HelpThing:Recommendations

From WikiThing

The Library Suggester compares your entire library to the thousands of other member libraries on LibraryThing, and provides a list of some books you might like. Recommendations based on single books (people who have X are also likely to have Y) are available on each work's recommendations page).

[edit] Fiction vs. Non-Fiction

You can generate recommendations for fiction or non-fiction (based on information garnered from library records, not LT member tagging, which is why occasionally a fiction book might show up as non-fiction or vice-versa).

[edit] Recommendation format

All recomendations take the format:

Num. Title by Author.

Number of copies. Number of reviews. Average rating. Why?

Clicking on the title or the author will take you to the respective work or author page. Clicking on "Why?" will show what books you have that most strongly predict that you will like the recommendation. Clicking on one of these titles will take you to their work page.

[edit] Recommendations based on a subset of your library

There are four different algorithms that LibraryThing uses to calculate similarity. For the first three, you can choose to generate recommendations based only on a subset of your library via tags. For instance, if your reading tastes have changed over time, and you want to generate recommendations based only on what you read last year instead of based on your boxes of children's books, you could tag your books "recently read", and then restrict the recommendations to that tag. This could also be useful for personalized recommendations if you have multiple people's books cataloged in one account. Books that you own but that are not included in a tag subset will still be excluded from the recommendations.

[edit] Authors already in your catalog

If you want to increase your exposure to new authors - yes, you've already heard once or twice that there's a new Harry Potter book out that you haven't bought yet, stop recommending it! - you can choose "omit authors already in my catalog". To undo this, click on "show all". At the moment, there is no way to omit individual books from the recommendations (i.e. no "not interested" button).

[edit] People with your books also have

This algorithm generates recommendations based on patterns of book co-occurrence in libraries.

For example, let's say there are 1000 LT members. 500 of them have book X, and 100 of them have book Y. If these two books were unrelated and entirely randomly distributed, you would expect 50 members to have both. That is, if everything is random, 1/2 of any subset of LT members should have book X, so if the subset is "members with book Y", then we would expect the overlap to be 1/2 * 100 = 50. If significantly more than 50 people own both X and Y, that indicates that ownership is not random, and that people interested in one book would be interested in another.

What this algorithm does is sum this kind of calculation over all of the books in your library compared to all other books - which is why it can take a while to load!

[edit] Similarly-tagged books

This algorithm compares books not on the basis of similar ownership, but on the basis of similar tags, so you can get recommendations based more specifically on subject matter. It uses the most common tags applied to a work by all LibraryThing members, so even if you don't tag your books, you can still generate recommendations. It also weights based on tag obscurity - so the fact that two books share the tag "fiction" counts less towards making them similar than if they share the tag "books about books" or "dystopian steampunk".

[edit] Special-sauce recommendations

This algorithm combines elements of the previous two (shared ownership and similar tags), as well as some other mysterious factors and possibly pixie dust, and preferentially weights by book obscurity - so that even if you like young-adult fantasy, it's less likely to recommend Harry Potter and more likely to recommend a rarer book.

[edit] Most popular books you don't have

This one's pretty straightforward, which is why it doesn't work with a tag subset and is the quickest to load. You can see the top 1000 books in LibraryThing on the Books Zeitgeist page (here), and if you don't have it in your catalog, it'll pop up on this list. Also, the "Why?" link is missing from these recommendations.

[edit] Why don't recommendations take rating into effect?

Doing book-to-book, library-to-library comparisons to get recommendations is already computationally expensive. Adding ratings to the mix would exponentially increase the number of calculations, and hence the time, that the recommendations engine has to do its job. Additionally, only about 10% of books on LibraryThing have ratings - either people haven't read them yet, don't remember them well enough to rate them, or just don't bother rating any books in their catalogs. The idea of generating recommendations based on ratings is out there, but is unlikely to be implemented any time soon due to the computational costs.