I was going through my list of recommended books, and the book "Secret Life of Bees" was recommended based on two other books being books I have read (Five People You Meet in Heaven and Water for Elephants). Since I loved both of those books, I went to the LibraryThing page for Secret Life of Bees. I thought I'd give the "Will I Like It" feature a try. Supposedly, that feature is very confident that I will not like Secret Life of Bees. Why would LibraryThing be recommending something that it says I won't like?

But then, just for fun, I went to the LibraryThing page for Water for Elephants to see what that Will I Like It feature says if I would like the book, given I have already rated it. It is again very confident that I won't like the book despite my high ranking of the book.

So, obviously, the Will I like It feature isn't very accurate, and probably should be eliminated--I mean, what if people actually make decisions to read a book based on that? Or at least get rid of the fact that it says it's very confident about it. Maybe its confidence level should always be very low or something....this is actually a disservice to those who may want to use it.

This thread (http://www.librarything.com/topic/66776#...) in "Recommended Site Improvements" should clarify the issue somewhat.

On a general note though: you can never entirely trust any recommendation - automatic or not, on LT or elsewhere.

The book-to-book recommendations are based on books you have. And you have those two books.

Will I Like It is based on your whole library. As your library mostly consists of science fiction and fantasy, it was compared with other people with similar libraries. Those people do not, apparently, often have The Secret Life of Bees.

So, what we're talking about is the difference between recommending by a single book, which has potential problems, and recommending by some average of your whole library, which also has problems.

Actually Tim, the way I read it was that he first tried the Recommendations feature on the Recommendations tab, which recommended The Secret Life of Bees, which he doesn't have. Then he went to The Secret Life of Bees Will I Like It feature and was told he wouldn't like it. Both of these features as I understand it use your whole library not a single book.

I'd support the removal of Will You Like It.

The difference between the Recommendations list and Will You Like It is that the recommendations list shows me books I may not know about. If I'm already looking at a book, I don't need a computer to draw the obvious conclusion that if I own a lot of fantasy, and this book is fantasy, then I'll "probably like it".

It feels like the Recommendations list is leading me to new things, while Will You Like It actually discourages me from branching out--because if it isn't obviously the kind of book I read all the time, I probably won't like it. I think LT can be better than that. It should be about finding obscure connections and recommending books that aren't obvious, rather than telling us to stick with what's safe and familiar.

Tim, I think there needs to be some sort of pop-up or explanation with the features.

People are basically using a hammer and expecting it to be a screwdriver - and not understanding that the "will you like it?" isn't precision like the book to book recommendation tools.

Personally, I think it's clear, but when individuals are asking to remove a functioning feature based on their individual outliers, I think some clarification is probably in order.

But even if it were something like:

"Will You Like It" is based on the presence of books in your library - and just judges whether or not you'll like a book based on your current library. So, if you have a library full of biographies, knitting books and sports mysteries and you wonder why the bar says you definitely won't like Wuthering Heights, it's because your library shows nothing similar.

I dunno - or even if it were a when to use explanation:

"Will You Like It?" - Use it if you're wondering how similar this book is to your current library!

Book To Book Recommendations - you've just read the best book ever and you want to find something very similar to read, but don't know where to start.

I vote for keeping it. I think it's fun to see if it matches what I think of a book *after* I read it.

Yeah, yeah. I know it's supposed to precede the reading, but that's not how I use it.

I would also be sorry to see "Will You Like It?" go. It sometimes (often?) gets things wrong, but it's still fun.

Yeah, I like it too. Come on people, use your sense of humor.

>4

Ah. Not quite. The recommendations tab uses all your books, but it's only polling the book-to-book recommendations for them. That is, it's not doing anything other than aggregating all the book-to-book recommendations. It weighs them a little, and it shuffles them, but it's just aggregating the book-to-book recommendations.

Maybe it should do more, or maybe it shouldn't. It doesn't because the math is horrendous. Will I like it attempts a holistic calculation on your whole library, but it does it for only one book and its result is a number, not a complex data type. And it still takes a few seconds or more to calculate. Doing it on all your books would require a military supercomputer.

and not understanding that the "will you like it?" isn't precision like the book to book recommendation tools

Neither are precision. There is no precision in this sort of thing. It's an algorithm, just like Amazon's, just like anyone's except that Amazon never ever tells you the full answer—they never use the calculation to tell you whether you might like a book.

