Recommendation System-Tag-based recommendation

Article Directory

Recommendation System-Tag-based recommendation

self introduction

Yang Xiaodong is curious about the world.

You can use Tensorflow to build, modify and train some basic models.

Work with Ye Yanwei and Liu Hanqian to complete a large-scale project.

In the future, I hope to have a bookstore of my own.

Tag-based recommendations

Recommendations based on tags are very common in our lives. Music sites, Douban, video sites, these sites related to entertainment, the shadow of tags can be seen everywhere. Movie tags, book tags of Douban, music tags of NetEase Cloud Music, video tags of station B. It is a content-based recommendation algorithm.

The recommendation system is only necessary when it encounters information overload. Obviously, when the bookstore is a small bookstore, there are not many books, and there is no need for a recommendation system, just simple classification.

Here we mainly talk about a simple tag-based recommendation algorithm.

SimpleTagBased

$i)=\sum_{b} n_{u, b} n_{b, i}$
$n_(u, b) $Is a user$ u $Fight over standard sign$ b $The secondary number,$ of N_ {B, I} $Is a composition product$ i $He is playing too marked to check$ the number of b $. Use tags to connect users to items. Explainable, it will make readers feel that tags are reasonable, and it makes sense to recommend books based on tags.

Insert picture description here

The characteristic form of the label establishes the connection between the user and the product, and the label is the link between the user and the product.

TagBasedTFIDF

$i)=\sum_{b} \frac{n_{u, b}}{\log \left(1+n_{b}^{(u)}\right)} n_{b, j}$
SimpleTagBased can not reflect the user's personalized interest, here we can learn from the idea of TF-IDF, and improve this formula. $n_{b}^{(u)}$ Recorded tag $b$ How many users have used it

TagBasedTFIDF ++

$i)=\sum_{b} \frac{n_{u, b}}{\log \left(1+n_{b}^{(u)}\right)} \frac{n_{b, i}}{\log \left(1+n_{i}^{(u)}\right)}$
In the same way, we can add punishment for popular items. $n_{b}^{(u)}$ Recorded tag How many users have used $b$ , $n_{i}^{(u)}$ Recorded item How many users have $i$ marked.

Examples of cultural and entertainment recommendations-videos

$i)=\sum_{b} {Q}\frac{n_{u, b}}{\log \left(1+n_{b}^{(u)}\right)} \frac{n_{b, i}}{\log \left(1+n_{i}^{(u)}\right)}{R}$
Let us consider the quality factorvideo itself $R$ (it can be the number of views with time decay, or the CTR after going online), and then consider the quality of the label itself $Q$

Examples of cultural and entertainment recommendations-Douban

Guided users, Douban used tags to cluster the recommendation results of users, showing the recommendation results of users under different tags, increasing the diversity and interpretability of recommendations.

Finally, the page is organized to improve the diversity of the recommendation results, because readers’ interests may be broad for a long time, but one day is more specific. We want to hit the user’s interest of the day on a certain day, and show through the tag cloud All readers’ interests allow readers to select relevant tags based on their interests today, and obtain reference results, which improves the diversity of recommended results and makes it easier to satisfy the diverse interests of users.

Summary and personal views

Tag-based recommendation is a content-based recommendation algorithm.

The advantages of content-based recommendation are as follows:

Independence between users.
Good interpretability
New items can be recommended immediately

The disadvantages of content-based recommendation are as follows:

Feature extraction of item is generally difficult. A high-quality community will contribute high-quality tags to achieve a virtuous circle. The tags of some movies and books are human subjective, high-level extractions.
Unable to dig out the potential interest of users. The final organization page solves this problem to a certain extent (not in essence).
Could not generate recommendations for new users. Let users choose some tags when registering.

Based on the above analysis, I think that for entertainment recommendations, label-based is a good choice. Entertainment works are different from e-commerce products. Movies, books, and videos on entertainment websites have content. They have their own characteristics that are different from other works. No author wants his books to be exactly the same as others, and these works also have Some commonalities are temperature, and the user's tags, such as bookmarks, footnotes, summarize his content. These works will easily resonate with people, and high-quality communities will also have high-quality evaluations and tags. Customers and readers are also willing to share their feelings and evaluations, and these tags are their understanding, and their own content is a very good feature, which is a high-level extraction of human subjectivity.

So I think that for entertainment recommendation, tag-based recommendation is a good choice

Willing to share their feelings and evaluations, and these tags are their understanding, and their own content is a very good feature, which is a high-level extraction of human subjectivity.