that context are not known in advance, right? We are inferring the context. So you can't, you can only indirectly measure the impact of context awareness, right? So, you may build, let's say, you may learn your latent variable models, and then incorporate that to those topics, let's say your latent variable models as part of your recommendation algorithm. But at the end, you can only measure, let's say the accuracy of the system as a whole, and maybe compare it to a system that doesn't do this on that particular approach. You cannot point to specific elements of the context. But it does allow you to do certain types of reasoning with these latent variable models that are not otherwise easy to do in the context of explicit representation. And it also allows you to deal with systems that tend to be more dynamic than knowing everything about context beforehand. So, I guess that's where I'll stop. I'll be happy to answer other questions that you might have. >> So, how do we, How of these system is distinguished between the stable aspects of a users preference and the contextual things that change from that? Maybe I've really like some particular actor but any given movie that their in, I might bore less depending on when I'm watching it. >> Right, so I mean that is exactly the challenge. So it's the modeling parts or the learning part in this contextual modeling is exactly what figuring out that particular question. So you have to, for example, it's entirely possible that you have different contexts that you have defined, right? But it turns out that's the users end up or the particular user for instance end up getting similar ratings to the item. Even doing their in different context. So the challenging context in that situation actually is irrelevant, right? And you find datasets or domains that are like that where having context awareness doesn't necessarily add anything to the system. And you would know that by doing that kind of evaluation. So compare the context aware version of the system with, let's say the standard non contextual version and that tells you whether some context actually changes. You can actually measure this by using statistical machine learning methods to identify combinations of contextual conditions where rating actually changes. All right, as result of that content for the same item. So that's third of easy representational approaches to actually being able to figure out by doing some analytics on the rating or preference data. And the interactional setting is a little more difficult because don't actually use some specific set of conditions you are comparing to each other but there you have to do things like, change points detection for example. Just to try to understand if there's some significant change in the users behavior in a particular, let's say session, particular portion of a session than let's say, previous portion of a session. So by looking at change point analysis, things like that, you might be able to the size with these. >> Thank you very much for being with us. >> Sure, thank you.
