Has there been a change to Recommendations?

Am I going crazy, or has there been a change in the Recommendations that display when you click on a book? It seems that there are fewer than 10 recommendations for many of my books now. For example, when I look at my own copy of Rose Madder, I only see 3 recommendations now when I could have sworn there used to be 10. This is not a bug report, necessarily, but if something has changed, I'd like to hear about it. If I missed another thread on this, please point me to it.

I"m seeing three on that particular book, but ten on those I spot-checked in my own catalogue. They may have tweaked a display threshold, if you're sure you saw ten before for that same book.

You mean the auto-generated LT recommendations on the work page? We've been voting those up and down for a while now. I think once a recommendation reaches a certain level of negativity it gets dropped. You can vote and see other people's aggregate votes by clicking on the Rate Recommendations link at the top of the recommendations list.

>3 aulsmith: Yes, I mean the auto-generated recommendations. I'm seeing fewer recommendations on many of my books, not just the one I cited, so I was wondering if there had been some site-wide change.

Whereas I'm wondering why I suddenly have 185 new recommendations on my recommendations page.

>3 aulsmith:

I got the impression that sturlington saw a sudden change; rating recommendations started a long time ago.

>6 lorax: Right. But I started to see a decrease in December. I am only assuming it was related to ratings. It could, of course, be a bug.

>5 rarm: Which recommendation page? There are several. See the left side of the page to tell us which one you're on.

I think it was a sudden change in the last few days. A lot of my Stephen King books have fewer recommendations, and I look at those frequently.

Hmm, if it's limited to King books, it could be that another fan went through and down-rated a bunch of recommendations, pushing them below the threshold.

I'm looking through a bunch of my books now to see if I can see other instances. I own a lot of Ursula K. Le Guin and see changes to those as well. Many books still show 10 recommendations, though. I can't really detect a pattern.

ETA Seems like genre fiction (SF, horror) is more affected than literary fiction or nonfiction.

Okay, I say bug.

It doesn't seem to be related to ratings. I looked at the recommendations for The Stand, which now shows four on the main page and 184 on the rating page. On the main page, one of the books has a rating of 10 and another a rating of 1 and there are enough books with higher ratings to fill out the page.

When I first went to the work page for The Stand it showed me three LibraryThing recommendations, but after clicking on Rate Recommendations and then back to Compact it is now showing me 6 recommendations? Can anyone confirm that?

>12 jjwilson61: I just tried it. It was showing 3 when I first went to the page. I clicked on Rate Recommendations and then back to Compact and now it is showing 4. Should I report this as a bug?

Hey. No, there's been a change. For a while now we've been working on a new recommendation system, and generating them for past books. In turn, these trickle down to individuals' recs. It went live yesterday. The change will continue to roll out for some time. There's a major change in how recommendations are displayed coming soon as part of this.

I'll look at some of the bugs above.

I'll also report on some of the stats. While some books have fewer recommendations, many books have far more. And the overall quality is—we think—a good deal better.

More later.

Yeah, there's been a change in how same-authors are handled, which is causing the low number displayed in some contexts.

Basically before the recommendations themselves were limited by authors—indeed, limited in each category of recommendations and then limited when they were assembled into the "combined" recommendations. So, for example, a work could have no more than three recommendations for books by Stephen King. The exclusion was deep—at the data level.

Now, however, the limiting is supposed to be done on the display level. So it will store all the Steven King recs, but only show X number—or knock those after X back to lower rankings. This is a much better solution for a number of reasons. For starters it means the overall recommendation by member, which are assembled from these combined recommendations, are not being artificially hampered by a limit of three books by an author per work.

Anyway, I'm working on it.

Okay, as a temporary fix, more recommendations are being fetched for works. You can now see more on books like Rose Madder.

>15 timspalding: Thanks for clearing that up, Tim. I look forward to seeing how it develops.

>7 aulsmith: I meant recent recommendations/recommendations by date; usually there's a few new ones but today I got a big influx. And now based on Tim's explanation, I understand why.

I previously complained that I was seeing practically nothing but LGBT titles in the recommendations; now I'm seeing hardly any of those, and instead a lot of stuff related to the authors with the heaviest presence in my catalogue (e.g. I have over a hundred books by or about P.G. Wodehouse in my catalogue, and the top half-dozen recommendations on my list are all Wodehouse-related). Likewise for Georges Simenon and Terry Pratchett.

