Pages

Friday, May 8, 2015

What Can We Learn from the Apps on Your Smartphone? Topic Modeling and Matrix Factorization

The website for The Burning House begins with a simple question:
If your house was burning, what would you take with you? It's a conflict between what's practical, valuable and sentimental. What you would take reflects your interests, background and priorities. Think of it as an interview condensed into one question.
But what about the more mundane and everyday stuff? As an example from an earlier post, I borrowed a quote popularized by the Iron Chef, "Tell me what you eat, and I will tell you what you are."

Do the choices we make reveal some underlying trait or situational demands that will enable us to predict future behavior? If one gathers up all the valuables and leaves behind the wedding album to burn in the house, can we guess more accurately using this information how long the marriage will last?


What about the apps on your smartphone? The list is long and growing with games, lifestyle, travel, music and entertainment, plus utilities, education, books, reference and business. Such a categorization imposes an organization that may not reflect the patterns we see in actual app download and usage. As an analogy, we can divide the supermarket into aisles marked with different signage (bread here in the middle of the store and butter over there along the wall), yet the market baskets rolled through the checkout reflect a very different associative network. Your phone is the basket, and your apps are the items purchased (or frequency of app usage equals amount of item bought). What is in your basket reflects your personal combination of wants and needs (e.g., a special trip before the big game on TV or your primary shopping for the week).

An alternative approach might treat this as a form of topic modeling with your phone as a document and app usage levels as frequency of word occurrences. Topics are the latent variables that generate the pattern of app co-occurrences (e.g., similar interests, needs or networks). The R package stm for Structural Topic Modeling may provide a gentle introduction for the social scientist, although Latent Dirichlet Allocation (LDA) still remains one step beyond the statistical training for most. This, of course, will change as more researchers are motivated to learn the mathematics given the promise of easier to use R packages.

Meanwhile, nonnegative matrix factorization (NMF) accomplishes a similar task with decomposition techniques from linear algebra. We start with the assumption that there are underlying usage patterns, such as, taking pictures, emailing them or posting to Facebook. Several apps are likely to achieve the same goal. The purposes that organize app usage are the latent variables. These same latent variables are also responsible for similarity among users. Users are similar because they use the same apps, but they use the same apps because they share the same motivations or purposes represented by the latent variables. The apps work together for like users to achieve the same ends.

To simplify, we can think of this as a joint factor analysis of the apps and a cluster analysis of the users. Thus, from a single data matrix NMF delivers two matrices: (1) "factor loadings" showing the relationship between the latent variables and the apps and (2) cluster membership weights reported for each user indicating the contribution of each latent variable to that user's app usage profile. We now have a matrix factorization or decomposition:

[ usage data ] = [ user latent profile ] x [ app latent profile ].

Do you play games on your phone to pass the time while waiting? These games are likely not the same ones you would play if you wanted to compete against others. More than one game can accomplish the same goal, so multiple apps "load" on the same latent variable. A user might do both at different times, therefore, that user belongs in both "clusters" (with each user cluster defined by a different latent variable). In NMF, the whole can be generated as the sum of the parts, which I have illustrated in my building blocks post.

Nielsen reports that less than 30 apps are used by the average smartphone owner. Can we agree that the usage data matrix will be sparse, given the number of available apps? However, we expect to find user-by-app blocks with higher densities. Such "clumping" occurs in high-dimensional spaces when users are heterogeneous with different groupings of wants and needs. These user clusters seek out blocks of apps that serve their purposes, creating joint clusters of uses and apps. This is what we uncover with NMF and LDA.

Note: My post on Brand and Product Category Representation ends with a list of examples containing the R code needed to run the R package NMF and perform the type of analysis reviewed here.

Monday, May 4, 2015

Clusters May Be Categorical but Cluster Membership Is Not All-or-None

Very early in the study of statistics and R, we learn that random variables can be either categorical or continuous. Regrettably, we are forced to relearn this distinction over and over again as we debug error messages produced by our code (e.g., "x must be numeric"). R will reminds us that if the function expects an argument to be a factor, our input ought to be a factor (although sometimes the function will do the conversion for us). Dichotomous variables do give us some flexibility for sex can be entered as a factor with values 'male' and 'female' or coded as numeric with values of 0 and 1 indicating degree of 'maleness' or 'femaleness' depending on whether male or female is assign the value of 1. Similarly, when the categorical variable has many levels, there is no reason not to select one of the levels as the basis for comparison. Then, the dummy coding remains 0 and 1 with the base level coded as 0s for all the comparisons (e.g., Catholic vs Protestant, Jewish vs Protestant, and so on).

