In this module, we're going to be talking about a general case of learning recommenders. So many algorithms are built directly from some algorithmic structure and insight into what might produce a good recommendation, such as find items similar to ones the user likes or decompose the matrix for latent features. And then we use metrics to to evaluate how well the algorithm did. But there are many other algorithms, including many modern algorithms, that use the metric directly in the process of building up the recommendation algorithm. And what they do is they learn the best recommender, within a family, for some particular metric. We've already seen this when we looked at the FunkSVD matrix factorization method, where we directly learned matrices Q or P and Q to minimize, Or have least a squared error. So what we want to do here is see how we can generalize that approach to many different kinds of recommender algorithms. The structure of learning a recommender algorithm is expressed in this formula. So we have theta, which is the set of parameters and or recommendation model, but often its a parameter such as for the FunkSVD the theta is going to be our P and Q matrices, That are going to be used to produce the recommendations. And then we have this function, sum function, which is either the error or the utility computing predictions or recommendations using a particular model and parameters. And what we do is if g is an error function then we compute the theta that minimizes g, argmin means find the theta for which g of theta is smallest. If it is a utility function that is it's high when the recommendations are good, rather than high when there's lots of errors, then we want to maximize g. The approaches are equivalent. It's just a matter of whether we're looking for the smallest error or the largest utility. There are three components of these kinds of algorithms. There's a model, such as matrix factorization or linear regression, or nearest neighbor. The model generally has some parameters, theta, the coefficients in a linear model, the matrices in a matrix decomposition. And the model, plus the parameters, produces the recommendations of the predictions. We then have a utility function or a loss function that measure how good or bad our output is with respect to some training data. And then we have an optimization algorithm that uses this to find the best parameters. And they're organized like this. So the green boxes are the pieces of this whole puzzle that we select as a part of the process of engineering the recommender. We have some model, let's say matrix factorization. And it has some parameters, PQ. Now the model will take the parameters and it's going to produce some output, predictions. The loss of utility function compares the output to training data ratings. And it computes how good or bad the output is with respect to the training data. The optimization algorithm takes that loss/utility or utility function and computes a new set of parameters to try. And we iteratively generate output using the model and parameters, measure how bad it is, improve the parameters, until we are satisfied with the parameter values that we have chosen. As we saw in FunkSVD, this is often through some kind of a convergence criterion, a threshold on how much the parameters are changing, or just we're going to train 40 times, 100 times, whatever. But these pieces work together to let us train the algorithm. For a simple example, we can look at estimating ratings with a single value. So we have our scoring function is going to be predict the rating with a single global value. So our parameters are the value B. Our error function we want to minimize the squared error. So we, the G is going to be the sum over the users and items of the error, which is the rating minus this baseline value. Now with this problem, our optimization algorithm is, look up that statistics tells us that b equals mu. Now we have the best value. There's not a lot of sophistication to the training. But the basic principle is there. We want to find the value that minimizes the error. Now this approach can be applied to many different kinds of models. We can do a bias model, which takes a global bias, mean rating, and then per user, per item, biases. And effectively what this gives us is a personalized mean. How much better or worse are you going to like this item than average, adjusted by how much better or worse you like items, on average. And in this case, the parameters are going to be b, b sub u, for every user, b sub i for every item. So if we have, 1000 users and 500 items, then we're going to have 1 plus 1000 plus 500 = 1501 parameters. That's a lot of parameters. If you want to apply the gradient descent approach that we saw in FunkSVD, that's 1501 derivatives. Fortunately, they all follow common patterns and a lot of things are going to be zero. But the parameters explode pretty quickly here. You can also use a linear regression that takes some features of users and/or items and puts them together in this linear combination in order to produce our scores, in which case, we want to learn the coefficients for each of these features. Or we can do matrix factorization, which we've seen worked out in detail, where we're learning matrix, or learning matrix values. We may also include the baseline values themselves in our learning process. So we've got these different models, we can then apply them to different metrics. We can look at prediction accuracy metrics like we've seen, RMSE, or simplifying that to the sum of the squared error. We could optimize for classifier accuracy, accurately classifying things as good or bad, consistent with the user's preferences. We often model this as what's called a logistic regression, which is a linear regression that's wrapped up in the logistic functions, that it's good for predicting boolean values. We can also optimize directly for rank accuracy, where we want to optimize the results of ranking to put good things towards the top. This generally requires some extra intermediate steps in order to actually make that optimizable. We're going to talk more about that in one of the later videos. And with the model and the error metric, then we need an optimization method. Gradient descent is very, very common in recommendation. It's also sometimes called steepest descent or ascent. There's another algorithm called expectation maximization that can also be used. Any optimization method for finding the best values, given a utility function may be applicable and some are going to be easier or harder to apply to different kinds of error functions and to different kinds of recommendation models. One common theme in a lot of these is they do use derivatives. So a lot of them need the derivative of the loss of the utility function in order to compute updates, because most of these are iterative methods, where they start with a guess for the parameters and then they see how bad the recommendations are with the guess and iteratively improve the parameters until we get to a minimum value for the error. To do this, we usually need to be able to take the derivative of the error. This makes it hard to optimize error metrics that aren't differentiable. So sometimes what we'll do then is we'll find a differentiable approximation to the error metric. Many different algorithms can be framed in this way. And to do this, we need to determine a quality measure. How are we going to judge good as far what the algorithm has produced? Then we need to devise prediction recommendation model and determine its parameters. We need to map all this to an optimization method that's going to find the best parameter values. We need to train, test, cross-validate, measure how good our final train recommender is. And it's useful to have a smooth differentiable error response that we can plug it into our learning algorithms. This approach is the heart of supervised machine learning, which is where we're trying to learn how to predict a bunch of examples that we have classified, such as things the user bought or ratings. Many modern recommendation methods are built on these principles, and it's a very general approach that can result in a wide array of recommender designs. Training these kinds of recommenders can be an expensive process. This iterative process of computing recommendations, measuring area and improving the parameters can take a while. And how interpretable the outcome is, how explainable the recommendations are, depends heavily on the model design. Some are going to produce very explainable recommendations. Others, it's going to be a little harder to explain to users where the recommendations came from. But the end resulting recommenders can be extremely effective.