What is rather odd is that I only see 28 books on the first page of recommendations instead of 100, and there are gaps in the sequential numbering (1,2,3,4,21,24,31,32, etc.). The Filter is et to "none", so there must be a book-eater at work somewhere.

>19 thorold: Me too. But Tim said in >14 timspalding: above that changes in the display end are coming soon, so I'm hoping for some filters to do things like limit the number of the same author displayed on your list.

Tim, if you're listening it would be really nice to be able to get an option to show the books that you had previously said that you didn't want to see with an option to reverse that decision.

>20 jjwilson61: Agree, that would be nice.

Tim, I just want to say that I'm disappointed in the new recommendations I'm getting / how my recommendations have been reordered. 1. The recommendations are way more obvious: I own a book by ______, here's every book they wrote at the top of your recommendation list! 2. It looks like more emphasis is being put on all the books in all collections, less on ones I've actually rated. For me personally, it's disappointing because I list our whole family's collection, but only rate my own. I still got mostly recommendations that were MY recommendations, not my wife's or my kids'. Now, my top recommendations are heavily skewed towards my wife's stuff. I'm sure my situation is not the same as the majority of users, I just wanted to say that at least in my case, it's less useful.

Ratings have never been based on ratings but they are affected by what Collections you put them in. Each collection can be marked as contributing to recommendations or not so if you put your wife's or kid's books in their own collections then they can be excluded. Be careful though because if you also include them in some collection that *is* counted for recommendations like Your Library perhaps then they will still be used.

BTW, whether or not ratings should be a factor has been argued extensively. Here's one thread that I found, there are lots of others: http://www.librarything.com/topic/62630.

>22 waitingtoderail: I'm seeing this too. The percentage of books that I don't have because I've chosen not to read those series despite reading other books by the author has gone way up.

Yup, definitely much worse than before.

I don't seem to be getting more recs for books by authors owned, but I've got a definite spike in recs of mid-20C science fiction ...

>23 jjwilson61: That may be true, but I'm just saying, in MY case, the top recommendations have changed from ones which would be things I would be interested in to multiple works by authors my wife reads. I understand that including all our books together will lead to recommendations from their books, but this hasn't proved to be an issue until now. Whatever has changed in the algorithm has caused the recommendation engine to give more emphasis to unrated works, it just has.

Hold onto your hats. Here's what's improved and what needs improvements:

The only system throttled recommendation by author at the work level. So if book X was correlated to book A, B, C and D, the system chopped off the C and D. The new system throttles them at display only. This is a much more accurate, less arbitrary way of processing recommendations.

At present the recommendations for individual members are relatively more full of same-author recommendations, because the system is no longer able to rely on recommendations to be chopped off at two-per-author on the work level. It needs to be throttled back for user recommendations in a similar way—by display.

How is this better? Well, consider if lots of works are correlated to an author's works A, B, C and D. As things stood before, all the works that recommended A, B, C and D were forced to recommend only A and B. C and D were expunged from the data. This meant that C and D couldn't appear on the member level, because they weren't in the work level. If the user had A and B, they'd never heard about C and D. This was a distortion of what the data could say.

The old way did, however, have the virtue of limiting books by an author, albeit rather savagely. The new system needs to add author limiting of its own, not reply on the limiting at the original data level.

That's probably not clear, but it's the best I can give now.

On Monday a series of new display options will become available. I'll also make the old ones available. I will be looking forward to feedback on how much better or worse the recommendations are work by work . Then I can move on to aggregated, member recommendations. I'm sorry the system is too large to change in one fell swoop.

How is this better? Well, consider if lots of works are correlated to an author's works A, B, C and D. As things stood before, all the works that recommended A, B, C and D were forced to recommend only A and B. C and D were expunged from the data. This meant that C and D couldn't appear on the member level, because they weren't in the work level. If the user had A and B, they'd never heard about C and D. This was a distortion of what the data could say.

I'm not convinced that C and D were worthwhile recommendations, though. Because my recommendation list is suddenly full of box sets (or just fake lazy cataloguing entries) of books that I own.

Anyway, I'm working on it now. If you could see the error I found, you'd laugh your head off. (Basically, the first recommendation was treated as a negative, and all subsequent ones as a positive.)

Give me the night.

Recommendations of more works by the author I've already read are generally less interesting, though. If I have read A and B, then either I liked the books and am on the lookout for more books by that author (on author or series pages) or I didn't like those books and won't touch more.

