Our Ctrip AI: from recommendation to system cloudiness, from algorithm to formula, how does it work?

Original: Li Sheng

Ctrip group

From a big data platform to a shared personalized platform

From an intelligent customer service robot to an intelligent shopping guide to a small poetry machine

From a company-level data warehouse to a marketing delivery algorithm

How does ad algorithm work, etc.? I'll tell you.

First

Let me tell you

Which technical services are difficult to migrate to cloud

Which technical services are difficult to migrate to cloud

What does this mean?

This is a service closely related to business

Difficult to achieve intersectoral cloudiness

For example

Something like PaaS or Docker

In fact, it is completely unrelated to business

Relative

Cloud ecology and cloud architecture are easier to create

But it looks like personal recommendations

In part, business AI technologies and business are closely related

What is complexity of cloud processing?

Technically extract common features of a business

Clarify responsibilities between platform and different users

Technical boundaries

Simultaneously improve reuse and integration of modules, data and business knowledge

So for user

Creating Big Data and a Big Industry Group Knowledge Base

Key business benefits

Facilitate convergence of similar industries to form a new closed ecosystem of big business

In targeted industry clusters

On one side

Business is highly dependent on cluster

Full Big Data and Big AI Knowledge Map Top Level Service

Move directly

Optimization and modernization of relevant enterprises

Even creating a new big business ecosystem

On other side

These services must be compatible

Applicable to dozens of completely different business models

Segmented industries with different structure

This is why cloud is essential

High business

Cloud for related service products

Although it has always been a deep field of cloud computing

In fact, cloud architecture develops architecture, algorithm and data

Design possibilities create near-limitless challenges

But our Ctrip

A lot of innovative work has been done in this area

For example, a group personalized platform

Once encountered over 100 landing scenarios

Annual income over 200 million yuan

But only 7.2 people were used

This includes services, algorithms, data, and products

and supports

Almost 300 million calls a day

One scene iteration can be completed in 1.3 man-days

Or a new online scene

The same applies to robotic customer service platform Ctrip

Closed to business

Simultaneous support for over 60 business units

Linked online business in over 20 industries

Trust these companies

Deep architecture optimization for highly coupled systems

Innovation of many new technologies

The implementation and reconstruction of architecture implemented modularity of service platform

Reusable, customizable and compatible with various services

Let me tell you

There is a big difference between OTA and traditional e-commerce

For example, Amazon

Jingdong has mineral water, clothing and electrical appliances

Products such as mobile phones

There are tens of thousands of categories

But their customer service process, whole transaction process

Mostly

Everything is decided in a standardized way

AND OTA

Online Travel Agent, Online Travel Agent

Not same

For example

Booking a hotel is completely different from booking a flight

User requirements vary

The process of finding and booking flights, hotels and train tickets

The inventory system and structure is almost completely different

In general

Systems of different BUs are relatively independently developed by different teams

But traditional e-commerce

Usually tens of thousands of categories are used in one system

Inside Ctrip

Our company

Some Public Part Maintenance Products

Especially for artificial intelligence

There are also some personal recommendations

Although

For employees of a company or group

But it should actually be on a technical level

Encourage compatibility and self-service across 20-30 industries

But these tricks

This is also closely related to business

Now let me talk about this in discussion form

How to do cloud work on a system with high business connectivity

O

Topic topics will be divided into four parts

Ctrip Artificial Intelligence Practice

AI cloud

Group personalized cloud

The Future of Travel Technology

Wait...

One,

Ctrip Artificial Intelligence Practice

First, a small project for a poetry machine

Let me start with a "simple" AI system

We have a Ctrip poetry writing robot. For poetry and distance

This element

Mainly to solve problem of poetry

You are uploading landscape photos from your travels

This will help you identify content of photo

Write another Tang poem for AI ​​user

Example: below

Below is a photo of this person

Even though I can't write symbols

But Ctrip

Many users will also use it to create a text version of Meitu Xiuxiu

Take some photos. Help make "beauty" with words

Then

Take a picture of West Lake and he can recognize it

Because of photo

Not much information on its own

It's just sun and water. AI will parse it

Help you write a poem about West Lake

In following way:

