Recommendation System Practice (item bright)-Chapter 4th using user tag data

　　Tag application: one is to allow the author or expert to tag the item, and the other is to allow common users to tag the item (UGC). When a user tags an item, this label describes the user's interests, and the meaning of the item, thus connecting the user with the item. Tags are an important feature representation.

4.1 UGC tag system representative applications

The biggest advantage of the tag system: give full play to the group intelligence and obtain keywords that accurately describe the item content information. Accurate content information is an important resource to Improve the Performance of Personalized Recommendation Systems.

• Delicious/citelike/Douban/Hulu

4.2 recommendation issues in the tag system
　　Main problems:1. How to Use the user's tag-based behavior to recommend items for the user (tag-based Recommendation)

2. How to recommend tags suitable for an item when a user tags an item (TAG Recommendation)

(1) Why are users labeled?

Two dimensions:Social DimensionTo help others find information;Function dimensionTo facilitate future search.

(2) How do users tag?

Tag popularity: the number of times a user tags an item.

　　The popularity distribution of tags also shows a very typical long tail distribution.

(3) What kind of tag do users use?

• Indicates what an item is.
• Indicates the type of the item
• Indicates who owns the item
• Express users' opinions
• User-related labels
• User task

4.3 tag-based recommendation system

(U, I, B) indicates that user U tags item I B. A simple tag-based recommendation algorithm:

1. Calculate the tags most commonly used by each user.
2. For each tag, count the items that have been tagged the most.
3. For a user, first find the frequently used tags, and then find the most popular item with these tags to recommend to the user.
   The formula for user U's interest in item I is as follows:

　　Algorithm Improvement:

　　(1) Penalty for popularity

The above formulaTend to give a large weight to the hot item corresponding to the hot tagTherefore, hot items will be recommended to users, thus reducing the novelty of Recommendation results. And,Excessively large weights are assigned to hot tags, so they do not reflect users' personalized interests.. The book draws on the idea of TF-IDF, the formula is improved:

Appropriate punishments for hot tags and hot items will not reduce the offline accuracy of Recommendation results while improving the personalization of Recommendation results.

　　(2) data sparsity

There will be very few tags in new users or new items B (U) limit B (I). To improve the recommendation accuracy, we may need to expand the TAG set, for example, if you have used the "recommendation system" label, we canAdd the similar tags of this tag to the user tag set.For example, tags such as "personalization" and "collaborative filtering.The essence of tag extension is to find tags that are similar to each tag, that is, to calculate the similarity between tags..

　　(3) Tag cleanup

Not all tags reflect your interests. You need to clear these tags. The website allows users to provide feedback on tags, integrating the knowledge of experts and the majority of users.

　　(4) label-based recommendation explanation

Tag Cloud: Improves the diversity of Recommendation results and provides the interpretation function.

4.4 recommend tags to users

(1) Purpose: To facilitate the user to enter tags and improve the tag quality;

(2) method:

• The most popular tag on recommended items;
• Tags that are commonly used by users are recommended;
• Weighting the first two items

 

Recommendation System Practice (item bright)-Chapter 4th using user tag data