Introduction to the main recommended algorithms

Reprint: http://blog.sina.com.cn/s/blog_602feaa80100fjq9.html

In the recommendation system Introduction, we give the general framework of the recommendation system. Obviously, the recommendation method is the most important part of the recommendation system, which determines the performance of the recommended system to a large extent. At present, the main recommended methods include: Based on content recommendation, collaborative filtering recommendation, based on association rule recommendation, based on utility recommendation, based on knowledge recommendation and combination recommendation.

First, based on content recommendation

Content-based recommendations (content-based recommendation) is the continuation and development of information filtering technology, which is based on the content information of the project to make recommendations, and does not need to be based on user evaluation of the project opinion, It is more necessary to use machine learning methods to get the user's interest information from the case of the characteristic description of the content. In the content-based recommender system, the project or object is defined by the attributes of the relevant characteristics, and the system is based on the characteristics of the user's evaluation object, learns the user's interest, and examines the matching degree between the user data and the project to be predicted. The user's data model depends on the learning method used, the decision tree, the neural network and the vector-based representation method. Content-based user data is a historical data that requires a user, and the user profile model may change as the user's preferences change.

The advantages of the content-based recommendation approach are:
? 1) No other user data is required, no cold start and sparse issues.
? 2) can be recommended for users with special interests and hobbies.
? 3) can recommend new or not very popular projects, no new project issues.
? 4) by listing the content characteristics of the recommended items, you can explain why those items are recommended.
? 5) There are relatively good technologies, such as the classification of learning technology has been quite mature.

The disadvantage is that the content can be easily extracted into a meaningful feature, requiring a good structure of the feature content, and the user's tastes must be able to be expressed in the form of content features, and can not be explicitly judged by other users.

Second, collaborative filtering recommendations

Collaborative filtering recommendation (collaborative Filtering recommendation) technology is one of the earliest and most successful technologies used in recommender systems. It generally uses the nearest neighbor technology, uses the user's historical preferences information to calculate the distance between users, and then uses the target user's nearest neighbor user to the commodity appraisal value to predict the target user to the specific product preference degree, the system thus according to this preference degree to the target user to recommend. The biggest advantage of collaborative filtering is that there is no special requirement for the recommended object, which can handle unstructured complex objects, such as music and movies.

Collaborative filtering is based on the assumption that a good way to find content that is genuinely interesting to a user is to first find other users who have similar interests with the user, and then recommend the content that they are interested in to the user. The basic idea is very easy to understand, in daily life, we tend to use good friend's recommendation to make some choices. Collaborative filtering is the use of this idea in the e-commerce recommendation system, based on other users of a content evaluation to the target users to recommend.

The recommendation system based on collaborative filtering can be said to be from the user's point of view to carry out the corresponding recommendation, and is automatic, that is, the user obtains the recommendation is the system from the purchase mode or the browsing behavior implicit obtained, does not need the user diligently to find the suitable own interest the recommendation information, such as fills in some survey form and so on.

Compared with content-based filtering methods, collaborative filtering has the following advantages:
1) Ability to filter information that is difficult to automate machine content analysis, such as artwork, music, etc.
2) share the experience of others, avoid incomplete and imprecise content analysis, and be able to filter based on complex, difficult-to-articulate concepts such as information quality and personal taste.
3) have the ability to recommend new information. You can find information that is completely different from the content, and the user's content of the referral information is not anticipated beforehand. This is also a big difference between collaborative filtering and content-based filtering, and many of the content-based filtering recommendations are user-familiar content, and collaborative filtering can uncover potential interest preferences that users have yet to discover.
4) can effectively use other similar user feedback information, less user feedback volume, speed up the pace of personalized learning.

Although collaborative filtering as a typical recommendation technology has its considerable application, but the collaborative filtering still has a lot of problems to be solved. The most typical problems are sparse issues (sparsity) and extensible issues (Scalability).

