Clever Geek Handbook

What is the prediction function applied to Recommendations, Tanimoto coefficient for elementary CF is used

I am building a recommendation system that uses element based filtering. But I have a problem with the prediction function. I don't know what function can be used when calculating the similarity between different elements (Cinema) using the Tanimoto coefficient (Jaccard similarity coefficient) ?. the following example might explain my problem. Suppose User1 watched movie 1, and when we calculated the tanimoto ratio between movie 1 and all other movies, we found that the top 5 similar movies were 527 595 608 1097 and 588. Each of these movies has its own similarity to movie 1 how to follow:

User: 1

Watched Movie---Similar Movie----Tanimoto Coefficient score
          1--------527-------- = 0.33242
          1--------595-------- = 0.3377
          1--------608-------- = 0.3523
          1--------1097-------- = 0.3619
          1--------588-------- = 0.42595

So what's the next step after calculating the similarity? please i need help with this.

PS: I found that all the top 5 (527,595,608,1097 and 588) were viewed by user 1, so they cannot be considered a featured movie.

Many thanks

+3


source to share


1 answer


First, in both methods - from user to user and from element to element, we have defined two functions: similarity and forecast. Similarity measures for you how close to each other two objects (users or elements) are. In your case, Tanimoto was chosen. What you are missing is the forecast function. Since you have nearby entities (in i2i - elements), you have to predict the value of the rating (or in implicit user feedback - something happens). The simplest form is to use a weighted average function, where weight is a measure of similarity:

enter image description here

The average should only be calculated for items not rated by the user. This is one of the simplest solutions for a specific user using item2item.

A quick example. Having a rating matrix such as R:

xEIUy.png



We are trying to predict the rating for user 1 and point 1. Tanimoto's similarity measure is used in the calculation below.

enter image description here

So we predict that User 1 gives a rating of 1: 4/5.

For performance reasons, we keep indexing most of the similar Top-N items, but these items should still be new to the user for whom the recommendation will be generated.

+2


source






More articles:


All Articles