Product Development Plan

Axion Technologies K.K CEO

Takushi Yoshida

Dec 2019 - Revised on December 31, 2023

​

​

​

1

Image Via Qualcomm Media Center

1.User Understanding

2

Company takes all, employees take nothing

Rise of big techs, Japan lost the “tech war”

End of Lifetime employment

3

Target Users Feel Anxiety

Japanese Business person feels anxiety because ...

- I made user interview it’s at appendix

Public pension is nearly dead

They need to acquire new knowledge and update skills.

Pelsona; 20~40s next gen business person

※Face photos were generated by GAN, developed by facebook AI team.https://carpedm20.github.io/faces/

Hisato Takahashi

20, male, student at Keio Univ (Faculty of Commerce)

​

He feels anxious about job hunting in the near future due to frequent news about Japan's economic decline over the past three decades. Both TV and newspapers seem outdated, and he finds the Internet unreliable as well. He feels there's nothing trustworthy to believe in.

Ken Nakagawa

27, male, Banker at biggest Japanese Commercial bank

​

He believed he had a stable job as a banker, but then large-scale layoffs began at all three of the biggest banks. Services like Alipay and the rise of Bitcoin could disrupt traditional banking in this century. He's on the verge of losing everything he's invested in his career.

Miu Imazu

31, female, Marketerat IT company

​

She's a hard worker who has devoted most of her time to her career since completing her master's program. However, she's now seriously considering marriage and needs to find a time-efficient way to catch up with the state of the IT business.

​

​

​

Toshi

42, male, Head of marketing, at CPG maker

​

He is a traditional Japanese marketer facing the need for drastic change as the game has fundamentally shifted. Lacking knowledge in the new digital marketing methodologies, he is unfamiliar with statistical and computational approaches.

Ken Nakagawa Storyboard

※Face photos were generated by GAN, developed by facebook AI team. https://carpedm20.github.io/faces/

​

Ken Nakagawa

27, male, Banker,at the biggest Japanese Commercial bank

​

​

7:00 crowded train. Read boring articles with Nikkei news app for topic with colleagues and boss. App says banking industriy is declining sharply. He quit and moved to Instaglam. He learned that a college friend got married.

​

12:00 Lunch. The bank was born as a result of the merger of three big bank, and the factions of those remains. He has lunch with the faction he belong to. He is so tired of fitting to talk to old people, the lunch time is very stressful.

​

​

​

​

​

Average Japanese business person are so afraid that their job opportunity last in next decade.

8:00 office. The company has an implicit rule that subordinates must seat before the boss comes. He has to get so many inefficient things done, but those works never improve productivity.. He is so frustrated

​

Ken Nakagawa Storyboard

Ken Nakagawa

27, male, Banker,at the biggest Japanese Commercial bank

​

​

​

15:00 smoking rooｍ. People are talking on layoff. Two competitors lay off 9500 and 4000 each, but The bank will lay off 19000. He is afraid of annual income decline, and you also have to look to change jobs

18:00 with boss. Boss keep saying sarcasm on spreadsheet Ken made over 20 minutes. There were just one spelling mistake. It cost too much. Unproductive...

​

19:30 drinking with people from other department. He promise to play golf with them to get support the project his department proceed. He sacrifice whole weekend in next month when he plan to visit Disneyland with her girlfriend. Ouch.

22:00 tain. Neighbor is reading web magazine featuring bitcoin. The headline is “Does the banking industry end?"

The world is changing. He need to explore the next era. He download axion and started learning on “Tech economy”

People spend so much time for unproductive works. They overfit the firm’s “protocol” and suffer from potential lack of skill.

2. User’s Pains

7

8

Environment: Failure of Ad-model in News Consumption

• Fake news. A system with an incentive to distribute on very poor information

• Privacy concern. Side effect of attention market
• Information explosion. The amount of information far beyond human cognition causes serious matching problems

​

Monumental fake news

Total traffic for broadband subscribers. Source: Ministry of Internal Affairs and Communications

9

​

�Pain #1 : Too much information

• People reach the limits of cognitive function. They spend so much time on consuming “snack content” with digital device.
• Digital connected devices have become dominative in media time spent for this 10 years.
• Total media contact time is 411.6 minutes, the highest ever.

Via メディア環境研究所

10

Pain #2: Low-End Contents Bubble by News Apps

Strategy: Publishers Keep high-end contents in their hand and hand low-quality contents to news app.

Incentive: Advertising revenue is not enough to create high quality content. So they are driven to make crick-bait contents

Publishers

News Apps

Situation: Every news app distribute low-end news. Their lineups are very similar because inventory source. News app also has a feature to be favor of crick-bait, and sensational information.