This photo was taken on Mount Fuji

Because it's getting late

That's why I don't see any information

The system skips background

Some season and time data is displayed in Scenic Spots Knowledge Base

Of course, including weather conditions of day

It was snowing day this photo was taken

But I can't see it in picture

Because picture itself doesn't contain that much information

Write poems based on knowledge base

You see following:

Or this input

"Go Degang" of three words that AI helps write poetry

This is entirely based on character knowledge map for writing poetry

You see following:

We use

Blind test in Shanghai

It was a blind test with professional poets in Shanghai

The judges didn't know beforehand

Whether each poem was written by a human or artificial intelligence

Blind human-machine test

No. 3, slightly better than poets in Shanghai

You see this:

Again

Now you can also write quatrains, rhythm poems, pagoda poems, Tibetan acrostics

Also searchable

User takes photo

Artificial intelligence extracts ancient Tang poetry that best matches photography

Also year and author

You will see following:

Two,

Ctrip Artificial Intelligence Practice

Helpdesk robot

When another AI application

Larger piece

Mostly a helpdesk robot

Currently at session level

Approximately 80% of conversations can be resolved by bots

In remaining 20%

Most customer support teams should be helping hotel or airline guests

Manual

Fight for some rights or deal with urgent matters

For example

Due to a problem, illness, or temporary change in user's schedule

Our trip

We need to help guests negotiate with hotels and airlines

Help your customers get most out of it

Paid night in room

Can I recalculate commission today or return it?

If this location is resolved, total could reach 94%

In following way:

This

Basic algorithm

This is also solved by deep learning

Sure

Recognition of some named objects

Including user intent recognition, CNN-like methods

I won't talk about these methods

This is LSTM+CNN, this is a matter of Q standard fit

This is what QA will get

Let me tell you

When biLSTM can run forward and backward

These two directions receive contextual information

The lCNN model does not deform when translated, scaled, etc.

Could be better

Feature area detection, good reliability

final

Compare multiple experiments

biLSTM+CNN found to perform best offline

I made a map with curves and waves

In following way:

Ctrip Cloud AI Practice

Implementation of AI in Ctrip

The business effect is very good

But biggest technical challenge is actually cloud processing

Each business unit of Ctrip is practically its own industry

The architecture of business-related IT technologies is also different

Other customer service systems

It's completely different

Contrasting technology and business

The product itself is divided into different teams in different business units

This is a very independent team

In this case

How to support more than 60 business lines in a group

Actually

Equivalent to over 60 companies in over 20 industries

Therefore

Full set of similar cloud solutions

While there are some in industry, they are not particularly developed

Including cold start of each business unit

Because it has many customer service scenarios

Even for a business unit

For example, a hotel

There are after-sales and pre-sales consultations

We'll see later

Which hotel would you choose?

Or choose a hotel in which city?

We did it

Cloud system operation

I made some abstractions and formed a body annotation

Students who are familiar with this customer service robot also know

We will coordinate human trainer with background platform

So

The thing is appropriate. We will give an artificial simulator

What to do every day

Some KPI reports to review every week

The data and algorithms of cloud platform will be listed for them

Enable standard Q (body)

How to mine library, corresponding platform will also be indicated

In following way:

What we do

Sure

Including logical cloud business management

Business management process

His flight and hotel may be completely different

In this case, how to use cloud platform

Let every line of business be an AI coach

Full business editing in this self-service service

In following way:

Artificial Intelligence

There are two main things. One of them is a small poetry machine.

There is also a customer service robot

For Ctrip

There are many artificial intelligences

Our company

Ctrip artificial intelligence has also been reorganized and repositioned

What to do

Actually artificial intelligence

This is a very laborious job

Even very small functions

For example, anti-yellow images or anti-political photos

It also requires a lot of manpower and material resources

For example

One million yellow pictures should be marked

Or sensitive images can implement this feature

But in fact, this is a very small functional point for direction of business

To create a large system will require many function points

In this case

How to implement artificial intelligence of whole group

How to update artificial intelligence of this line of business?

Our company

It also uses a concept similar to that proposed by Google Li Feifei

Democratic artificial intelligence

About our AI technical staff

More focus

How we help different business units

How to use artificial intelligence to improve OTA user experience quickly and efficiently?

For example

For business departments such as air tickets and train tickets

How to improve user experience?