Categories vs. Dimensions or Continuous vs. Discrete

The debate over psychiatric classification has bought the battle into the news, as has changes in the admission policies of women's colleges to accept transgender applicants. I have discussed the issue under both clustering and latent variable modeling. It seems just too easy to dissolve the boundaries and blur the distinctions for almost any categorization scheme. For instance, race is categorical, and one is either European or Asian, unless of course, they are some mixture of the two. I have borrowed that example and the following figure from a video lecture by Katherine Heller (also shown as Figure 1 in her paper).


We are familiar with finite mixture model from the mclust R package. Although I have shown only the contour ellipsis, you should be able to imagine the two multivariate normal distributions in some high-dimensional space that would separate Europeans (perhaps the blue ellipse) and Asians (which then must be in green). As geographical barriers fall, racial membership becomes partial with many shades or groupings between the ideal types of the finite mixture model.

Both the finite mixture model (FMM) and the mixed membership model (MMM) permit data points to fall between the centroids of the two most extreme densities. For example, a finite mixture model will yield a probability of cluster membership that can range from 0 to 1. However, the probability from the finite mixture model represents classification error, which increases with cluster overlapping. This is not unlike the misclassification from a discriminant analysis, that is, the groups are distinct but we are unable to separate them completely with the available data. The probabilities from the partial or mixed membership model, on the other hand, do not represent uncertainty but a span of possible clusters arrayed between the two extreme ideals.

The analysis of the voting record from United States senators presented toward the end of Heller's paper might help us understand this previous point. One might infer a latent continuum that separates Democrats and Republicans, but the distribution is not uniform. Democrats tend to bunch at one end and Republicans clump at the other. In between are Democrats from more conservative southern states and Republicans from more liberal northern states. One might argue that the voting dynamics or data generation processes are different for these clusters so that it makes sense to think of the continuum as separated into four regions with boundaries imposed by the political forces constraining their votes.

Interestingly, we learn something about the bills and resolutions in the process of accounting for differences among senators. Some votes are not strictly party-line. For example, senators from states with large military bases often vote the same on appropriations impacting their constituency regardless of their party affiliation. More importantly, the party accepts such votes as necessary and does not demand loyalty unless it is necessary to pass or defeat important bills. Legislation comes in different types and each elicits its own voting strategies (the switching interpretation).

Implementations of Mixed Membership Models in R

One of the less demanding approaches is the grade of membership (GoM) model. The slides from April Galyardt's NIPS 2012 workshop illustrates the major points (see her dissertation for a more complete discussion). R implements the GoM model with the gom.em function in the package sirt. For a somewhat more general treatment, R offers a battery of IRT mixture model packages ( e.g., psychomix and mixRasch). However, nonnegative matrix factorization (NMF) accomplishes a similar mixed membership modeling simultaneously for both the rows and columns of a data matrix using only a decomposition procedure from linear algebra.

Simply put, NMF seeks a common set of latent variables to serve as the basis for both the rows and columns of a data matrix. In our roll call voting example, we might list the senators as the rows and the bills as the columns. Legislation that yielded pure party-line votes would be placed at the extremes of a latent representation that might be called party affiliation. Senators who always vote with their party would also be placed at the ends of a line separating these pure types. We might call this latent variable a dimension representing the Democrat-Republican continuum, although the distribution appears bimodal. Any two points define a line, so we can always infer a dimension even if there is little or no density except at the ends.

Some votes demand party loyalty, but other measures evoke a "protect-my-seat" response (e.g., any bill that helps a large industry or constituency in the senator's state). Such measures would move some senators away from the edges of the liberal-conservative divide as they switched voting strategies from party to state. Alternatively, the bill may elicit social versus economic concerns, or provoke nervousness concerning a primary challenge. Each voting strategy will group bills and cluster senators by generating latent basis vectors. You can think of a NMF as a joint factor analysis of the votes (columns) and cluster analysis of the senators (rows). Each latent variable is a voting strategy so that senators who switch strategies depending on the bill will have mixed memberships as will bills that can be voted for or against for different reasons.

Wednesday, April 29, 2015

Modeling the Latent Structure That Shapes Brand Learning

What is a brand? Metaphorically, the brand is the white sphere in the middle of this figure, that is, the ball surrounded by the radiating black cones. Of course, no ball has been drawn, just the conic thorns positioned so that we construct the sphere as an organizing structure (a form of reification in Gestalt psychology). Both perception and cognition organize input into Gestalts or Wholes generalizing previously learned forms and configurations.

