Mahout recommendataionon Implicit data with multiple factors. - mahout

Question

Could you please provide me some details on mahout recommendations using data with multiple factors.? I have data with user id, book, language, category etc. Say suppose, a person read the book with category as thriller, in french language. Now considering all those facts i need to recommend a book to him. Could you please give me some insight on picking the right path.?

Answer 1

Just the thing for Mahout 1.0 where we create models for a search engine to index and query.
The models are called indicators and are lists of similar items for each item. Similar in the sense that they were purchased by the same people. This is the essence of a cooccurrence recommender.
The collaborative filtering data is the book read or the ID. If you recommend a book you can show other IDs with the same titles for multiple formats (ebook, recorded, papeback etc.) The metadata can be used to skew recs toward a certain category. The language is probably a filter unless you think your audience are usually multilingual.
Create the CF type indicator by feeding purchases into Mahout 1.0 spark-itemsimilairty. out will come a list of similar books for each book. Index those in a search engine. Then the simplest query is the user's history of books purchased. This will yields unskewed recommendations as an ordered list of books.
Now to skew results towards the user's most favored category index the categories for each item in a separate field in the index. So the index has a field for "indicators" and one for "categories". The "docs" are really items/books in your catalog. the skewed query is (pseudo-code):
query:
field: indicators; q: "book1 book2 book3 book10" //the user's purchase history
field: categories; q: "user's-favorite-category user's-second-favorite-category"
field: language; filter: "list-of-languages-of-books-the-user-has-purchased"
You can put as many categories in the query on that field as you wish, perhaps all the user has purchased from. Note use of a language filter, you may want to use this as a skewing factor rather than a filter. In this way you can seamlessly integrate collaborative filtering recs skewed or filtered by metadata to get higher quality recs. Any metadata can be used that you think will help.
BTW you will get even better recs if you add in other actions you have recorded like views of book details. This will call for a specially processed indicator called a cross-cooccurrence indicator and is also calculated by spark-itemsimilairty. In fact you can include just about any action the user takes--the entire clickstream as separate cross-cooccurrence indicators. This will tend to greatly increase the amount of collaborative filtering data you can use in making recs and therefore improve quality.
This idea can even be extended to actions on items that are not books, like categories. If a user purchases a book they also, in a sense, purchase a category. If you record these "category purchases" as a secondary action and create a cross-cooccurrence indicator with them you can use them both to skew results and as a purchase indicator. The query would look like this:
query:
field: indicators; q: "book1 book2 book3 book10" //the user's purchase history
field: category-indicators; q: "user's-history-of-purchased-categories"
field: categories; q: "user's-favorite-category user's-second-favorite-category"
field: language; filter: "list-of-languages-of-books-the-user-has-purchased"
Read about spark-itemsimilarity here: http://mahout.apache.org/users/recommender/intro-cooccurrence-spark.html This includes some discussion of how to use a search engine (Solr, Elasticsearch) for the index and query part.