[Recommendation System thesis notes] Introduction to recommender systems: algorithms and evaluation

This paper is short. As the title says, I will briefly introduce some algorithms and Evaluation Methods of the recommendation system.

The recommendation system was previously a keyword-based filtering system and later developed into a collaborative filtering system, solving two problems: 1. Manually review the documents with a large number of keywords; 2. Filter some non-text files, such as music, based on people's taste.

After that, the research field of the Recommendation System has been split. On the one hand, focus on the commercial value of the actual problem; on the other hand, some machine learners use a large number of technologies in the recommendation system.

This kind of forks promotes the development of the recommendation system. Many researchers have realized that they have ignored two key points:

1. Provide a single recommendation under different types of recommendation systems;

2. Evaluate the recommendation system in a broad sense and encourage researchers to create comparable results from different aspects;

 

 

Some well-known papers in this field:

Herlocker's paper: how to correctly evaluate new recommendation algorithms and systems.

1. Is it worthwhile to spend time researching recommendation algorithms;

2. Are all algorithms equally good;

For the above two problems:

1. The evaluation experiment shows that not all measurements have the same recommendation results, and the correct grouping of measurements may affect the recommendation accuracy;

2. Which prediction scheme is evaluated based on the user's purpose, most directly reflecting the applicability of a recommendation system with a specific purpose.

Middleton's paper: the personal data of user entities is very helpful for content-based technology to be applied in the recommendation system.

There are three reasons for Middleton's paper to become famous:

1. It shows a type of domain that can be promoted to other fields (for example, an entity exists and the recommendation system can effectively guide the user's interest space );

2. It shows how an existing external entity can be processed in the recommendation system.Cold Start Problem(For example, a recommendation system purely based on system filtering cannot provide too much value to their early customers. In fact, before new users fill in their personal data, the recommendation system cannot provide too many valuable recommendations );

3. This work requires a very detailed business evaluation of the effectiveness of individual data in the recommendation system.

Hoffman's thesis: In the latent semantic model, a model-based collaborative filtering algorithm, the latent probability Semantic Analysis and the maximum Expectation Algorithm are used to construct a concise and precise dimension reduction model. This model subconsciously assumes that the weights of user preferences as a vector are distributed on some potential factors. In addition, their experiments show that their algorithms are very accurate and time-consuming.

Huang's thesis uses a different method to solve the sparseness problem in the recommendation system-associate retrieval ).

Using data from China's online library, they explored how a diffusion activation algorithm (spreading-activation) could improve the quality of recommendation systems to help users mine and transmit associations. If both users read or love similar books but are not the same, their associations will be lost. Huang's paper shows that a diffusion activation algorithm can be used to help the recommendation system, especially to give appropriate recommendations to new users.

Deshpande & karypi: a record-based recommendation system is used to solve the Top N Problem in the recommendation list, rather than for all.

Their papers show the similarity between item entries or item sets during group buying promotions to provide effective recommendations. In addition, in order to evaluate these two key technologies, the paper uses a variety of datasets to verify the results.

[Recommendation System thesis notes] Introduction to recommender systems: algorithms and evaluation