It is because we are familiar with pictures like the following that we impose an organization on the black objects and "see" a ball with spikes. You did not need to be thinking about "spikey balls" for the figure recruits its own interpretative frame. Similarly, brands and product categories impose structure on feature sets. The brand may be an associative net (what comes to mind when I say "McDonald's"), but that network is built on a scaffolding that we can model using R.

In a previous post, I outlined how product categories are defined by their unique tradeoff of strengths and weaknesses. In particular, the features associated with fast food restaurants fall along a continuum from easy to difficult to deliver. Speed is achievable. Quality food seems to be somewhat harder to serve. Brands within the product category can separate themselves by offering their own unique affordances.

My example was Subway Sandwich offering "fresh" fast food. Respondents were given a list of 8 attributes (seating, menu selection, ease of ordering, food preparation, taste, filling, healthy and fresh) and asked to check off which attributes Subway successfully delivered.


The item characteristic curves in the above figure were generated using the R package ltm (latent trait modeling). The statistical model comes from achievement testing, which is why the attribute is called an item and the x-axis is labeled as ability. Test items can be arrayed in terms of their difficulty with the harder questions answered correctly only by the smarter students. Ability has been standardized, so the x-axis shows z-scores. The curves are the logistic functions displaying how the probability of endorsing each of the eight attributes is a function of each respondent's standing on the x-axis. Only 6 of the 8 attribute names can be seen on the chart with the labels for the lowest two curves, menu and seating, falling off the chart.

The zero point for ability is the average for the sample filling in the checklist. The S-curve for "fresh" has the highest elevation and is the farthest to the left. Reading up from ability equals zero, we can see that on the average more than 80% are likely to tell us that Subway serves fresh food. You can see the gap between the four higher curves for fresh, healthy, filling and taste and the four lower curves for preparation, ordering, menu and seating. The lowest S-curve indicates that the average respondent would check seating with a likelihood of less than 20%.

What is invariant is the checklist pattern. Those who love Subway might check all the attributes except for the bottom one or two. For example, families may be fine with everything but the available seating. On the other hand, those looking for a greasy hamburger might reluctantly endorse fresh or healthy and nothing else. As one moves from left to right along the Ability scale, the checklist is filled in with fresh, then healthy, and so on in an order reflecting the brand image. Fresh is easy for Subway. Healthy is only a little more difficult, but seating can be a problem. Moreover, an individual who is happy with the seating and the menu is likely to be satisfied with the taste and the freshness of the food. Response patterns follow an ordering that reflects the underlying scaffolding holding the brand concept together.

Latent trait or item response theory is a statistical model where we estimate the parameters of the equation specifying the relationship between the latent x-axis and response probability. R offers nonparametric alternatives such as KernSmoothIRT and princurve. Hastie's work on principal curves may be of particular interest since it comes from outside of achievement testing. A more general treatment of the same issues enables us to take a different perspective and see how observed responses are constrained by the underlying data generation process.

Branding is the unique spin added to a product category that has evolved out of repeated interactions between consumer needs and providers skill to satisfy those needs at a profit. The data scientist can see the impact of that branding when we model consumer perceptions. Consumers are not scientists running comparative tests under standardized conditions. Moreover, inferences are made when experience is lacking. Our take-out only customer feels comfortable commenting on seating, although they may have only glanced on their way in or out. It gets worse for ratings on scales and for attributes that the average consumer lacks the expertise to evaluate (e.g., credence attributes associated with quality, reliability or efficacy).

We often believe that product experience is self-evident and definitive when, in fact, it may be ambiguous and even seductive. Much of what we know about products, even products that we use, has been shaped by a latent structure learned from what we have heard and read. Even if the thorns are real, the scaffolding comes from others.

Wednesday, April 22, 2015

Conjoint Analysis and the Strange World of All Possible Feature Combinations

The choice modeler looks over the adjacent display of cheeses and sees the joint marginal effects of the dimensions spanning the feature space: milk source, type, origin, moisture content, added mold or bacteria, aging, salting, packaging, price, and much more. Literally, if products are feature bundles, then one needs to specify all the sources of variation generating so many different cheeses. Here are the cheeses from goats, sheep and cows. Some are local, and some are imported from different countries. In addition, we will require columns separating the hard and soft cheeses. The feature list can become quite long. In the end, one accounts for all the different cheeses with a feature taxonomy consisting of a large multidimensional space of all possible feature combinations. Every cheese falls into a single cell in the joint distribution, and the empty cells represent new product possibilities (unless the feature configuration is impossible).