III. recommendation based on association rules

The recommendation based on Association Rules (Association rule-based recommendation) is based on the association rules, the purchased goods as the rule header, the rule body as the recommended object. Association rules mining can find the correlation of different goods in the sales process, and has been successfully applied in the retail industry. The rule of management is to count the percentage of transactions in a transaction database that have purchased a trade set X and purchase a commodity set Y, and the intuitive meaning is how much the user is inclined to buy something else when they buy some. For example, when buying milk, many people buy bread at the same time.

The first-Step association rule discovery of the algorithm is the most critical and time-consuming, which is the bottleneck of the algorithm, but can be done offline. Secondly, the synonym of commodity name is also a difficult point of association rules.

Iv. recommendation based on utility

The utility-based recommendation (utility-based recommendation) is built on the utility of the user's use of the project, and the core question is how to create a utility function for each user, so The user data model is largely determined by the utility function used by the system. The benefit of a utility recommendation is that it takes into account utility calculations for non-product attributes, such as the reliability of the provider (Vendor reliability) and the availability of products (product availability).

V. Based on Knowledge recommendation

A knowledge-based recommendation (knowledge-based recommendation) can be seen in some way as a reasoning (inference) technique that is not recommended based on user needs and preferences. Knowledge-based methods differ significantly from the functional knowledge they use. Utility knowledge (functional knowledge) is a kind of knowledge about how a project satisfies a particular user, so it can explain the relationship between need and recommendation, so the user data can be any knowledge structure that can support inference, it can be the query that the user has normalized, It can also be a more detailed representation of the user's needs.

Six, the combination of recommendations

Because of the pros and cons of various recommended methods, in practice, combinatorial recommendations (Hybrid recommendation) are often used. The most researched and applied combination of content recommendations and collaborative filtering recommendations. The simplest approach is to use a content-based approach and a collaborative filtering recommendation method to produce a recommended prediction result, and then combine the results with a method. Although there are many recommended combinations in theory, they are not always effective in a specific problem, and the most important principle of combinatorial recommendation is that the weaknesses of the recommended technologies can be avoided or remedied by the combination.

In combination, researchers have proposed seven combinations of ideas:
1) Weighted (Weight): Weighted multiple recommended technical results.
2) Transform (Switch): Depending on the issue background and actual situation or requirements to decide the transformation using different recommended techniques.
3) Mixing (Mixed): At the same time, using a variety of recommended technology to give a variety of recommendations to provide users with reference.
4) feature combination (Feature combination): Combining features from different recommended data sources is used by another recommendation algorithm.
5) Cascade (Cascade): first with a recommended technology to produce a rough recommendation, the second recommended technology based on this recommendation to further make more accurate recommendations.
6) Feature expansion (Feature augmentation): A technique that generates additional feature information embedded in the feature input of another recommended technique.
? 7) meta-level (META-LEVEL): The model produced using one of the recommended methods as input to another recommended method.
Seven, the comparison of the main recommended methods

Various recommended methods have their own advantages and disadvantages, see table 1.

Table 1 Comparison of main recommended methods
Recommended method	Advantages	Disadvantages
Content-based recommendations	The recommended results are intuitive and easy to interpret; No domain knowledge required	New user issues; Complex attributes are not handled well; To have enough data to construct the classifier
Collaborative filtering recommendations	Novelty interest discovery, do not need domain knowledge; Improved performance over time; Recommended personalization, high degree of automation; Ability to handle complex, unstructured objects	sparse problem; Scalability issues; New user issues; The quality depends on the historical data set; The recommended quality is poor at the beginning of the system;
Rule based recommendation	To discover new points of interest; No domain knowledge	Rule extraction is difficult and time-consuming; The synonym of product name; Low level of personalization;
Recommendation based on utility	No cold start and sparse problems; Sensitive to changes in user preferences; Ability to consider non-product features	User must enter utility function; The recommendation is static, the flexibility is poor; attribute overlap problem;
Based on Knowledge recommendation	Can map the user's needs to the product; Ability to consider non-product attributes	Knowledge is difficult to obtain; The recommendation is static

Category: Data mining and machine learning

Introduction to the main recommended algorithms