Recommended Search Methodology --- direction of the 2022 search gameplay

The concept of vector has been talked about for almost a year, and the position of vectors in the process of playing search in the future is becoming more and more important, just like the seven-day spiral of that year as the core basis; I said that vectors are not words I made, vector recall algorithms are also real, do not be ignorant to limit your thinking. Ask more about the next degree or read some books that I recommend to everyone, and slowly you will know what is going on, you can't keep up with the changes of the times, which does not mean that the times will not change.

Talking about the development of search, the first generation of search is the development of categories, as long as your products are placed on the right categories and the leaf categories have traffic.

The second generation began to develop mainly on keywords, mainly in the era of machine statistics, behind each keyword is linked to the pit production, whose pit production is large, the ranking is high.

The third generation of search to statistical machine learning methods, based on user query, recall, L2R these three processes, to a certain extent to improve the efficiency of user acquisition, but this service model is still a series of information thrown to the user, users ultimately still need data to filter and identify to get their most needed information, so the third generation of search services are defective in terms of efficiency.

With the development of Web technology, human beings have experienced the Web 1.0 era with web links as the main feature to the Web 2.0 era with data links as the main features, and the current Web technology is moving towards the semantic network based on knowledge interconnection proposed by Berners Lee, the father of the Web, in 2001, that is, to the Web3.0 era.

body

What is a semantic network:

The goal of semantic networking is to build a World Wide Web that can be understood by both humans and machines, making the network more intelligent, and improving more accurate and fast services on the basis of parsing user queries.

To do this, all the data composed of online documents needs to be processed and stored together to form a large, usable database.

To do this requires strong data capabilities and web content intelligent analysis capabilities, first of all, it is necessary to semantically label these Web data, but due to the huge volume and heterogeneity of Web data, the scope of the field is large, so how to automatically add appropriate tags to the web page content has become one of the technical pain points, plus the Web data that has been labeled, how the machine thinks and reasons is also an urgent problem to solve.

Due to the existence of the above technical problems, all of them were not greatly applied 10 years ago when the semantic network was proposed, but with the strength of the algorithm and the great improvement of the ability of data processing, coupled with the accumulation of 10 years of research process, the mature ontology model modeling and formal knowledge expression methods and the ontology language of the World Wide Web laid the foundation for the emergence of subsequent knowledge graphs.

Overview of the Knowledge Graph:

A knowledge graph is a structured semantic knowledge base used to express concepts in the physical world and their relationships in symbolic form.

Its basic constituent units are the "entity-relation-entity" triplet, as well as the entity and its related genus value pairs, and the entities are interlinked through relationships to form a network of knowledge structures.

As can be seen from the definition, the knowledge graph is a semantic knowledge base with sufficient domain knowledge, the most important component of which is the triple.

A triplet is a semantic relationship that represents the semantic relationships between entities, all with direction and explicit semantics.

The above concept is quoted in the book "Alibaba B2B E-commerce Algorithm Practice", and the author is Alibaba's official CBU Technology Department.

Why has always emphasized that the search logic has completely changed, the change is that it is no longer a single keyword weight ranking, no longer a single pit property weight.

The core of the search ranges from text correlation matching, behavioral data matching, to semantic vector recalls and scene recalls.

Behind it is the gradual transformation of consumer demand for goods into the demand for scenes.

Redefine the demand generation of goods through scenarios, which requires in-depth mining of user and behavior data and product knowledge.

This can be difficult to understand, but let me give you an example:

In the past, the database built by the search system was based on keywords and commodities, and now the generation of the semantic network is to build a database of semantic knowledge bases and commodities through knowledge graphs.

That is to say, there are two databases behind the current search, one is the machine learning algorithm of keyword statistics, and the other is the semantic knowledge database, which is recommended according to the semantic recall.

Semantic knowledge base, there must be many students who are confused.

Angry classmates may say that I made up "words" in a nonsense.

You may wish to buy the book "Alibaba B2B E-commerce Algorithm Practice" that I mentioned above.

Not understanding is another matter, but it is enough to prove that the official is doing that, you understand or can't understand that it is your own: "knowledge side" problem, so read more books, study more, don't shoot arrows every day and then target yourself to make up a set of theories.

That is to say, when we search for keywords, the core is the accuracy of identifying intentions, which used to be based on user identity and behavioral data preferences, but the recalled goods still need to be screened and screened by the users themselves through data and do not fully achieve "accurate identification of shopping intentions", but after building the semantic knowledge base database, it is different, according to the semantic recall is a vector recall.

The vector is the continuation of the label, and the existence of the label is to restore the real shopping intention of the consumer.

Vectors solve the problem of semantic similarity.

Therefore, semantic similarity is actually a problem of vectors.

Vector recall is an algorithm, behind the algorithm is semantic similarity, the problem solved is to identify the user's precise needs.

The keyword pit production that we are most familiar with can only solve the problem of word recall

That is, it is in line with the word weight and relevance problem, but it cannot solve the problem of semantics, so the recalled population is not accurate, at this time, if there are too many keywords, the system will use the brush hand as a user for identification, and judge the preference for recall recommendation according to the behavior trajectory data.

The traffic you make up is similar to the brush hand, do you know why the more you brush, the more inaccurate it is? In particular, it is now a real-time shopping chain recommendation, which will be based on real-time behavioral data preferences.

The original crowd is not accurate, and the recommended crowd will be even more inaccurate and enter the dead cycle.