Anyway, the problem isn't even the algorithm. If the user in question had gone to the world's smartest librarian or bookseller and said "Will I like this book?" the librarian/bookseller might well have asked "What else do you read?" If the user had replied "I mostly read science fiction and fantasy" the librarian/bookseller might not have recommended The Secret Life of Bees.

I'd support the removal of Will You Like It.

No.

Message edited by its author, Jul 4, 2009, 11:34am.

Just adding my voice to posts 7, 8 and 9 above - I really enjoy the "Will You Like It?" feature and would be sad to see it disappear from the site.

I think "Will You Like It?" is a fun feature and the sometimes strange results will often give me a much-needed chuckle.

>6 Or what about just a rewording of it? Rather than "Will you like it?", something like "Is this your usual type of book?".

I understand how the feature works, I just don't agree with the principle behind it. I don't think people will only like books that are the kind they always read, and I don't think telling them that they won't like anything different is helpful. You could go to the world's dumbest librarian too, and say, "I only read science fiction and fantasy; will I like this crime thriller?" and they could say "Probably not." But what's the point? You don't have to ask anyone to know the answer to that question. LT should be the smart librarian who can help people with non-obvious questions, not the one who gives the safe answer and discourages people from branching out.

A book like the one mentioned by the OP is exactly the sort that LT should be highlighting--something that isn't their usual type, but that the recommendations algorithm suggests anyway. Instead, people see a different kind of book on the recommendations list, and consider giving it a try, but are immediately told by LT that they won't like it just because it's not what they usually read. I just don't see how this is a service.

Sure, it can be fun for the powerusers to see how accurate or inaccurate it is after the fact, but people who don't understand exactly how all parts of LT work might actually expect it to be reliable.

#12

Me too. Fun (often funny) to see what it says.

I thought the idea was whether people with libraries similar to yours owned the book?

people who don't understand exactly how all parts of LT work might actually expect it to be reliable.

I agree with stephmo that, in light of this, some on the spot info might be a good idea.

I also like "Will You Like It". It's like a coin-operated fortune teller. Fun to play with, but I wouldn't actually make decisions based on it.

It's Zoltan.

18: Thanks for that link, and all that it leads to.

v

16> I'm a fan of Magic 8 Balls myself. Must be why I like the feature :-)

>18

Are you saying I'm a dwarf?

Will You Like It is a fun feature. I even enjoy it when it tells me that I will love "Spammy Spam Spam".

> 21: I have reason to believe there is at least one munchkin in your household.

Sure its fun for power users, but Tim keeps telling us he's trying to do right by the ordinary user not just the power users.

I agree with 12 and 22 - the entertainment value is a good thing.

>5It should be about finding obscure connections and recommending books that aren't obvious

I suspect that creating an algorithm for this would be very, very difficult. It's just not the sort of thing computers are designed for - it requires a human brain.

Whew! good to know Will You Like It will remain.
I think its fun. Recommendations may not be precise,but any list of books
provides me with reading considerations.

Frankly, I don't care about the hows and whys, nor do I expect miracles.
Even I don't know what books I might like from one day to the next.

I like Will You Like It, it's just a fun feature. I'm as pleased when it says I will like something I hate as when it says I will hate something I know I liked.

#22 Sylvia - It told me that I would love Spammy Spam Spam too!

If I want a reliable opinion I usually ask Booksloth, her recommendations have never let me down yet.

I also enjoy "Will you like it?" and would hate to see it go.

>25 Not really. Something like that could be accomplished just by combining the basic recommendations with Will You Like It--for each book on the recommendations list, check against the Will You Like It value; the ones with a low value (ones that we'll supposedly dislike, but which are actually just not our usual kind of book) are the unexpected and therefore interesting ones.

I think having a brief explanation is a good idea, though, or, better yet, a link that says "What's this?" in small letters, which would lead you to a more thorough explanation.

I don't think people will only like books that are the kind they always read, and I don't think telling them that they won't like anything different is helpful. You could go to the world's dumbest librarian too, and say, "I only read science fiction and fantasy; will I like this crime thriller?" and they could say "Probably not." But what's the point? You don't have to ask anyone to know the answer to that question. LT should be the smart librarian who can help people with non-obvious questions, not the one who gives the safe answer and discourages people from branching out.


But that's not what it is.

