Navigation

Turning White Wine Into Red: Recommendation Failures & The Alchemy Of Context

By on January 29, 2015 in Current System Paradigms, Featured

“[N]o event or object is ever experienced in perfect, objective isolation. It is instead subject to our past experiences, our current mood, our expectations, and any number of incidental details—an annoying neighbor, a waiter who keeps banging your chair, a beautiful painting in your line of sight. With something like wine, all sorts of societal and personal complications come into play, as well. We worry, for example, about whether our taste is ‘good.’ — New Yorker Magazine  quote about Columbia University neuroscientist Daniel Salzman.

Context rules … and it’s impossible to escape the biases it brings to the recommendation process.

Context is the sum of what we see, feel, smell and taste feel, as well as the effects that  surroundings and other people have on us. It is vital to the discussion and enjoyment of wine, music, books, films and many other products. It helps us to express why we like things, why other people enjoy things.

But context plays by its own rules and ultimately spoils the accuracy of  recommendations based on it.

Understanding the power of context is vital because it creates a powerful alchemy that can turn white wine red, make cheap wine expensive, make weak coffee strong and exert the same force on every other type of product recommendation.

Big Data And Face-To-Face Both Fail To Tame Context

Big data attempts to circumvent the power of context by creating even bigger masses of context. It accumulates and processes masses and masses of purchase data and personal information built around the basic idea that “people who bought this bought this” and that people who bought one product would buy a similar product.

Significantly, despite the massive amounts of purchase and personal information that big data captures and crunches for current recommendation systems, all it really has is general lifestyle context. And while that context can offer suggestions for inference, it lacks precision and solid information on satisfaction and future purchases. Loads of data may offer the illusion of accuracy, but suffers from the same biases as other forms of context.

Personalized face to face recommendations — “hand selling” — uses pretty much the same strategy  whether that is in that store or, as with wine, with a restaurant sommelier. Questions tend to be about, “what did you like about your last [fill in the blank] or an attempt to figure out what characteristics (sweet, hard, easy to use) might be matched to a new purchase.

Situational Context

Your perception of quality and your overall experience are colored by the context in which you experience something. In other words, the context in which you hear a song or drink a wine affect your opinion about its quality, as can the way others talk about them.

Did one hear a song at a concert, or through headphones? Was wine consumed at dinner with friends, or by itself at the kitchen counter after being brought home from the grocery store?

Context can alter perceived characteristics and change verbal descriptions and conversations regarding tannins, rhythm, whether something is soft or hard, and other characteristics that people find significant. Was a guitar riff particularly awesome, or perhaps the intense tannins and fruit made for an enjoyable glass of red?

The Great Divide Between Preference And Context

Context is a human construct that is the sum of all of the qualitative sensory experience connect with wine. This construct subject to countable biases just by genetics, education, culture. Itt is a uniquely individual experience that can only be experienced by one person at one time in one setting. It is analogous to the paradox found in those who study human consciousness.

While everyone in a group of people can taste the same sugar/water solution of a given concentration,  there’s no guarantee that any two people in that group will experience the same degree of sweetness. And it is likely that each person’s verbal expression of the taste will vary greatly.

Despite this, is possible for any of those people to express whether or not they like that taste or not.

This illustrates the divide between preference and context.

Because the  communication of a personal experience requires translation into words which are inherently inaccurate and bound by education, culture, etc.

The fact that even wine experts can describe the exact same fragrance and taste so differently as different fruit or different other characteristics, illustrates the inherent and accuracy.

Studies show that the ability to describe taste and smell sensations is heavily dependent upon education. this means that many people will offer many different descriptions of a given wine, which makes it very hard to make recommendations and decisions based on those contextual descriptions.

Price As Context

Price and expectations also work as context. Multiple studies have shown that even expert tasters who blind tasted the same wine twice rated it higher when told one was more expensive (When Does the Price Affect the Taste?).

Psychologically, we are led to believe that expensive is better.

In addition, there is a self-justification involved because we want to think we got our money’s worth on purchases. Regardless of whether the product is wine, a concert ticket, a new car, pricey wine glasses, or top-of-the-line headphones, our expectations tend to create the impression of that better experience.

For more on the psychological aspect of situational context, please see: Anxiety, Stress and Social Pressure Sabotage Choice

Context does have great value in conversation and social interactions, but if context is all you have for  recommending wine or music, then those recommendations can never escape its fundamental flaws and bias.

Music lovers and experts have no trouble deciding what they like or don’t like. For that matter, no one needs to be an expert or describe the context to know they like Coke better than Pepsi. In fact, if people tried to make decisions on soft drinks the way they do with wine or music, the results would be the same kind of inaccuracy, frustration, and confusion.

This piece from the New Yorker magazine offers more interesting insights into context: What We Really Taste When We Drink Wine.

Alchemy: White Wine Into Red

How can context turn white wine into red?

That was a question that researcher Gil Morrot, at France’s renowned international research center at Montpellier, set out to answer.

In a paper, “The Color Of Odors,” (see references for this and other papers, below) Morrot tested 54 wine tasters in their ability to describe the fragrance of two panels of wine: white and red.

The wines were actually the same. The only difference was that the “red” wine was actually the same white wine that had been artificially colored with a compound that was odorless and tasteless.

The study found that the panel described the wines differently. The artificially colored white wine was treated as a red wine.

Morrot’s paper cites other research showing that human beings have an excellent ability to detect and discriminate odors, but typically have great difficulty in identifying, or associating words with specific smells. He notes that humans will rely first on visual cues and modify subsequent perceptions based on that. These two phenomena, he says, are determined by the structure of the brain.

He also cites numerous other studies which show that opaque glasses hinder the identification of whether a wine is red or white and the fact that wine reviews use words and modifiers that tend to match the color of the wine being described.

Another paper, “The Odor Of Colors” confirms Morrot’s conclusions and finds — not unsurprisingly — that wine experts are better than the average drinker. This emphasizes the gulf that exists between expert descriptions and why they are so poorly understood by the average drinker.

Even if the expert is dead on accurate, the average wine drinker usually doesn’t have the capacity to translate the words he or she hears into something they can relate to their own experience.

The study does not address the issue of whether people become wine experts because they are better at this ability, or whether they are better at this ability because they became wine experts.

Strong Coffee Into Mild

Another published study, “Does The Colour Of The Mug Influence The Taste Of The Coffee?” notes that the color of a coffee mug affects perception of intensity, and that the color of a dinner plate will affect perception of the taste of the food on the plate.

What all of this illustrates is that matters of taste reflect a cross-sensory perception in which context can exert a powerful influence. In these cases, the cross-sensory perception issues are only one part of the overall context in which an experience is experienced and described. And this hobbles the effectiveness of context to make accurate recommendations.

Context Further Hindered By Misinterpretations, Genetics … More

Oral and written communication of experiences are inherently personal representations of internal perceptions. This means that context– along with genetics, experience, culture, vocabulary, education, and other factors — guarantee that even if two people were to experience the same feeling in their heads, their descriptions would be different.

Sadly, all recommendation systems in use today rely on attempts to use contextual information to try and make a recommendation. But, no matter how big the data gets and how complicated the algorithms grow, these systems can only make a guess because they do not develop clear, unequivocal expressions.

Solving Context Issues By Moving Beyond Inference To Expression

Recommendations need to move beyond context-based inference algorithms, filtering and clustering and instead, develop clear intent, perceptual-based recommenders..

Even the most accurate existing inference based recommendation engines can be substantially improved by overlaying intent-based results.

Further Reading

Backgrounders On Issues Of Internal Context