From The New Publishing Standard:
Over at Book Riot, Arvyn Cerezo takes us through the process and then explains why they will still recommend a book you have absolutely no interest in.
Machine learning systems called recommender systems, or recommendation systems, use data to assist users in finding new products and services … These algorithms, however, need a decent amount of data to choose a recommendation strategy in order to produce meaningful and personalized recommendations. This data may include past purchase histories, contextual data, business-related data, user profile-based information about products, or content-based information. Then, all of these are combined and analyzed using artificial intelligence models so that the recommender system can predict what similar users will do in the future.
All very clever, but…
The limitations of content-based filtering include its inability to comprehend user interests beyond simple preferences. It knows some basic stuff about me, but that’s as far as it can get. What if it recommends a racist book? What if it recommends a book that might trigger readers without some heads-up? What if it recommends a book that is problematic? The keyword is nuance, and algorithms can’t tell the difference between two books that have similar stories.
And don’t we know it? Fifteen or more years buying books on Amazon and it will still recommend books I would eat shards of glass than read.
I always figured that was just Uncle Jeff getting revenge for one of my less complimentary posts about Amazon, but it seems in fact it’s just that the recommendations system is as useless today as it was fifteen years ago.
Cerezo concludes:
“With all the pitfalls of algorithms — and AI in general — it seems like nothing beats book recommendations done by an actual human being. They are more accurate and more personal. Most of all, you can also find hidden gems that you really like rather than the bestsellers (and what everyone’s reading) that these machine learning systems always spit out.”
Two points arise.
First, “rather than the bestsellers (and what everyone’s reading) that these machine learning systems always spit out” is fundamental to the problem. Algorithms – especially for a commercial operation like Amazon – have the sole purpose of selling more books. They and the company do not give a flying fig about our personal preferences.
Link to the rest at The New Publishing Standard
BooHoo, Amazon presents books it thinks that the person who signed in will want to buy based on their past buying, browsing and searching habits.
As far as “personal preferences” are concerned, PG supposes that some people have “personal preferences” in books that they don’t want to buy or read or do something with, but is Amazon somehow required or expected to understand someone’s personal preferences that have not been reflected in their previous and current activity on Amazon?
If PG was as concerned about Amazon and his personal preferences, he would open a new Amazon account and be careful not to let anyone else use it. Within a few weeks, Amazon would understand PG’s personal preferences by what he did on the site with the new login ID.
As far as “book recommendations done by an actual human being,” without being a snob about it, PG has never met a person working in a bookstore who would have been likely to give him a good and precise suggestion for a book that PG would like to read. The most PG has ever received is something like, “Our twentieth-century history books are over there,” or “Fantasy and Science Fiction is on aisle three.”
To be fair, if PG in his current instantiation ran into PG at age thirty working in a bookstore, current PG doubts his thirty-year-old self would understand much about PG, the elder’s preferences in books.
If PG was good friends with a bookstore employee and had spent hours talking about books with that person, the results might be better if PG showed up when the bookstore was open and the employee was working at the time.
If you want something done right, you have to do it yourself.
That certainly applies to finding interesting books.
I just read the new Katie Porter book in one sitting. It was not recommended to me, I tracked it down myself.
Goodreads, where I have shelved nearly 2500 books, has never once recommended me a single book I’m interested in, I think because the whole of their process is “one person who read this book you read also read this other, completely dissimilar book, would you like it?”
Amazon does slightly better, but I’m sure used to do much better.
My friends don’t do particularly well either, though.
(Incidentally, your CAPTCHA is… over the top.)
Even my close friends, who read, and have a pretty good idea of the sorts of things I like, only have about a 50/50 success rate in recommending books to me. And vice-versa, I suspect.
I just took a look at my recommendations: something I have not done in several years. They still rely on the technique of “This author whose book you bought: Did you know they have written other books as well? Amazing!” This is not the worst technique, but there isn’t much value added here. Then there are the books that are in fact in my “saved for later” list. I suppose the idea is to remind me that they are there, but that is a different thing from recommending a book. If I am looking for recommendations, the value added is zero. Finally, there is a small number of books that categorize well with something I have bought. I recently, for example, downloaded the memoir of the executive officer of the USS Wahoo, a very successful submarine in WWII. So there in my recommendations is the memoir by the same officer, but as commander of the USS Tang. This actually makes sense as a recommendation. While it is a “this guy wrote another book!” recommendation, wartime memoirs aren’t the sort of thing you necessarily expect multiple books, so I had not thought to look.
This all is pretty typical of my experience. There are a few good hits, but they are buried under a pile of junk. Even there, better search categorization would be more useful. I searched for World War II submarine memoirs and got, in addition to the usual ads and the books somehow related to the topic but not in fact what I asked for, a couple of actual hits well down the list.
And yes, the recommendations are worst than useless for fiction.
Why are they bad at recommending books? Are they supposed to be good? The non-fiction recommendations I get usually make sense, if one goes strictly by subject matter. But for fiction I never, ever forgave Amazon for a mistake they made 20 years ago when they thought I would enjoy Gormenghast because I enjoyed Lord of the Rings. *Shrug*.
In all seriousness, fiction is trickier, because a combination of elements go into stories. Not just plot elements, but a style of writing, an approach to themes, and how well the story is told. Some stories have a certain je ne sais quoi that just resonates. Details, a specific type of tree in a forest. It’s harder to pinpoint down simply going by genre / subject matter, and I honestly think it may be asking too much of a machine to get it right in this scenario.
That said, I get why the OP (or anyone) might think a bookstore can nail this, because YouTube manages to do so with music. At least for me it does. But I interact with songs differently than I do with e-books: I play the entire song on repeat, or a section of it, and I hit the like button and bookmark it. Whereas, if I’m going to re-read a book I tend to buy the dead-tree edition so I can re-read the parts I like best — random access for the win! Except that usually just results in me re-reading the whole thing anyway 🙂
Not only that, but the e-book equivalent to playing sections of a song on repeat might be highlighting text, which some people do. Some people, not me. That’s a side effect of me using mental rather than literal bookmarks in dead tree books, hence the preference to random access. So, if you never re-read e-books, never highlight them, or do anything besides hitting the like button, the algorithm might not have as much to work with. Just a thought.
I have spent a lot of money on Amazon recommendations. Amazon likes me. They call me up and we talk about books and the current state of American literature. Don’t they call you?