The retailer, on the other hand, was probably thinking more of supply and demand when they filled this cooler with cheeses. It's an optimization problem that we can simplify as a tradeoff between losing customers because you do not have what they are looking for and losing money when the product spoils. Meanwhile, consumers have their own issues for they are buying for a reason and may infer a means to a desired end from individual features or complex combinations of transformed features. Neither the retailer nor the consumer is a naturalist seeking a feature taxonomy. In fact, except for the connoisseur, most consumers have very limited knowledge of any product category. We are simply not familiar with all the offerings nor could we name all the alternatives in the cheese cooler or the dog food aisle or the shelves filled with condensed soups. Instead, we rely on the physical or online displays to remind ourselves what is available, but even then, we do not consider every alternative or try to differentiate among all the products.

Thus, the conjoint world of all possible feature combinations is strange to a consumer who sees the products from a purposefully restricted perspective. The consumer categorizes products using goal-derived categories, for instance, restocking or running out of slices for your ham and Swiss sandwich. Thus, attention, categorization and preference are situated within the purchase context defined by goals and the purchase process including the physical product display (e.g., a deli counter with attendant is not the same as self-service selection of prepackaged products). Bartels and Johnson summarize this emerging view in their recent article "Connecting Cognition and Consumer Choice" (see Section 3 Learning and Constructing Value in Context).


Speaking of cheese (in particular, Brillat-Savarin cheese), we are reminded of the above quote popularized by the original Japanese Iron Chef TV show. Can it be this simple? I can tell you what is being discussed if you give me a "bag of words" and the R package topicmodels. R-bloggers shows how to recover the major cuisines from a list of ingredients from different recipes. My claim is that I learn a great deal by asking if you buy single wrapped slices of processed American cheese. As Charles de Gaulle quips, "How can you govern a country which has 246 varieties of cheese?" One can start by identifying the latent goal structure that shapes awareness, familiarity and usage.

Much is revealed by learning what music you listen to, your familiarity with various providers in a product category, which brands of Scotch whiskey you buy, or the food choices you make for breakfast. In each of those posts, the R package NMF was able to discover the underlying latent variables that could reproduce the raw data with many columns and most rows containing only a few responses (e.g., Netflix ratings with viewers in the rows seeing only a small proportion of all the movies in the columns). Nonnegative matrix factorization (NMF), however, is only one method for uncovering the hidden forces that structure consumption activities. You are free to select any latent variable model that can accommodate such high-dimensional sparse data (e.g., the already mentioned topic modeling, the R package HDclassif, the R package bclust, and more on the way). My preference for NMF stems from its ease of use and successful application across a diverse range of marketing research data as reported in prior posts.

Unfortunately, in the strange world of all possible feature combinations, consumers are unable to apply the strategies that work so well in the marketplace. Given nothing other than hypothetical products described by lists of orthogonal features, what else can a respondent do but rely on the information provided?


Wednesday, April 15, 2015

Recommending Recommender Systems When Preferences Are Not Driven By Simple Features

Why does lifting out a slice make the pizza appear more appealing?
We can begin our discussion with the ultimate feature bundle - pizza toppings. Technically, a menu would only need to list all the toppings and allow the customers to build their own pizza. According to utility maximization, choice is a simple computation with the appeal of any pizza calculated as a function of the utilities associated with each ingredient. This is the conjoint assumption that the attraction to the final product can be reduced to the values of its features. Preference resides in the features, and the appeal of the product is derived from its particular configuration of feature levels. 

Yet, pizza menu design is an active area of discussion with many websites offering advice. In practice, listing ingredients does not generate the sales that are produced by the above photo or even a series of suggestive names with associated toppings. Perhaps this is why we see the same classic combinations offered around the world, as shown below in this menu from Bali.


I am not denying that a vegetarian does not want ham on their pizza. Ingredients do matter, but what is the learning process? Does preference formation begin with taste tests of the separate toppings heated in a very hot oven? "Oh, I like the taste of roasted black olives. Let me try that on my pizza." No, we learn what we like by eating pizzas with topping combinations that appear on menus because others before have purchased such configurations in sufficient quantity and at a profitable price for the restaurant. Moreover, we should not forget that we learn what we like in the company of others who are more than happy to tell us what they like and that we should like it too.