>30 timspalding:
There's definitely something weird in the current implementation: if I do the "remove authors held" filter I go from (nominally) 1000 recommendations to 158! (And most of the 158 turn out to be books about authors in my collection). It definitely needs some kind of author throttle. The weight given to any author in my collections for recommendations should probably decrease as a function of the number of books by that author I own, at least above a certain threshold (say three or four books). If I already own twenty books by X, then I'm likely to be saturated with X-lore: I certainly know about the N-20 books by X I don't own yet, I've read all the biographies and critical works, and I know about the other authors who write in the style of X.

>32 thorold: If I already own twenty books by X, then I'm likely to be saturated with X-lore: I certainly know about the N-20 books by X I don't own yet, I've read all the biographies and critical works, and I know about the other authors who write in the style of X.

I don't agree with those last two statements.
I collect Terry Pratchett, but have not read any works about him, nor know about many other authors in his style.

I don't use recommendations. Is there a way to dismiss an author ?

Okay, I've got a first draft of a new member algorithm working.

It uses the combined recommendations as before. That means that it looks at all your works and gives points to every book they recommend. The higher the rank (i.e., the first recommendation, not the second), the more points it gets. If multiple works recommend a work, those points are also counted, but less and less with each subsequent recommendation, so your James Bond books don't take over everything.

In addition, it does the following post-value modifications.

* Downgrades books by authors you have a number of books by
* Downgrades books by authors showing up many times in your recommendations
* Downgrades books with work relationships indicating they are contained by or contain books in your library.
* Upgrades books publishing recently

All of these are progressive effects, not binary. So book number four by an author gets its "value" demoted by 5%, book number five by a bit more, etc.

After all that it sorts by value and chops off the top 1,000. Demoting means they'll show up lower in the ranks, or slip off the end at rank 1,000.

Thorod's no-author number is now 339. 82 of his recommendations are marked as new.

I didn't do ratings, but will come back and do them. There's some complicated math there, since both your ratings and other books' ratings need to be assessed according to the number of standard deviations from their mean, so differences in how you and others' spread your ratings across the stars don't mess things up.

I'm interested to see what people think of this, and to see how it chews up memory, but I haven't given it an extensive smoke test. I had hoped to avoid doing this until the new work-level recommendations were rolled out, but you guys seemed upset, so I got this out for now.

My goal is to get it looking good across many collections, and then provide you with tools so you can adjust how much you want to value:

* Newness
* Not being by authors you have-ness
* Not being by authors already recommended-ness
* How much your ratings should matter
* How much the books ratings should matter

>34 timspalding: I'm personally not very concerned about taking my own ratings into account. If a single book is an outlier in my collection, it shouldn't affect the recommendations very much anyway. If I have lots of books of a similar type that I keep reading even though I don't think they're great, maybe they should be in the recommendations after all.

For the books' ratings, I'd be happy if I could just set a minimum threshold myself and exclude any books below that rating (say, 3.8). If you want something more sophisticated, maybe you could upgrade and downgrade books based on how far their rating is from the average. So if the average rating is 3.8, a book with a 3.85 or 3.75 rating would get only a minor adjustment, while a book with a 4.3 average rating would get a significant boost and a book with a 2.7 average rating would hopefully not appear at all.

Also, I think it would be nice if the algorithm downgraded books in series that we've already started (i.e., where we have some of the series books catalogued already). If I've read the first few books of a series, I don't need the recommendations algorithm to suggest that I continue with it.

Take a look at yours, Zoe. I'm afraid that was is "new" may not work for you, because I used your books as a guinea pig, and regenerated them as I was working.

If you want something more sophisticated, maybe you could upgrade and downgrade books based on how far their rating is from the average. So if the average rating is 3.8, a book with a 3.85 or 3.75 rating would get only a minor adjustment, while a book with a 4.3 average rating would get a significant boost and a book with a 2.7 average rating would hopefully not appear at all.

Right. That's the point about standard deviations. You can't judge by ratings, because people have different average ratings. And you can't judge by distance from the average, because some people rate with a larger "spread" than others. So you need to judge by number of standard deviations from the average. The same applies both to your ratings and to others.

The combined recommendations on the work level now include ratings as a minor factor, using this logic. But your own ratings are not factored in.

And the "why" doesn't seem to be working.

Also, I think it would be nice if the algorithm downgraded books in series that we've already started (i.e., where we have some of the series books catalogued already). If I've read the first few books of a series, I don't need the recommendations algorithm to suggest that I continue with it.