We also did an integration in corresponding field

Including a small poetry machine

Lots of imports

Where

Including BAT, Face++ and others, and open source AI capabilities

Democratic and open promotion of Ctrip's achievements in artificial intelligence

In following way:

Our company

A single access platform will be created in public area

We have a systematic approach

A relatively scientific method for model evaluation has been developed

enable

Evaluation, standardization and comprehensive selection

For example

Image recognition: integrates Ctrip introspection

And API BAT,

We will comprehensively evaluate which one is better

How to complement each other

To provide best image recognition in industry for every business unit

This is our Ctrip group

Cloud platform of artificial intelligence (Ctrip Brain)

You can see

Includes features of each module

Some algorithms with ads

There are also some traditional pictures

Some AI for text, voice, etc.

In following way:

Above

Common Functions

For example, classification of places of interest

Send a photo. Find out what's in this photo

Then write poetry

Actually

This thing is very useful on OTA sites

For example, I want to run a spring marketing campaign

We need to collect all photos of spring

Artificial intelligence of entire Xiaoshi machine

It will be dismantled and distributed to various business lines

In following way:

This is artificial intelligence

Let me recap customer service section

Customer Service

This is closely related to business

In this case, our technicians

The entire customer service system will be integrated

There is also an abstraction at algorithm level

After abstraction

Into artificial intelligence, which each functional module supports each line of business

Coach's work

Guarantee when joining a new business line

To our entire technology development team

It costs as little as possible

Work on a new line of business

Support for AI trainers via self-service platform

This part of people are very knowledgeable

For technical team

Especially for social technology students

On contrary, I don't know much about business

Can some tickets be refunded?

I don't know if I can rebook

We do what we do to support these businesses

Get rid of it on platform

Subscriptions

Group personalized cloud platform

Interaction with business

The highest level - personalization

Because we need to promote 20 or 30 business line products at same time

Recommendation logic

It's completely different

Raw data and user behavior data are completely different

In this case

We have also created a corresponding unified recommendation system for groups

100 scenes released

Annual income 210 million yuan

Many deep learning methods are now also available

In the following way:

In smart appointment decisions

If user decides where to play

The decision cycle will be much better than in e-commerce or traditional internet companies

For example

The results of our statistics

If user wants to travel within country

On average, it takes 21 days to make a decision if you travel to Southeast Asia

will make a comprehensive decision within 40 days

460 visits

For example, user selects Southeast Asia

Can make decisions

My son loves to play on beach, my husband loves to play outdoor sports

My wife loves shopping

Together we will recommend attractions

In this way, interests will be combined and we will make some recommendations

We will tell him where price is cheaper

Because

The travel market is not same as e-commerce

The tourism market changes depending on season

The price index changes a lot every day

In this case, we will make big data for it

There are also personalized solutions using artificial intelligence

In following way:

Such a solution

Once decision is made, look at change in price curve

Comprehensive price index including flights and hotels

Here you will be advised on which day to go

User selects approximate time period

For example

I want to go to beach next month

My husband wants to go out and exercise outdoors

Integrated launch of these rides

Suggest cheapest day next month

Corresponding price chart

I mostly do this kind of work

Internet also showed best results

In following way:

This is a recommendation for a scene in a line

Where user goes once

I will recommend things like food, accommodation and travel. This is corresponding aggregation page

In following way:

Entire recommendation system

What I just saw is a product landing form

See how its structure is arranged

We will summarize data of business department

We will use a virtual "standard product"

Standardize user behavior and order data, then make recommendations

In following way:

Architectural

We have adopted this model, our bottom level

These things are in white

Most search engines

Including all reverse flow closed loop and offline mode

Online compute engine

We will make a single set in group, and it will be repeated centrally

Yellow border at top

Basically we make some adjustments for a certain scene

In following way:

Actually

We are still working on some things

We will make it a complete cloud

Let product manager set it up

Let's put it in its entirety

After online and offline organization becomes modular

Who uses our business

You can make your own configuration in this interface

This allows some combinations

For example, use your browsing history to make recommendations

This can be customized

In following way:

In addition

About future of tourism technologies

Future Studies

We can see

According to data, share of mobile Internet users reached 95.1%

much more than

Broadband users in some countries