News recommendation: Google News, Sohu News, today's headline research and analysis

This article from Google News, Sohu News, today's headline recommendation system, analysis of news and information industry, recommend the system to adopt the main strategy.

1, Google News

Rec (article) =if (article) XCF (article)

IF (article) content filtering

GoogleNews classifies news articles into predefined topic categories, including international, sports, and entertainment.

In the log analysis, according to the user's search and clicks behavior, constructs the Bayesian framework, predicts the user news event interest.

CF (article) Collaborative filtering

The user clustering, so that users have the same interest in other users to help complete the personalized recommendation process.

2, Sohu News

Features: A wealth of content classification tree system, can be repeatedly optimized, operation intervention color heavy.

Content Category: Label words, themes, channels, channels, topics, hotspots, regions, sources, etc.

User: User grouping, creating user portrait, corresponding to content classification.

News Warehousing: Content synchronization, extraction (based on the content characteristics of the text, based on the title, summary of the key words).

Recommendation: Calculate the distance between news keyword and label word, subject, channel, channel.

Correction: According to conversion rate and other indicators, repeated revision of user portrait and content classification.

3. Today's headline

Featured: A rich recommendation logic

Recommendations based on similarity of similar article topics: by obtaining similar articles that users have read articles about.

News based on the same city: for users with the same geographic information, it is recommended that the most popular articles in the city match.

Based on the recommendation of the article keywords: for each article, extract keywords, as a description of the content of the article a feature. Then with the user action history of the text

Chapter keywords for matching recommendations.

Based on the popularity of the popular article recommended: According to the user in the station reading habits, find the most popular articles, for all users who have not read the article

Recommended.

Based on the relationship of social friends reading habits recommended: According to the user's outbound friends, get outside the station friends to forward comments or published articles to recommend.

Based on the user's long-term interest keyword recommendations: Through the comparison of short-term and long-term reading interest topics and keywords to recommend.

Based on similar user reading habits of the list recommendation: The calculation of a certain period of user action similarity, the reading content of the intersection of recommendations.

Based on site distribution source content recommendation: Through the user read the article source distribution for users to calculate 20 users like to recommend the news source.