Research on the needs of website users: how to make personalized recommendations to users?

But through the recent period of thinking, there are some new discoveries and summaries.

Personalized recommendation is a highly technical topic, but this paper only focuses on the "truth" in the recommendation process and does not involve any profound technology.

The discussion in this paper mainly refers to the basic characteristics of Zite and several Weibo filter products at present.

In order to avoid unnecessary disputes, except the official function of Sina Weibo, other specific product names are not specified.

1. Network reading recommendation based on SNS With the development of technology in recent years, a number of industrialized technologies have been formed in natural language processing, data mining, machine learning and other fields, laying a technical foundation for highly personalized content recommendation.

By processing, we need raw materials to make products.

So is personalized recommendation.

Personalized recommendation is nothing more than selecting the content that the user may like and presenting it to him/her.

Liking is influenced by two main factors: (1) the degree of relevance between content and users' interests; (2) the quality of the content itself (popularity)

In other words, who is the user? What kind of things does he like? Is this the type that users like? It is very lucky to be able to identify users' interests and find the characteristics of the content in general terms. Social networking provides almost all the above raw materials for personalized content recommendation: users have accounts to identify themselves; Users' interests can be obtained by mining historical behavior data such as users' postings and network relationships. The type characteristics of content can be obtained by natural language processing technology.

This is perfect! Zite can be said to be such a timely product, which perfectly combines these.

By logging into Twitter or GoogleReader account, Zite will analyze users' reading preferences (interests), and then select news/articles with high matching with users' interests according to users' preferences, which really greatly improves the efficiency and quality of reading.

Second, Weibo's information filtering and recommendation is along the above ideas. It seems to be a good product form to deal with Weibo, an information treasure house that tends to explode, filter according to users' preferences and push the content that users care about.

However, if you think about it carefully, you will find that there are two problems to be solved: 1, popularity. We must pay attention to a phenomenon: Weibo is a lightweight and highly participatory public information source.

Judging from the number of entries, the daily UGC in Weibo may be higher than the whole Internet in the previous month, but the content is uneven! So the quality of content becomes a problem.

For example, a user who likes jokes will be very happy if he is recommended a high-quality joke. But it's also a joke. Recommending a joke of average quality may be a kind of junk information.

It can be seen that Weibo information recommendation needs to pay more attention to one factor: the quality (popularity) of the content.

The accuracy and efficiency of natural language processing technology based on machine learning and other methods in text classification have reached the industry standard.

However, the recognition and processing of semantics is still immature.

At present, there is no natural language processing technology to evaluate the viewpoint and literary talent of an article.

A reliable quality evaluation method depends on the feedback from the masses, that is, popularity.

It is not difficult to evaluate the popularity of articles (especially for Weibo). It can be judged by how many people in Weibo forwarded this comment.

However, Weibo, as a real-time and instant SNS product, on the one hand, emphasizes the quality of content, on the other hand, there are also issues of participation and stickiness that need to be considered.

The real-time performance of recommendation systems that rely on expert user feedback will inevitably be greatly reduced.

This is why the timing of filtering recommended products by Weibo is quite confusing (including the interest reading function provided by Sina Weibo).

This naturally sacrifices the feelings of many users.

Zite, a product form, does not require high real-time performance, so it is not necessary to consider this issue.

2. There is another difference between users' expectation of high Weibo filtering and Zite reading recommendation.

As a reading recommendation, as long as some popular articles are selected and pushed to users, users' expectations can be well met.

If you miss some important hot news, or mix a small amount of content that users don't care about, as long as there is a certain relevance, it will not significantly reduce the user's reading experience.

But Weibo is different.

The attention of Weibo users has been the result of users' preliminary screening, and he certainly doesn't want to miss any important information of these users; On the other hand, as a product in the form of a filter, users' expectations of its irrelevant content filtering ability will be relatively harsh.

Based on the above two points, the user satisfaction and passing line of Weibo filter are relatively high! 3. Let's talk about popularity. In addition to popularity, there are also some factors that are difficult to concretize that affect users' preferences.

For example, many users may have this experience. A Weibo with low forwarding volume may be something I like and recognize very much.

An important feature of SNS is that users' preferences depend largely on their friends.

The user's recognition of content often depends on his recognition of information sources.

You have to admit that even the users you care about have completely different weights in your mind.

The question is, which concerns are users' favorites? Many times, only the feet know whether the shoes are good or not.

Users generally don't tell you! 4. Is Facebook's EdgeRank algorithm a silver bullet? As the originator of social networks, Facebook also faces the problem of information overload.

It currently uses a recommendation algorithm called EdgeRank.

In terms of principle and workflow, it's really simple: any action you make to your friends is called an edge (including comments, forwarding, likes, etc. According to different actions, a score will be calculated for every action you do. The cumulative score of all actions expresses the relevance between you and the friend, and this total score will affect whether the friend's Weibo is easy to appear in your timeline.

This algorithm is effective for the relationship-oriented SNS.

But for social media like Weibo, is it still effective? 5. Users have different subtle preferences. Highly personalized recommendation, no matter how the product form changes, the ultimate goal is to deeply grasp the user's interest.

In order to thoroughly grasp the characteristics of users, in addition to strengthening technical means to identify, it is also a very important factor to fully collect user data.

As the saying goes, a clever woman can't cook without rice.

Many important user characteristics, such as users' attention to certain details; Or the user pays special attention to some specific users, which is not enough to get a clear judgment from his historical data.

This also increases the difficulty and complexity of Weibo filtering.

Let's see how Zite solves this problem.

Zite is oriented to long web pages, and users first see the article classification and title, not the text.

Users click on the title to start reading before entering the article.

This natural process actually hides a shocking secret-"I am interested in the content of this article"! Yes, no matter how clever the user interest identification algorithm is, it also needs to keep running-in feedback with users and listen to their voices! But Weibo doesn't filter well. Weibo is short in space. Forcibly adding a click-through product process is undoubtedly suicide in user experience.

So will giving the user a "like" button solve this problem? My answer is pessimistic, and the user's participation motivation may be difficult to guarantee.

This needs to combine the first point (popularity) and the second point (user expectations). For a product whose initial state can't meet users' needs, it is difficult to ensure users' participation motivation.

The patience of users is the most expensive resource in every new product promotion.

Conclusion: To improve the satisfaction of users' needs, we must have high-quality user interest identification ability. And meet some needs of users at the first time, at least reaching the passing line; Finally, the product also needs to have a strong ability to collect user characteristics (preferences), so that each of your trial users will eventually become a loyal fan of your product, thus infecting people around you and helping the product spread rapidly.

Do a good job in social recommendation, three-point technology, seven-point product and operation.