Many students said that I saw that there are a lot of people brush keyword pit production also brushed up ah, this I do not deny, can do the basic is the explosive data model is very thorough understanding, plus the general early stage will drive, with the drive to take the initiative to mark the single product into the pool, with the brush non-search way pit production forced search passive to the single product marked into the pool, if you use this two ways into the pool more accurate, plus the real visitor clicks behind is greater than the number of clicks of false visitors, the preference of behavioral data will change.

In fact, it is a matter of proportion, if the real visitor is difficult to be greater than the false, and then the recommended search traffic is not opened to brush the number of buyers who pay incrementally, the form of data is good-looking, and the basic fourth week will slowly fall out.

Behind this is the proportion of positive feedback of real clicks.

To solve this core, it is necessary to lay out the "semantic vector words" to brush the structure of the keyword and open the recommended search visitors.

In the end, it really doesn't matter what keywords to brush, brushing the semantics and structure is fundamental.

Take the through train, the through train is the most text-oriented,

But now you put half title and full title into it or synonymous, and the typo keywords will also be displayed.

This fully shows that the current through train to show, not only pay attention to relevance matching, but also pay attention to semantic recommendation.

As long as there is text, the system recalls the recommended similar semantics from the semantic knowledge base database.

This is something that could not have happened in the past, and it used to be that as long as this keyword was not searched by anyone, it would not be displayed.

Some students say that this account weight is related, this point does exist, but the most fundamental is that the semantic knowledge base of the semantic network knowledge map has now been constructed, and it will be recalled and recommended according to semantics.

It is said that the fundamental reason is that the volume of historical behavior data that has been upgraded and precipitated by the algorithms in these decades is large enough to allow the system upgrade to develop into a semantic recall.

Now it has really reached the needs from the demand for goods to the scene.

From statistical machine learning algorithms to truly intelligent scene semantic recommendations and even cross-category product collocation and recommendation.

So don't just stay in the previous search cognitive system.

Why has it always been emphasized that the underlying logic of search has changed completely, but everyone just can't feel it, just feels that it is difficult to do.

Behind the difficulty is the lack of understanding of the underlying logic.

I wonder myself why people don't feel the change.

In fact, from the operation end, no matter how your underlying logic changes, my operation means are still those few tricks.

Brushing, driving, optimizing keywords, how to write the title is basically these points, so how will you change the underlying logic or need to do it through these means, it is indeed not false.

The means of operation have not changed, what the underlying logic will be changed to pale, it is not as good as you tell me how to brush, how to drive, how to optimize the title and optimize the keywords to be real.

I summarized some of the changes I made from the operational side of the search underlying logic changes.

First: the product layout and the rhythm of playing goods are not the same, before the single link to push the main model, now is the multi-link layered traffic operation.

Second: the way the title is written has indeed changed a lot, in the past, the attributes were stacked as much as possible, so that the word segmentation was as rich as possible to get as much keyword traffic as possible, and now it is basically based on the word system layout, playing a vector or locking the display range to focus on a vector to deal, so that the system recommends. So short titles, and other structured forms of headings began to appear.

Third: keyword optimization, the previous optimization around the keyword search popularity, combined with the number of online goods to play weak keywords, now optimized is the combination of keywords structure and semantic accuracy behind the keywords.

Fourth: the biggest feeling is the node problem, before playing the explosive model to pay attention to what weight update point, 3, 5, 7, now more nodes, more detailed most of the students died in the processing of nodes without guidance.

Fifth: through the car, the previous through car for the search service to do the direction has been the word system, the more keywords the better, pull up the search; now the fewer the keywords, the better, pay more attention to the single product active "marking", and guide the "vector" problem.

The above is that it is difficult to rise without paid search, and now this situation is broken, so you may only find that many links will come to a lot of search traffic without driving.

But my point is that it is inevitable that the payment for good products is less and less, and good products should reduce promotion fees. Sixth: the high-traffic value closed-loop system will be more deeply rooted in the hearts of the people, and solving the vector problem will solve the problem of directionality, accuracy, and semantic accuracy of traffic, and the value will be higher with high traffic value.

I write this article may have quite a few students can not understand, in 2018 we began to talk about labels why not the same situation?

But what about the facts?

You can doubt, but don't give up the courage to validate and learn.

Anger often comes from ignorance, or from invisibly touching the interests of others.

Those who agree with my point of view may wish to participate in our offline class on September 15th to thoroughly understand what the vector recall and semantic vector word layout are.

Also suspect that there is a semantic recall recommendation, see if your home has "small degree" and "small love", how they are recognized and recommended.

If you like my article retweets is the biggest support.

This article is provided by the chief operating officer of the Seven Treasurers - Ghost Brother (WeChat: qdbz888) only represents personal views! For more details, please see the circle of friends.

Sign up,

Classes start on September 15, 2021

Location: Guangzhou

Sofitel Guangzhou Shengfeng (988 Guangzhou Avenue, Tianhe District, Guangzhou, China)

Arrange your trip in advance.

Cost: 29800/person (self-consumption)

Multi-person discount (self-consumption)

In the past, offline members only needed to pay 10,000 yuan for retraining, send core information, and organize the company's internal training to enjoy special treatment

Payment terms: Pay 9800 deposit to learn the latest online course vector recall explosion search game 1.0, earn back tuition in advance to strengthen basic theoretical learning. Students who participate in the vector recall pop search game do not come to participate in the offline deposit and do not refund as online members to learn the tuition fee of the vector recall burst search game.

Note: For example, all online members who want to participate in offline courses only need to make up the difference of 20,000. Offline members only need to pay 10,000 for retraining.

Scan the ghost brother QR code registration

Listen to free membership classes

Label Overlay Gameplay 6.0 First Revealed You can click to buy to watch