One Slope to Rate them All

RecSys Week 2: A CF schema for cool guys, and items.

If I gave 3 stars to a MacBook Air, how many would I give to a MacBook Pro? And if i gave 4 start to a MacBook (that’s a thing now), how many would you say that I’d give to the MacBook Pro now? Let’s say on average a MBP get’s 1 more star than the MBA, and the MBP get’s 0.5 stars than the MBP. In this scenario the answers to the ratings could be 4 and 4.5.

These predictions were calculated by adding the average difference between the models, to ones I had already rated. This is the idea in which the Slope One Method, the recommender system proposed by the authors, is based on.

The method basically predicts the rating of an item with a linear function with slope one, you guessed, using as argument the rating of another item, just like in the example. But in that case, two different ratings where generated, which one do I choose? There are different functions to integrate the information, some just take the average and others use a weighting influenced by the amount of ratings.

I think that a big plus in the work of the authors is the proposition of a weighted scheme that allows to keep the simplicity and also take into account the amount of ratings, which makes recommendations more trustworthy.

A problem is the space, since we are not looking at similarity between items, but the difference, it is hard to determine which one has more weight when making a prediction, so usually the whole matrix of deviation must be stored. This also means that it can be quickly updated when adding new ratings, so this method is in a middle point between space and efficiency.

To conclude, I’d like to add that in the experiments the authors got performances very similar to the methods that were commonly used and recap that this system uses the intrinsic popularity of the items to generate predictions.

Thank you for reading, any correction, commentary or related reading I’ll thankfully accept. This post is a commentary on the paper Slope One Predictors for Online Rating-Based Collaborative Filtering (2007), written by Daniel Lemire and Anna Maclachlan. This is a weekly post for the Recommender System course (IIC3633), Pontifical Catholic University of Chile.

Related posts:

The age-old discussion: breast V bottle. If you ask me, there is only one right answer. That answer? Whatever method works for you and your baby is the only “right” way for you! As a certified…

Remember those times when we had to get our Tablets, iPad’s, and Laptop’s during KHDA week back in school? “Bring Your Own Device” (BOYD) week was the best. Although we started off getting a bit…

SEO is the procedures or techniques through which you can appear in Google on the first page and in the first result in the search engines such as Google search engine, Yahoo and Bing … etc, for…