Finding the perfect gift easily ranks as one the most stressful tasks of the holiday season. Success is cause for celebration: The recipient is surprised and impressed by how well you know her. Failure is cause for regret: Your missed-the-mark gift is met with a feigned smile and forced thank you. And failures not only take an emotional toll, but they’re bad for business: According to Deloitte, of the $64 billion spent through e-commerce on holiday gifts last year, $19.4 billion of it—roughly 30%—was returned. As much as technology has advanced in recent decades, it has yet to alleviate some of the challenges of gift giving. But now that we trust algorithms to create playlists, choose movies, and even recommend soul mates (with varying degrees of success), isn’t it time to set them loose on gift-giving?
Some companies have tried. In 2011, Etsy launched its short-lived Gift Finder, which recommended gifts based on your Facebook posts, profile updates, and “likes.” That same year, Walmart released Shopycat, which also relied on Facebook data, only to shutter the service after a couple of months. Then there was Hunch, a site that built a “taste graph” of recommendations by asking the user questions. Ebay purchased Hunch for around $80 million in 2011—and shut it down in 2014. Clearly, no one has nailed the machine-aided gift-selecting process just yet, either due to trying to mine the wrong data set (Facebook tells us too little about what you could want), or because we put the onus on the user. But with advanced data sets and machine learning, might now be the right time to give AI-gifting another shot?
The anatomy of the gift.
In order to develop the formula for the perfect gift, we first have to consider the age-old dynamics of gift-giving. A giver’s motives are usually pretty clear—to express fondness for and deepen the relationship with the recipient. If all goes well, the person on the receiving end feels appreciated, closer to the giver, understood, surprised, and delighted. (Yup, the exchange is all about social manipulation.) Where gifts go horribly wrong is when the receiver feels misunderstood, wrongly perceived, and possibly even insulted. How could you possibly think I’d want this?
The formula for the perfect gift begins to emerge: It’s a little about the giver (her taste, personality, and intentions). It’s a lot about the gifted (her personality, where she is in her life, and her aspirations). And it’s about their relationship (their shared experiences and common interests). Need plays a very small role, which is why socks almost always make for a crummy gift.
Now, let’s look at the emotional moments in the giver’s journey. First, she needs to define the gift by catching hints and storing the ideas as they come up. Then, she goes on the hunt—either online or in stores—to find the thing within her budget. She might track the delivery, wrap the gift when it arrives, and add a card expressing the sentiment or rationale behind the gift. She presents the fruit of her thoughtful efforts in person or remotely, and awaits the response as the wrapping paper is peeled back. The journey is a mix of emotional highs and lows—she gets excited over the initial idea, becomes anxious if she can’t find it in time or for the right price, and then is struck by uncertainty over whether the receiver will even like it. It’s also an emotional journey for the gifted, ending positively in delight or negatively in disappointment.
When thinking about an AI solution for gift prediction, those peaks and especially valleys represent great opportunities for algorithmic optimization.
The perfect automated gift.
Getting the right data.
Using a hybrid approach.
Teaching the algorithm.
To get the right data, we could look at two different kinds of inputs: Explicit and implicit. Explicit means asking the user to provide information by asking straightforward questions such as, How much do you want to spend? Who are you shopping for, and what do they like? When it comes to explicit inputs or affinities, you might ask something like, Does the person like sports? That’s a blunt instrument, because when you’re looking for spontaneity and discoverability, asking lowest common-denominator questions like that one negates the magic and leads to predictable or even offensive results. Since questionnaires tend to rely on demographics—the gender and age of the recipient, for example—they typically capture what is most generic about the person, rather than pointing the way toward gray areas, where a lot of the greatest surprises pop up.
Implicit inputs, meanwhile, are the ones we uncover by looking at users’ digital footprints. The best places to look for implicit data in the digital world are the ones that a user already curates—like Instagram, Pinterest, and Spotify—to form an aspirational image of herself. (These services may not expose their data to third parties, but they could certainly create great recommendations themselves if they chose to.) It’s rich data to mine, because not only do these streams contain overt information about the user’s tastes, they’re also often consciously made public by the user. In addition, there can also be metadata, such as the geographical tagging or time stamps that can serve as secondary clues to the user’s wishes.
Monitoring contextual signals as people interact with content and entertainment, though, can easily cross the “creep line,” the point at which a company’s use of personal data feels like a violation of privacy. Finding the relationship data—the stuff that connects the gifter to the gifted—is also tricky. You could look at Facebook exchanges, but people don’t often communicate with their true friends that way. WhatsApp, text messages, and email cross the creep line. That means that some of the information needs to come from the user.
So we’re not looking at a completely automated gift-giving solution, but at a machine-aided and user-curated automation process.
Relying on pure automation, without violating user privacy, would drop us into the generic gifting space, where the selections are safe, harmless, and uninspired. But we’d rather keep the magic than delegate the personal touch entirely to an algorithm to blah results. Argos, a U.K.-based retailer, offered an interesting hybrid approach, blending machine learning and personal curation: The user provided some key pieces of information, then the system offered recommendations, which the user then sorts through with a Tinder-like swiping gesture, providing feedback and inputs (more of this, less of that) before arriving at the right product. The feature has, unfortunately, been disabled, but it did offer an interesting model for training an algorithm with a user’s help.
The goals are to optimize the path to the perfect gift—and slash the gift-return rate. But there’s a critical obstacle to training models to achieve this outcome: The delight or disappointment at receiving a gift happens offline. That critical moment indicating success or failure is invisible to any algorithm written to automate gifting, and yet that’s a crucial piece of information to improve a recommendation engine over time.
So how can we figure out whether the algorithm has done a good job? We can look at a number of factors: Was the gift returned? Did the recipient brag about it on social media? We could send her a quick follow-up asking her, confidentially, just how much she liked it. We could see whether the (public) relationship status between the giving and the recipient changed on social media.
Ultimate success lies not only in making the gift giver’s life easier but in making the gift recipient happy. Without that last set of inputs—the stuff that will make the AI smarter—we might be doomed to repeating the same mistakes as the AI-aided gift services that came before. Remember, when it comes to training algorithms, it’s not just the thought that counts.
With research by Michael Krulwich