Features like this -- I'm not saying "Will I like this" is one -- can be scarily accurate. The old Alexlit, which was really Web 2.0 before its time, did this. They can identify the particular crime thriller that someone who reads a lot of SF and fantasy will like. They can see that a lot of people who read SF also read Patrick O'Brian. I'd rather see a library-average comparison than book-to-book. Any idiot can tell that if I liked book 1 I'll like book 2, and that's all that book-to-book gets me. I don't think I've ever come across an unexpected gem from my "Recommendations" tab.

Really, for anyone thinking that predicting the accuracy of liking a book is "easy" - there's a cool million dollars just waiting for you:

http://www.netflixprize.com//rules

I know it's not books, but you should be able to apply your logic to movies no problem. Honestly, there's a reason Netflix is doing this - and is willing to publicly give out data sets for people to dig into in hopes of obtaining a methodology that will result in a mere 10% increase in the betterment of it's predictor and pay 1 million dollars for it.

Because predicting the human brain is very difficult.

We have a lot of fun toys on LT that pick and predict in different ways. We have to remember that when we ask for predictions to predict unpredictable events that we're really better off asking people for random suggestions and explanations as to why they're recommending it.

_Zoe_, I think you and I see eye to eye on the issue...

The Algorithm should be more based on how you've rated similar books. I don't need to know that books like The Secret Life of Bees, Water for Elephants, and the Five People You Meet in Heaven aren't sci-fi or fantasy.

But when I've rated a book (such as Water for Elephants) five stars, and then the algorithm tells me that "it's highly confident" that I won't like Water for Elephants, something is wrong. If I went to ANY librarian--the smartest in the world or the dumbest--and said, I really loved Water for Elephants, will I like The Secret Life of Bees, they'd probably be able to tell me.

And man, I'd hate to be so boring that because I read a lot of one type of book that I can't branch out into other types of books as well.

I don't think it's too much to ask that all the features talk to each other to enhance the usability and interconnectiveness of LibraryThing.

Message edited by its author, Jul 5, 2009, 8:55am.

>33 You're simply using the wrong tool, though.

If you loved Water for Elephants and want to find out if you'll like the Secret Life of Bees and it's a total outlier in your library, you're better off doing a book-to-book recommendation.

http://www.librarything.com/work/522063/...

Because when you go there - it's easy to see that the 4th recommendation is The Secret Life of Bees. Doing the thing you described would be exactly these results. It would also completely destroy the functionality of "Will You Like It?" as it works today just to make you happy on your outlier book.

The truth in your "smartest or dumbest librarian" scenario is that if you started out, "I've never read anything but science fiction or fantasy before, but I read Water for Elephants and loved it, do you think I'll like the Secret Life of Bees?" question is that you'd likely get one or two follow-up questions. Because they'd want to know what it was that you liked about Water for Elephants to make sure that you'd probably really still like the Secret Life of Bees.

Algorithms don't and can't ask follow-up questions. Algorithms also don't make judgments about boring or exciting. You are assigning that judgement to that tool. But the good news is that if this is the beginning of a trend and you do like both books and start adding more to your library, algorithms adjust. And soon, you'll find that it adjusts it's opinion of whether you'll like it or not. Or, it could be that it will have been an outlier and you'll never add another similar book and it will have been right all along. It won't make you boring - it will just be your taste.

How does the computer know what a book is about. Someone has to give it tags to tell it what it is about. The computer will not read the book and analyze it? It won't know that books have similarities unless we provide it with clues that it does. So people who have books in similar tags in their library generalizes. Without the tags Would it think that Because I have a Charles Dickens in my library, say Great Expectations, that someone who has A Christmas Carol in their library, also by Dickens should have the book recommended to me? If I rate Great Expectations and many don't rate books at all I as recall, the books are different obviously.

So is it suggesting similar books based on how are they defined, or books based on who owns what...

A mix of who own what, and which books have similar tags.

If your tags are differently defined by you then others, you would end up with things you have no desire to read. You would have to resort to a standard definition of a tag...

I don't think the system only looks at how you tag books, but rather at how certain books are tagged on LT-wide scale.

>38

Right.

is that if you started out, "I've never read anything but science fiction or fantasy before

Ha! I have you there: It's not the only thing I have read before, have never stated that's the only thing I have read, nor would I ever say that because it would be inaccurate. Perhaps the fallacy of what I'm coming up against. Are most of my books that way? Perhaps, but I have enough background to enjoy multiple different genres of books. Maybe I'm more diverse than most people. In fact, I'm pretty sure of it, which is why your very basic premise doesn't apply to me.

