The role and problems of the recommendation system

The role of the recommendation system

The recommendation System (Recommender BAE) is the use of information filtering technology to recommend to users the information they may be interested in. Http://www.aliyun.com/zixun/aggregation/8086.html "> Recommendation system is different from information classification and information search processing methods.

Information classification is organized according to the time/subject/category/user/task, such as organizing structured information, browsing classified information can help users understand the organization of information, but the presentation of this information is based on the user's known target information category, the user cognitive things in favor of fuzzy disorder, It is difficult to find information that is unknown and interesting from ordered information.

Information search is based on and group behavior related to the weight ranking information, with the task of the user to send a quick search to the content of interest and then in-depth reading, and then with a new task to continue to search, and the reality is that individual users have to adjust the keyword repeated search in the long tail of information.

With the rapid growth of information, duplication of information and excessive information caused by passive access to information overload, the search engine to actively obtain high-quality information will also cost more, recommendation system is to solve these problems the most potential method, its role is:

helps users quickly discover interesting and high-quality information and enhance the user experience. Increase the user time to use the product. Reduce the negative impact of users browsing to repetitive or offensive information. Provide personalized information, information recommended more accurate. The problem of the recommendation system

The basic principle of the recommendation system is to match from the database to the analysis of user behavior inferred from the preferences, according to the recommendations of the different methods can be divided into the following:

Collaborative filtering System (collaborative filterring) content-based recommendation System (content-based) Hybrid recommendation System (hybrid) based on user-product two Map network structure (network-based)

Most of the mathematical formula, for lack of technical knowledge of the design staff is somewhat obscure, individual attempts from the product design point of view from the data, data peripherals and users from three aspects to analyze, before the analysis needs to understand the following questions:

1. Key meta data. Metadata is data about data that can be used to describe and manage data, such as singers, albums, release Times, publishers, and categories of songs, "Black and white" from Warner's December 2008 edition of the "Orange Moon Moon". For Recommender systems, you need to find important metadata that affects user preferences. Assuming that the user is a fan of Square Datong, the singer is the key metadata, the user may also like the other songs on this album "Little Bug" and "100 expressions", for those who like to listen to new songs, release time may be more important, It is also possible because users like to listen to R&b.

Structured data

Unstructured data

2. Structured and unstructured. Structured organization between metadata (such as the nationality of the singer and singer of the song) can be easily obtained, but these metadata are usually just one of the key metadata, and unstructured metadata (such as rhythm, tone, and timbre) can also affect the user's choice, and the invisible link between the data can only be obtained through a large number of analyses.

3. Relevance. And the user's behavior, background, characteristics and other related, analysis of the data between the regular characteristics. A common purchase book site, 40% of users who bought the book, bought another book. The correlation between beer and diapers can be obtained by analyzing the data correlation of large consumers ' purchase orders.

4. Diversity. The key elements of the structure of the strength of the impact of product diversity, such as the category of books belong to a high degree of complexity leading to the diversity of books, and music relatively single. The diversity of products means that the implicit association between the data is more complex, it can increase the difficulty of analysis, and the recommendation system is more complex.

5. Timeliness. The speed of the data update and the user's demand for new data affect the timeliness of the data, such as popular forum posts than the blog in the timeliness of the article. such as micro-blog and news Such a high timeliness of data requires server data updates to high, time affects the recommended system important data. Data mining focuses on real-time analysis and provides the latest recommendations based on each operation of the user and the introduction of new data.

6. It is difficult to identify. It is difficult to ask users to articulate what they like in a few words, and the user's preferences change over time. Like Google's music recommendation, for most ordinary users, that rhythm and timbre choice to their favorite concert is more difficult. The significance of the recommendation system is to infer the user's preference according to the user's historical record, rather than let the user take the initiative to choose.

7. Label. User tagging is a manual workaround for organizing data, but it can also cause other problems:

non-automated solutions can increase user operations and make it difficult to discover the invisible link between data. The user fills in the label, because the ambiguity of the word will cause the label to be too many, the relation between the data will weaken, reduce the cohesion between the data. Users choose the recommended tags, easy to understand the words will lead to strong data cohesion, resulting in the structure of data bias, not conducive to users to find the content of interest.

8. Scoring mechanism. Usually a five-point system and a two-point system (like/dislike), the more points, the more trouble users choose, the need to eliminate the differences in user evaluation system. User Collaborative filtering content biased to popular, can be filtered to the content of low quality, but the user of the low scores of small content is not necessarily not interested. Through questionnaires, users will choose each problem, and through the network is not mandatory rating, users do not like the content is likely to not grade or jump directly to the next data.

Resources：

The five major problems of

recommendation system. The research progress of personalized recommendation system in Resys. Jianguo, Zhou, Wang Binghong. "Progress of natural science, 19th volume 1th, January 2009 The practice and thinking of watercress in recommendation field. Wang Shou Kun from web2.0 to recommendation Engine 2.0. Learning and Shizhi Source: http://daichuanqing.com/index.php/archives/1757