Users

As a result...

Users Suffering from bad and same contents !

It is like they are in the bubble of junk information.

11

​

�Pain #3 : Ads and Algorithm

I tried 25 different news apps from US, China and Japan. I was very frustrated because...

Addictive. Content selection algorithms are designed to increase advertising revenue, making you addict to the app.

Notification War. "Retention machine" is annoying your healthy life.

Not for me. “Popular articles” are not very interesting to me. Murder case, celebrity’s love affair ...

Snack. Nothing to enrich my life.

3. Solution

12

13

Netflix on News Consumption

• Recommendation of sureley selected news
• A service that aggregates high-quality information enables users to touch the correct information without any loss
• User-centric. Subscription bring clear relationship between customer and service provider.

​

​

​

Recommendation

Certain contents discovery better than hacked search and social

Netflix-nise

Huge long tail with external content

Subscription

Investment cycle to user’s benefit

High-Quality Inhouse Content

High-end lineup created by the high-end editorial

14

Potential needs for educational information on new economic rule !

​

​

​

4. User Experience

15

16

​

Product Concept

Not snack but supplement

• Capacity development. Users have good information every day, and develop their knowledge, without being exhausted with the stupid information
• Community. Conversations with people with the same knowledge and desire enrich user’s life, not socialize or show off yourself or your own.
• Mindfulness. You can change your brain by yourself. Control information that can be exposed to yourself, and try to interact with digital information that is full of happiness

​

Elegant UX

​

https://sketch.cloud/s/YMwLe/eP09Om/play

Bad UX

Continuous elegant experience. You can enjoy analysis on leading Asian digital economy with text, audio and video for ¥ 2,000 / month

​

Joyful. Meet content with customizations and recommendations. Intuitive content selection is facilitated with a modern UI that emphasizes images.

​

High learning effect. Transition from high frequency / short time / low learning efficiency experience to low frequency / long time / high learning efficiency experience with long form contents, smooth animation etc

​

Follow what you like. You can follow categories and authors you like to read.

Recommendation purely for you (not for platform’s advertising revenue) :

We provide contents based on recommendation system for the purpose of contributing to your development

We do not provide addictive “slot machine”. We provide supplements for people’s capacity development.

​

​

Hand Pick by Editors : Editors pick is very valuable in ML algorithm age. This provide

Reputation System：Readers can refer to article and author reputation before reading the article

We utilize the editorial, and wisdom of the crowd

Podcast: Most of commuters around tokyo spend 2 hours for commute train. Podcast must fit usage in super-packed train.

Video: Younger people are more likely to consume video on their smartphones. They don't care much about the video quality of the smartphone video.

Text, Podcast and Video

21

​

Reading history, Visualization

Giving fun to keep reading contents like fitness app

• Your interest
• Recommending users with similar interests
• Daily reading amount
• Reading history

​

​

​

fitness app like design

22

​

Self Generated Magazine

• Automatically generate a magazine composed of content that users are likely to like
• Contents selection based on content-based filtering or collaborative filtering
• Coping Spotify’s “My Daily Mix” or “Your Discover Weekly”

23

Magazine Inside

​

​

• Some us people show preference to handpick after Cambridge Analytica and Russian clacking toward Facebook feed algorithm
• Recommendation system of advertising model driven app is designed to make users addict to the app
• Believing good curators is one of best strategy in late situation

Medium Magazine via Medium

Apple News +

24

Dynamic Game Difficulty Balancing

Dynamic game difficulty balancing (DGDB) for good learning experience.

• We assuming that we can able to apply DGDB into news app.
• If the difficulty level can be quantified, it is assumed that content with the appropriate level of difficulty to the player's skill can be suggested.
• We hope DGDB take users into “flow” (feeling of complete and energized focus in an activity, with a high level of enjoyment and fulfillment).
• If the difficulty level of the content is gradually increased according to the reader's skill, the user can enjoy high efficiency learning experience.

​

Via gamasutra.com

5. Two-Sided Marketplace

25

Question: What Kind of Marketplace Is Favorable for News Readers, Creators, and the Society?

5.1 Two Sided Marketplace

Creators

Readers

Supply

Demand

Marketplace

A two-sided market occurs when two user groups or agents interact through an intermediary or platform to the benefit of both parties.

including writer、journalist、publishers...

5.2 Best Scenario: Network Effect

Same Side Network Effect

More publishers make the platform better, attracting more publishers

Publishers

Same Side Network Effect

Readers benefit from more Publishers

Readers

​

Publishers benefit from more Readers

More readers make the platform better, attracting more readers

5.3 Benefits creates stickiness

​

​

Creators

Readers

Supply

Demand

Marketplace

• Integrate tools, services, and programs to provide better services to creators and enable the building of a more effective business with Axion.