Yes. The series logic is happening on work recommendations now. HP 2 won't recommend HP 1. (There are exceptions for large series, because the 154th Dummies book shouldn't be excluding the 122nd one out of hand.) But it's not downgrading books going forward. I think it should, but only ones that are 2, 3 and more books forward. It's good to have it recommending the next book to you. You might not know it's out.

Yeah. Sorry. I'll look at the why.

>37 timspalding: Well, it's definitely different, and better, than it was a couple of days ago. I'm happy that it does seem to be recommending books with higher average ratings than in the past. The main problem now is that there are too many series continuations.

The main problem here is that I want to change the system so that it does this work progressively and in the background. It's doing so much more now, that it's a little worrying both in terms of the time it takes to generate and the memory/load. Such are my issues.

too many series continuations

K. I'll work on that.

>39 timspalding: Yeah, I can see the value recommending the next book in the series in case we weren't aware of it, though ideally I think that would be done through a separate module than recommendations—it would be nice if we could specify which series we want to follow, and be informed of when another book has come out.

For recommendations, I do think the new-in-series books should be capped at one. In a series where I've read books 1-4, I don't really need to see #6 as my third recommendation and #5 as my 22nd recommendation and a companion book as my nineteenth recommendation.

In general, though, I'm very happy with where this is going.

ETA: I see you already addressed this point, thanks. So much cross-posting :)

>42 timspalding: Well, I personally don't need my recommendations to update that often, if that helps.

The "why" appears to be broken at the moment.

Overall, though, it looks a lot better than it did yesterday. There are no longer a bazillion Doonesbury books on there.

>39 timspalding: Yes. The series logic is happening on work recommendations now. HP 2 won't recommend HP 1.

This sounds dubious to me. There are plenty of series where the order you read them don't matter but it still makes sense to order them (Oxford History of the United States for example).

Also, the tag filters at the right don't seem to be working.

Yes. The series logic is happening on work recommendations now. HP 2 won't recommend HP 1. (There are exceptions for large series, because the 154th Dummies book shouldn't be excluding the 122nd one out of hand.) But it's not downgrading books going forward. I think it should, but only ones that are 2, 3 and more books forward. It's good to have it recommending the next book to you. You might not know it's out.

How does this interact with book-pairs that are in multiple, differently-ordered series? Something like the Narnia books, where there's a publication order and a chronological order, that aren't the same? Or something like the Discworld books where there is a "main" series and several sub-series?

I am seeing more interesting recommendations though. I'm currently excluding most series books beyond the first of a series from recommendations so I don't get overloaded by more of the same. When this settles down I'll add them back in and see what happens.

This sounds dubious to me. There are plenty of series where the order you read them don't matter but it still makes sense to order them (Oxford History of the United States for example).

Right. It's a problem. It only does it when the books are numerically ordered, and the series is show (e.g., below 7).

Okay, it is what it is now. I need to talk to CH and work on the display part. So problems you see—why and tags—will not get fixed in next six hours.

>53 timspalding:

No worries, just wanted to make sure you knew about them.

Can confirm that both the "why?" feature and the "filter by tags" feature are broken. (In the latter instance, I was told I didn't have any available recs based on tags at the moment. Really.)

Edit: Meant to say I had none based on the tag knitting, which is just... bull.

>55 AnnaClaire: Recommendations by tags has been broken for a while.
Bug report http://www.librarything.com/topic/182070

>56 MsMaryAnn:

That's a different issue - it was only happening for some people.

My recommendations are a lot more interesting - thank you!

Tim, it looks way better now! :)

My new top recommendation is to a "work" of Harry Potter and the Half-Blood Prince which I already have in my catalog but this work is separated from the main work because the editions in it have no isbn or author so it isn't clear if they belong to the book or the movie. Can something be done about that or should I just "No Thanks" it?

>34 timspalding:
Yes, it's looking a lot better. Still a lot of obvious things, reasonably enough, but not dominated by any one type of book as the previous iterations were. I even saw a couple of things in the list I didn't know about and might want to follow up. Thanks! :-)

Tim wrote:
So you need to judge by number of standard deviations from the average. The same applies both to your ratings and to others.

I presume this means that my ratings only count (for others' recs) if I've rated some reasonably large number of books?

>62 AndreasJ: And there's a reasonable spread in your ratings. If you've only used ratings to mark your favorite books with 5 stars then it will have no affect on ratings.

I've fixed all those recommendations to box sets. This was a problem on the work level. It's now looking at all potential recommendations, finding all contains/contained pairs, and throwing out the one with fewer copies. This does the trick.