But do your collections reflect that? - which otherwise would be the equivalent of telling the librarian you only read science fiction and fantasy while that isn't true.

I'm surprised that people take the 'Will I like it' feature so seriously. It is fun. I don't skip over a book just because the line says 'No!' I read the reviews and descriptions to make my decision. I often us 'Will I like it' after the fact.

>40 The thing is, I looked at your library. This is the data that the feature can use - and that's precisely the data that you're giving it. Now you want the data to assume facts not present. 200/292 of your books either have the science fiction or fantasy tag. That's 2/3rds of your library. Over half your books are in the Star Trek series - it doesn't exactly scream "The Secret Life of Bees" fan, now does it? The algorithm doesn't have a mystery, "but take into account books that they have also read and likely enjoyed that aren't listed" function.

If Tim could program that into his algorithm, he'd be a bazillionaire.

I'm already a brazillionaire.

That's a lot of bras! Why so many?

Because this is LiBRAraryThing!

If someone starts cataloging bras, I'm going to hurt you.

Will this pass muster?

http://www.librarything.com/work/149992/...

LOL

That's a pretty awesome book from a LibraryThing perspective—good tags, good recommendations, etc.

Yes, but will we like it? ;-)

Well, LT thinks I will love it!

But LT thinks I won't like it. Ha, shows you what they know.

The recs are pretty good too - I have seven of the ten already. (Love those green check-marks!)

Message edited by its author, Jul 5, 2009, 6:28pm.

I won't like that book either, but i like whats in em...

And here, Kindle just won't do...

Message edited by its author, Jul 5, 2009, 7:28pm.

One point that appears to have been glossed over, the OP mentions that he's rated his book 5 stars.

Unfortunetly this has zero relevance to any of the recommendations or Will you like it features.

TIM, Please can you add in some rating weighting factor on at least some of these. Owning a book - especially now that collections includes unowned books - does not necessarily correlate well with liking it.

That makes sense, for the subset of people who take time to rate a book, we have about 12% of books rated? (Items rated) So if you are a rater, should it be easier to figure out if you like a book?

Hmmm...so basically it seems that some people are saying that "Will you like it?" should be renamed "Similar to your library?" or something of the sort so that it's clearer how the app works. While I would miss the brevity and fun of "Will you like it?" I can see their point somewhat. (Not that I really believe that a computer program can predict what I will like or dislike with any kind of certainty.)

Also reading_fox and DWWilkin bring up a good point. At the very least, how about if you rate a book one star (or 1/2) then it doesn't show you recommendations based on that book? At least on your Recommendations list. Or does this already happen and I've never noticed because I tend to look at recommendations only when I'm on the work's page?

>58 Conversely, you could easily create a "Stinkers" collection and simply mark it "Do Not Use for recommendations." Heck, I had my first abandoned book for the year and realized I should create a collection for that exact reason.

The other issue with the lack of ratings (only 10% is an issue) is that Tim's shared time and time again that the majority of the ratings are clustered around 4 and 5 stars. Yes, you can find 1 and 2 star ratings, but you'll come across a TON of 4 and 5 star ratings.

The point? The majority of things that are rated are liked. So you can extrapolate that owned is liked without taking the programming complexity of ratings into account.

The work for your wish is huge. If you put 10 rating data points to cover the half stars and then an 11th for what to do with no rating, this is no small task. It's not exactly an "if then" statement. And that's just for a shot at the 10% chance that there's data. Keep in mind, it's not just your rating that matters - it needs to be compared against normalized ratings that have a 10% chance of existing for other books. When, in fact, 90% of the ratings will fall back to the "meh, if it's like their library, they like it" algorithm.

And, to be honest, NONE of this would help the OP. His Water for Elephants is still an outlier. Why 5 stars? Does he like circuses? Tragic stories with uplifting endings? Runaway bestsellers? Historical novels? Depression era novels? Is that author a cousin? Let's face it, if it's a Depression era fixation and not a general historical novel thing, the Secret Life of Bees could be a huge miss. Five stars won't mean a thing - and it won't make anyone a bad person.

You'll never know until a few more similar books start to trickle in and form a better picture.

The further point is that, even if you don't like a book, if you own it for a reason—it didn't fall out of the sky onto your shelf—then it's still a data point.

For example, I own Serenity: Those Left Behind, a graphic novel that fills in the time between the Firefly series and the Serenity movie. It sucks. I hate it. But it's still a valuable data point. I'm an insane Firefly fan—I'd buy almost anything. If LibraryThing recommended another Firefly-based book, I'd probably get it. Hence, the low rating means the opposite of what you'd expect—it means I want related stuff all the more!