The expression "what grows together goes together" suggests that tastes are acquired initially by pairing what is locally available in abundance. It is the entire package that attracts us, which explains why pizza seems more appealing in the above photo. If we were of an experimental mindset, we might systematically vary the toppings one pizza at a time and reach some definitive conclusions concerning our tastes. However, it is more likely that consumers invent or repeat a rationale for their behavior based on minimal samplings. That is, consumer inference may function more like superstition than science, less like Galileo and more like an establishment with other interests to be served. At least we should consider the potential impact of the preference formation process and keep our statistical models open to the possibility that consumer go beyond simple features when they represented products in thought and memory (see Connecting Cognition and Consumer Choice).

Simply stated, you get what you ask. Features become the focus when all you have are choice sets where the alternatives are the cells of an optimal experimental design. Clearly, the features are changing over repeated choice sets and the consumer responds to such manipulations. If we are careful and mimic the marketplace, our choice modeling will be productive. I have tried to illustrate in previous posts how R can generate choice designs and provide individual-level hierarchical Bayes estimates along with some warnings about overgeneralizing. For example, choice modeling works just fine when the purchase context is label comparisons among a few different products sitting on the retail shelf. 

Instead, what if we showed actual pizza menus from a large number of different restaurants? Does this sound familiar? What is we replace pizzas with movies or songs? This is the realm of recommender systems where the purchase context is filled with lots of alternatives arrayed across a landscape of highly correlated observed features generated by a much smaller set of latent features. We have entered the world of representation learning. Choice modeling delivers when the purchase context requires that we compare products and trade off features. Recommendation systems, on the other hand, thrive when there is more available than any one person can experience (e.g., the fragmented music market).

To be clear, I am using the term "recommendation systems" because they are familiar and we all use them when we shop online or search the web. Actually, any representational system that estimates missing values by adding latent variables will do nicely (see David Blei for a complete review). However, since reading a description of the computation behind the winning of the Netflix Prize, I have relied on matrix factorization as a less demanding approach that still yields substantial insight with marketing research data. In particular, the NMF package in R offers such an easy to use interface to nonnegative matrix factorization that I have used it over and over again with repeated success. You will find a list of such applications at the end of my post on Brand and Product Category Representation.






Friday, April 10, 2015

Modeling Categories with Breadth and Depth

Religion is a categorical variable with followers differentiated by their degree of devotion. Liberals and conservatives check their respective boxes when surveyed, although moderates from each group sometimes seem more alike than their more extreme compatriots. All Smartphone users might be classified as belonging to the same segment, yet the infrequent user is distinct from the intense who cannot take their eyes off their screens. Both of us have the flu, but you are really sick. My neighbor and I belong to a cluster called dog-owners. However, my dog is a pet on a strict allowance and theirs is a member of the family with no apparent limit on expenditures. There seems to be more structure underlying such categorizations than is expressed by a zero-one indicator of membership.

Plutchik's wheel of emotion offers a concrete example illustrating the breadth and depth of affective categories spanning the two dimensions defined by positive vs. negative and active vs. passive. The concentric circles in the diagram below show the depth of the emotion with the most intense toward the center. Loathing, disgust and boredom suggest an increasing activation of a similar type of negative response between contempt and remorse. Breadth is varied as one moves around the circle traveling through the entire range of emotions. When we speak of opposites such as flight (indicated in green as apprehension, fear and terror) or fight (the red emotions of annoyance, anger and rage), we are relating our personal experiences with two contrasting categories of feelings and behaviors. Yet, there is more to this categorization than is express by a set of exhaustive and mutually exclusive boxes, even if the boxes are called latent classes.



The R statistical language approaches such structured categories from a number of perspectives. I have written two posts on archetypal analysis. The first focused on the R package archetypes and demonstrated how to use those functions to model customer heterogeneity by identifying the extremes and representing everyone else as convex combinations of those end points. The second post argued that we tend to think in terms of contrasting ideals (e.g., liberal versus conservative) so that we perceive entities to be more different in juxtaposition than they would appear on their own.

Latent variable mixture models provide another approach to the modeling of unobserved constructs with both type and intensity, at least for those interested in psychometrics in R. The intensity measure for achievement tests is item difficulty, and differential item functioning (DIF) is the term used by psychometricians to describe items with different difficulty parameters in different groups of test takers. The same reasoning applies to customers who belong to different segments (types) seeking different features from the same products (preference intensity). These are all mixture models in the sense that we cannot estimate one set of parameters for the entire sample because the data are a mixture of hidden groups with different parameters. R packages with the prefix or suffix "mix" in their title (e.g., mixRasch or flexmix) suggest such a mixture modeling.