It will take at least 24 hours to propagate.

>64 timspalding: Thanks!

A very minor oddity: the list shows 101 items per page, not 100. Page 1 is 1-101, page 2 is 100-200, etc. If you're not paying attention, you get the impression that a book is inexplicably appearing twice on the list.

Yeah, I'm going to be killing the display soon.

>67 timspalding: What are you doing with it?

Making it more graphical, optionally. And, again optionally, dividing it up into categories in the manner of BookPsychic.com.

>69 timspalding: Sounds good. I like covers :)

Although I'd also like a printable view that just had titles and authors.

While you're looking at recommendations:

When I set the "Remove authors held" filter there are still some books in the list by authors of books in my catalog.

Also, my big recommendations request would be for a filter by original publication date. The work page recommendations show the OPD, so the information seems to exist in some form.

Good ideas.

Okay, see here—the new features:

https://www.librarything.com/topic/187498

>60 jjwilson61: My list no longer contains the limbo edition of Harry Potter and the Half-Blood Prince. Did Tim jigger something or did it get combined with the book or the movie?

Limbo edition?

Limbo work would be more accurate. It contains all the editions that don't have an author or ISBN and so it can't be determined if it belongs to the book work or the movie work.

>78 jjwilson61:

Oh, right. That is the only manually removed recommendation in the system presently. Nothing recommends it.

Personally, I'm dubious about the work itself. But I didn't want to fight it.

Which would you combine it with then, the movie or the book?

This seems as good a place as any to mention this: At the same time as the surge in my recommendations, I've found that when you click "Why", it no longer lists any titles. Is there a way I can fix that?

Some info on movies:
https://www.librarything.com/topic/187552

>81 MiaCulpa:

Yeah, sorry. Working on it.

>82 timspalding: While you are working on recommendations can the feature "Filter by other members' tags" be fixed as reported here http://www.librarything.com/topic/182070. I used it often and really miss it.

Yeah. Sorry. It's on my list to fix.

About language. It may seem original language is given way too much weight when calculating recommendations? Maybe not apparant when looking at works originally written in English. But when looking at books from minor language groups, the recommended works often have nothing in common BUT original language.

Some examples, Swedish, Finnish and Norwegian:

• http://no.librarything.com/work/312089/recommendations/43347944
  
• http://no.librarything.com/work/5913777/recommendations/35250756
  
• http://no.librarything.com/work/449761/recommendations/40710389
  
  
  Other than that: Love the improved recommendations. Getting so many relevant suggestions I may end up cleaning out several bookstores!

>85 geitebukkeskjegg: I don't think it's original language per se causing the glut, rather it's that there's relatively few books in minor languages and the relatively few LT users who read those languages tend to form clusters with many similar books.

Still, actively suppressing same-language recs may be a way to deal with it. (Perhaps with a stronger effect the smaller the langage, LT-wise, is.)

>79 timspalding: A number of years ago, in a discussion of combining that had nothing to do with movies, you said something about there being times when combiners would just have to bite the bullet and mop up works with the same title when it wasn't clear what work exactly they belonged to. Of course, I'll never find where you said (or maybe implied it or possibly said something else entirely).

Anyway, a number of us have started to do what I call undifferentiated records where it's not really clear what the work the person with the entry meant. Personally I don't do them for movies (I lump the undifferentiateds with the book) but I do have them for other things, especially on voluminous classical authors.

So, if you don't want us to make these sort of limbo records, you should let us know and give us some other ideas.

>87 aulsmith:

Tim has a lot of ideas about combining that indicate he really doesn't understand the combining mindset. (He's suggested that we should combine abridgements for children with the primary work, for instance.) I wouldn't ask him to rule on this.

What would really solve this is the "low-quality data marker", of course.

>86 AndreasJ:

Yeah, it's not same-language per se. That effect CAN happen within the tag recs. But the tags have been assessed against same-language clustering, and it tones them down if they are. The major contributor to recommendations is the statistics of ownership, and there it's just math.

I'm not sure about suppressing same-author suggestions, or downgrading them. I can't see a way to do it that doesn't hurt good answers.

So, if you don't want us to make these sort of limbo records, you should let us know and give us some other ideas.

Yeah, it's hard. I don't really have an answer for you at present.

He's suggested that we should combine abridgements for children with the primary work, for instance.

I don't think so. If suspect I was misunderstood, or we're talking about something that wasn't that situation.

Also, the tag filters at the right don't seem to be working.

Now working.

>90 timspalding: Thank you!