Just don't become a leaf on the wind.

>24 Sure its fun for power users, but Tim keeps telling us he's trying to do right by the ordinary user not just the power users.

Tell that to this guy!

Message edited by its author, Jul 6, 2009, 11:14pm.

Ha! (to both of you!)

#59 - that only works if you remove that book from all other collections that you do want recommendations from. ie Your Library, owned, unowned but read etc this is inconvenient!

I knew someone was going to to come back with the 10% fallacy, it's been wrong ever since Tim first suggested it. The overall rating rate on LTs long tail might be 10%, but for most users it's way higher. I've done the sums for The top Zeitgeist books: they are rated by 30% of peope, 3% at less than 3*

That is 10000 people recieving recommendations for works they don't like. Hardly a small number of LTs users.

Regarding work - well yes it would be some, so is every suggestion made. I think this would be worth it. I would use the rating as a weighting on the popularity - ie if I rated a book 0.5, then it's only counts as 1/11th of a real book. Although ideally it would compare my rating against everybody else's rating. This I can appreciate might be a bit database intensive.

Message edited by its author, Jul 7, 2009, 5:41am.

Also, I would be likely to start rating many more of my books if that actually did something to my recommendations.

Here you go, Tim. Just read this and all of your problems will be solved! Effect of User Tastes on Personalized Recommendation

Sorry, had to fix the link.

Message edited by its author, Jul 8, 2009, 11:32pm.

60> Huh. It doesn't rain books up in Maine? We get those type of storms in Lowell all the time. (I think it's Massachusetts calling "come back Tim, come back.")

In any case, I have quite a few books that aren't much of a data point. I picked up books by lot at auctions and added quite a few books to my library that were pretty random. I guess they weren't completely unfiltered, but nigh so. For another example, I buy used Esperanto books sight unseen, five dollars a pound. I don't have Parizina because I'm a Byron fan; I have it because it came to me with a bunch of other short pieces in an envelope. Also, I buy a number of books because I can potentially scan them for Project Gutenberg; while pre-1923 material obscure enough to be already done is a set of data points, it's not one that's really useful for LibraryThing.

Of course, I have about 5% of my library rated, so that doesn't help me much. I sort of plan to create a "core library" collection and a "books I just sort of own" collection one day, but that takes time.

Message edited by its author, Jul 10, 2009, 10:01pm.

Thank you, #58, for making me feel better about my "Least Favorites" collection. I had originally meant it as a further kick in the teeth to the books I'm sorry I paid for, because one star was too good for them. Having them all excluded from recommendations makes me seem nobly practical-minded, WOOT.

The "will you like it" feature always seems to tell me I'll hate things that weren't written by Orson Scott Card. But, the confidence level is always "very low," so whatever.

(my disliked books also aren't much of a data point, except perhaps that you can learn that I can't throw books away or sell them... I kept every book I hated from college. Every last one. White Noise and Willa Cather and frakking Edward Carr included.)

>68

I have a really hard time getting rid of books as well (or I wouldn't have 2000+ of them I guess...)

I'm curious about your "Least Favorites" (mine is called "Fail (aka Anti-Favorites)") because they actually look interesting to me - were you forced to read these (for class or whatever) or did they look interesting to you and then let you down?

I'm a big OSC fan (for his books - which I love, not his politics - which I don't) so I find it hard to believe we are LT-opposites (but I have been wrong, oh, just sooo many times before.)

Message edited by its author, Jul 13, 2009, 1:32am.

3> Perhaps it could start recommending by groups of books; if it notices* that a bunch of books are all recommending the each other, then there's probably other books that show up in combination with those books but don't come up on the recommendations when looking individually. Thinking locally, a longer (hidden?) list of recommendations for each book could be looked at in those cases, for repeated values.

* For an irrelevant tangent of the day: using noticed here forced me to stop and conclude that, yes, using anthropomorphic phrases for computers is well-established acceptable practice, even if we know that a computer doesn't just notice anything without looking* for it.

70: I brought that up with Tim (13-book series that only recommends other books in series for top ten recs), but he thought it would be too server-intensive.

I really think it would be worthwhile to have longer recommendation lists, at least for books with enough copies that it would be meaningful. I'd say the recommendations are one of the most important features on LT, so it would be worth the server cost.