TechCrunch published an article yesterday about the challenges of personalization and why no one has been able to innovate beyond what Amazon did 10 years ago. Leena Rao makes a good effort in trying to understand the challenges, mentioning the need for intent-based data, making sense of social, and privacy concerns. All are true. But the framework with which she’s approaching the problem is wrong.
The right way to look at this is by splitting the world of products into two: products that age and products that don’t.
- Books retain value over time. A book you wanted to read last year is something you’d still consider buying today (hence, the existence of airport bookstores). Same goes for movies, which is why Netflix beat Blockbuster.
- Fashion items (shoes, clothing, accessories) do not. Softlines (the retail term for fashion items) are extremely seasonal; items go out of style within months and unsold ones end up on the discount rack.
You’ll notice that successful personalization tech is tightly focused around items in the first category. Books, music, video, kitchen appliances, gardening equipment, (to a lesser extent) electronics - all things that Amazon’s recommendation algorithms are good at. (I would know, I was the product manager for that team). That’s because these products have a long enough shelf life to reach a critical mass of purchase data. You need dense datasets to do personalization right.
Where does personalization suck? The second category. To make it even more difficult, items in this category tend to be ones that you can look at and within half a second decide if you like it or not. They are visual, tactile, sensual. They are also highly individual - a watch that I love is also something you might hate, even if we share the same taste in movies. Hell, I might even love one watch but hate another that almost looks exactly the same. People shop in this category by gut feel and emotion, not by attempting to maximize a list of requirements and system specs. The result is a very sparse dataset with items going out of style too fast for the algorithms to become useful. What you end up with is least common denominator recs (like white socks and undershirts) that completely lack joy and delight.
The solution, like Leena points at, is social, although she gets it slightly wrong. I’ll follow up this post with my thoughts on how social can really make personalization work.
Last year, Stanford Professor Sebastian Thrun offered an online course on Artificial Intelligence for any internet-connected students from around the world, and close to 160,000 people took part in the course.
This year, Professor Thrun has quit Stanford and will be running a 7-week course…
TechCrunch published an article yesterday about the challenges of personalization and why no one has been able to innovate beyond what Amazon did 10 years ago. Leena Rao makes a good effort in trying to understand the challenges, mentioning the need for intent-based data, making sense of social, and privacy concerns. All are true. But the framework with which she’s approaching the problem is wrong.
The right way to look at this is by splitting the world of products into two: products that age and products that don’t.
- Books retain value over time. A book you wanted to read last year is something you’d still consider buying today (hence, the existence of airport bookstores). Same goes for movies, which is why Netflix beat Blockbuster.
- Fashion items (shoes, clothing, accessories) do not. Softlines (the retail term for fashion items) are extremely seasonal; items go out of style within months and unsold ones end up on the discount rack.
You’ll notice that successful personalization tech is tightly focused around items in the first category. Books, music, video, kitchen appliances, gardening equipment, (to a lesser extent) electronics - all things that Amazon’s recommendation algorithms are good at. (I would know, I was the product manager for that team). That’s because these products have a long enough shelf life to reach a critical mass of purchase data. You need dense datasets to do personalization right.
Where does personalization suck? The second category. To make it even more difficult, items in this category tend to be ones that you can look at and within half a second decide if you like it or not. They are visual, tactile, sensual. They are also highly individual - a watch that I love is also something you might hate, even if we share the same taste in movies. Hell, I might even love one watch but hate another that almost looks exactly the same. People shop in this category by gut feel and emotion, not by attempting to maximize a list of requirements and system specs. The result is a very sparse dataset with items going out of style too fast for the algorithms to become useful. What you end up with is least common denominator recs (like white socks and undershirts) that completely lack joy and delight.
The solution, like Leena points at, is social, although she gets it slightly wrong. I’ll follow up this post with my thoughts on how social can really make personalization work.
