How to create a recommendation system for a movie?

Question

• What is the best approach?
• What are the algorithms used? What are their strengths and weaknesses?
• Why current recommendation systems for reproducing failures while providing good recommendations?

Answer 1

This is a very open question, which includes many different concepts.

As a starting point for discussion, we consider the k-closest neighboring algorithm . It is widely used in issues similar to your movie maker. One of the big problems with this algorithm is the human contribution to determining how many dimensions you use to segment your spatial space and select the properties of each of these dimensions so that each adds a value rather than duplicates the value of another dimension.

Directly related to the k-NN algorithm is a cluster analysis field. When you define data points for information that has clots within the more dispersed outliers, you can intuitively see that there are some kind of similarities in the folded points. You can easily group some of the scattered outliers with one or another lump, but there will be many points that lie between the lumps, which can fit into two or more competing clots. The only way to fix this dilemma is to add additional dimension parameters to your data points so that these uncommitted outliers are attached to a single click. (Follow the link to see a beautiful image of the compiled data.)

This brief introduction leads to the following concept: Pattern Recognition . This subject is mathematical and the subject of many studies in the field of theoretical computational science, statistics, artificial intelligence, machine learning and clairvoyance. This last one is half-hearted, but it indicates the essence of your problem: How can a computer predict what you will do in the future? The short answer is that it cannot. A longer answer tries to explain why your tastes and moods change in seemingly random directions into seemingly random times. A good pattern recognition system can select 20 films that you really like, and then recommend another one from the same group as the other 20 that you completely hate. Where did the system crash? Was it implemented in the algorithm, the initial choice of parameters for the dimensions of your spatial space, or was your profile confused because someone else used your Netflix account to order Howard Duck, Cruising, and Beaches?

The wikipedia page for "Pattern Recognition" contains many different algorithms and methods. You can start reading there to better understand individual strengths and weaknesses. You can also try asking this question on the Theoretical Computer Science stack to get long-haired answers.

Answer 2

The BellKor team won the Netflix prize . Thus, perhaps this approach may be the best approach.

To give a detailed, intuitive explanation of how these recommendation systems work, consider the following situation. I watch Star Wars twice a week. Now, if you had to recommend a movie that I liked, what movie would you choose? Movie with Harrison Ford? Sci-fi movie? Perhaps a movie made in the 80s?

The great idea of ​​recommendation systems is that the more they know what you like (for example, which genres, actors, etc.), the better they can give recommendations.

However, if your tastes contradict each other (for example, you like Saving Private Ryan, but also like films about pacifists), it will be difficult for you to recommend a film.

In short, many recommendation algorithms should know:

• What do you like: this is about understanding what function you can use when recording movies that you like. For example. what is the genre of the film, what actors in the film, etc.
• Which movies are similar to what you like. This involves finding a good similarity indicator based on the feature set that you use in the previous step.