• Improve the size and quality of accessible content
• Improvement of UX
• High-quality recommendations

​

5.4 Personalization is key

Creators

Readers

Supply

Demand

Investing in machine learning to improve content-user matching, enhancing engagement from both sides."

Personalization accuracy determines engagement.

​

​

5.5 Price is the most important indicator

​

​

Creators

Readers

Supply

Demand

Transactions between demand and supply are primarily managed through pricing, with recent adoption of 'dynamic pricing' that adjusts to supply and demand.

Price

5.6 Fixed Pricing: A Game of Increasing Benefits

Creators

Readers

Supply

Demand

Monthly

¥2,000

Expecting Benefits Commensurate with Price

Providing value equivalent to price to attract as many readers as possible

It has been demonstrated that when both sides reach a critical mass, positive feedback becomes stronger, leading to stickiness among participants.

Until reaching a certain scale, it is necessary to be willing to make upfront capital expenditures.

5.7 Optimizing Strategies of Netflix and Spotify

Netflix consistently operates near the break-even point, reinvesting the recurring revenue from paid subscriptions primarily into content acquisition and creation.

Over 70% of Spotify's cost of sales is allocated to payments to artists, resulting in the company continuously recording a slight deficit.

As a result, both entities are enhancing network effects in the two-sided market, thereby increasing their long-term cash flows and achieving sustained growth.

​

5.8 Keep solving matching problems!

​

​

Content Pool

Magazine

Newspaper

Search

Browse

​

Users

Creators

Readers

6. How to build Engagement

6.1 Characteristics of the two-sided market for professional content

​

​

• Netflix and Spotify, along with other successful two-sided subscription-based markets, demonstrate remarkable success.
• Both prioritize high-quality content and strategically invest to align closely with user preferences.
• Netflix specifically leverages original content as a crucial tool for attracting and retaining subscribers.
• A substantial portion of their monthly subscription revenue is strategically channeled back into content creation, showcasing a focused approach to balancing income and expenditures.

6.2 Strategies for Fostering User Attachment in Two-Sided Markets

ブロードバンド契約者の総トラヒック。出典:総務省 総合通信基盤局

Ranking

​

Content

Selection

​

​

Feedback

​

Content

Consumption

​

Mobile

Laptop

Desktop

Tablet

Recomendar Algorithm

Editors Pick

News Articles

Investigation

Special Report

Video

Podcast

V-Tuber

Visualization

​

Views

Rating　

6.3 Liquidity is the key to growth.

​

• Increased liquidity enhances user benefits, thereby generating liquidity network effects.
• The greater the quantity, the more users can reap benefits.

​

More

Liquidity

1

Users enjoy more content option

2

Creators

Supply

3

More Users

​

4

More Users pay more monthly charge

5

More

Creators

Axionにおける流動性ネットワーク効果

6.4 Correlation Between the Amount of Content and User Numbers?

Spotify adds around 20,000 songs monthly, growing its user base and recently achieving profitability with a catalog of approximately 40 million songs.

​

​

Netflix has acquired 158 million paid subscribers, with diversified content attracting an increasing number of users outside the United States.

Increasing the number of contents is the right strategy

while ensuring quality, of course.

Content is not homogeneous

• For Uber, homogeneity in services leads to increased liquidity. However, for content platforms, the diversity of types and varied user tastes must be considered. Merely increasing quantity isn't sufficient; the goal should be a wide array of content categories catering to the broadest range of users.