R can also capture the underlying organization shaping categories through matrix factorization. This blog is filled with posts demonstrating how the R package NMF (nonnegative matrix factorization) easily decomposes a data matrix into the product of two components: (1) something that looks like factor loadings of the measures in the columns and (2) something similar to a soft or fuzzy clustering of respondents in the rows. Both these components will be block diagonal matrices when we have the type of separation we have been discussing. You can find examples of such matrix decompositions in a listing of previous posts at the end of my discussion of brand and product representation.

Consumer segments are structured by their use of common products and features in order to derive similar benefits. This is not an all-or-none process but a matter of degree specified by depth (e.g., usage frequency) and breadth (e.g., the variety and extent of feature usage). You can select almost any product category, and at one end you will find heavy users doing everything that can possibly be done with the product. As usage decreases, it falls off in clumps with clusters of features no longer wanted or needed. These are the latent features of NMF that simultaneously bind together consumers and the features they use. For product categories such as movies or music, the same process applies but now the columns are films seen or recordings heard. All of this may sound familiar to those of you who have studied recommendation systems or topic modeling, both of which can be run with NMF.

Friday, March 20, 2015

What Consumers Learn Before Deciding to Buy: Representation Learning

Features form the basis for much of our preference modeling. When asked to explain one's preferences, features are typically accepted as appropriate reasons: this job paid more, that candidate supports tax reform, or it was closer to home. We believe that features must be the drivers since they so easily serve as rationales for past behavior. Choice modeling formalizes this belief by assuming that products and services are feature bundles with the value of the bundle calculated directly from the utilities of its separate features. All that we need to know about a product or service can be represented as the intersection of its features, which is why it is called conjoint analysis.

At first, this approach seems to work, but it does not scale well. We create hypothetical products and services defined by the cells in a factorial experimental design (see the book Stated Preference Methods Using R). The number of cells increases quickly with each additional feature so that we need to turn to optimal designs in R in order to limit the number of possible combinations. We have reduced the number of hypothetical descriptions, while the number of estimated parameters remains unchanged. Overall preference continues to be an additive function of the values attributed to each of the separate components.

Representation learning, on the other hand, is associated with deep neural networks, such as the h2o package discussed by John Chambers at the useR! 2014 conference. According to Yoshua Bengio (see his new chapter on Distributed Representations), "a good representation is one that makes further learning tasks easy." The process is described in his first chapter on Deep Learning. As shown in this figure from Wikipedia, the observed features are visible units and the product representation is a transformation contained in hidden units.


What do consumers learn before deciding to buy? They learn a representational structure that reduces the complexity of the purchase process. This learning comes relatively easy with so many sources telling us what to look for and what to buy (e.g., marketing communications, professional reviews, social media and of course, friends and family). Bengio speaks of evolving culture vs. local minima as the process for "brain to brain transfer of information." Others refer to it as a meeting of minds or shared conceptualizations.

Are you thinking about a Smart Watch? Representation learning would suggest that the first step is "getting a lay of the land" or untangling the sources of variation accounting for differences among the offerings. I outlined such an approach in my last post on precursors to preference construction. It is possible to go online and request side-by-side feature comparisons that look similar to what one might find in choice modeling. However, that step is often late in the process after you have decided to purchase and have narrowed your consideration set. Before that, one looks at pictures, scans specifications, reads reviews and learns from others through user comments. One discovers what is available and what benefits are delivered. As you learn what if offered, you come to understand what you might want and be willing to spend.

The purchase task is somewhat easier than language translation or facial recognition because product categories are marketing creations with a deliberately simplified structure. Products and services are simple by design with benefits and features linked together and set to music with a logo and a tagline. Product and service features are observed (red in the above figure); benefits are latent or hidden features (the blue) and can be extracted with deep neural networks or nonnegative matrix factorization. That is, we can think of representation learning as the relatively slow unsupervised learning that occurs early in the decision process and makes later learning and decision making easier and faster. Utility theory lacks the expressive power to transform the input into new ways of seeing. Both deep neural networks and nonnegative matrix factorization free us to go beyond the information given.

Finally, what happens when the consumer is pulled out of the purchase context and presented feature lists constructed according to a fractional factorial or optimal design? The norms of the marketplace are violated, yet respondents get through the task the best they can using the only information that you have provided them. Unfortunately, you do not learn much about bears in the wild when they are confined in cages.