6.6 Various Content Categories

43

Quantum Computer

Machine Learning

AR / VR

IoT

Fin-tech

ドローン

遺伝子編集

５G

低炭素

宇宙

Blockchain

Cryptocurrency

Datascience

Google

Facebook

Amazon

Apple

Microsoft

Baidu

Alibaba

​

​

​

Tencent

Huawei

UX

product management

productivity

startups

venture capital

Basic Income

都市

education

Future of Humanity

Digital Marketing

Biotech

Cloud

Cybersecurity

Enterprise

Telecom

Mobility

Security

Digital Transformation

​

Digital Economy 40 Categories

ビジネス英語

ビジネス交渉・心理学

ビジネスマナー

ビジネスライフ

ビジネス企画

ビジネス文書

プレゼンテーション

仕事術・整理法

環境とビジネス

リーダーシップ

Word・Excel・PowerPoint

ビジネスコミックス

経済学

経済思想・経済学説

経済史

財政学

各国経済事情

経済協力

一般

商業デザイン

セールス・営業

マーケティング・セールス　全般

商品開発

広告・宣伝

金融 ファイナンス

会計学

会計・会計学入門

財務会計

財務管理

財務諸表

税務会計

国際会計

各国会計論

キャッシュフロー

簿記

オペレーションズ

マネジメント・人材管理

統計学・数学

経営戦略

ロジカル・シンキング

アメリカMBA・名物教授

ビジネス英語

資格・就職・MBA

経営学

経営理論

経営管理

経営診断

企業経営

企業動向

CI・M&A

海外進出

売買契約

​

​

株主総会・取締役会・会社継承

環境とビジネス

歴史に学ぶビジネス

リーダーシップ

ビジネス人物伝

企業・経営

金融・銀行

世界経済

自伝・伝記

ビジネス法入門

建築関連法規

薬事法

不動産

会社法

債権・物権法

労働法

商法

手形・小切手法

株主・監査役

独占禁止法

登記法

税法

製造物責任法

​

business/economy 70 Categories

農林水産

建設・住宅

食品

資源・エネルギー

製造・加工

鉄鋼・化学

自動車・機械

電気・電子

病院・医薬品

サービス・小売

銀行・金融

不動産

交通

流通・物流

産業史

その他

​

Industry 16 Categories

https://docs.google.com/spreadsheets/d/10XuyA_sQEFcvg2di_eoXMpZ6SXcKKKpbRcGIemQQcPA/edit?usp=sharing

6.7 Content Acquisition could be like Puzzle

Group A

Various contents are allocated to various groups of users.

Group B

Group C

Group D

Group D

Determining the most effective order in which to combine contents becomes a considerably challenging question.

Look for approximations rather than optimal solutions while addressing realistic challenges.

7. Road Map

46

47

Product Development Plan

The Athletic

Stage 1

Stage 2

Stage 3

• Plain web blog for exploring PMF
• You can customize your feed through following
• Recommendation engine serve what customers want

Web App

Web App

Web / Mobile app

blog

Spotify

Seed stage

(6~12 month)

Series A

(12 month)

Series B

(12 month)

48

Stage1 Plan Web Blog

https://www.axion.zone/

• Increase the number of articles and prepare for future development
• Trying to be liked by search engines

​

49

Stage2 The Athletic

Follow teams, regions, authors etc. Tagged contents flow into your feed

Exclusive stories from an all star team of writers

Score of teams user follow and editors hand picked carsel.

Users can check scores, statistics and stories of following teams

Follow and check what you are interested in

50

Stage3 Spotify like

• The combination of Editors Picks and the recommended system is the latest trend
• Recommendations are made in a wide variety of combinations

Amazon Prime Video’s Various Recommendation

8. Engineering

51

52

8.1 Blog

https://www.axion.zone/

leveraging PaaS Ghost, we can run blog without engineers now, controlling things at dashboard.

[Ghost: The #1 open source headless Node.js CMS](https://ghost.org/)

​

53

8.2 The Athletic-like web app

Single page application

• Vanilla JS or modern JS library
• Authentication: Auth 0
• Subscription: Stripe
• Firebase or AWS Amplify
• CDN: Fastry
• Headless CMS: Contentful or Wordpress
• GraphQL
• Elasticsearch

​

プラットフォーム

• ウェブ＋アプリ

機能

• サインイン／ログイン = Auth0
• サブスクリプション決済 = Stripe
• データ基盤（Collection / Aggregation / Analytics）
• 集中ロギング（Fluentd, Elasticsearch, Kibana）
• 簡易な推薦システム（Most Popular等）
• 検索（Elasticsearch）
• Rating System（Authors / Items）
• Headless CMS（or WordPress + AWS S3）、会社側、寄稿者、パートナー向け
• クローラ
• CDNによる配信最適化
• ペイウォール

​

​

54

8.21 The Athletic-like web app

55

8.22 Subscription and Auth

Auth 0

• Secure ｍanagement of user identity
• Multiple login way like iD/password, Social Login, and passwordless.
• Providing advanced authentication functions such as single sign on (SSO) and multi-factor authentication

​

Stripe Billing

• Controlling subscription process and sending invoice on dashboard

​

​

56

8.3 CDN

• Handling a large amount of traffic
• Low latency
• Fast cache clear
• Need to be quick and quick, We frequently update article
• Comfortable browsing from overseas
• Axion plans overseas expansion from the beginning

​

Via 「CDNを活用した日経電子版の ネットワーク最適化とサイト高速化」 (日本経済新聞社 宍戸 俊哉)

​

​

​

57

8.4 Fastly

• Instant Purge lets us update stale content within 150 milliseconds or less.
• Number of enough POPs (point of presence)
• Flexible configuration change in VCL, manageable through coding
• SSL
• IPv6
• HTTP/2

​

58

8.5 Log Monitoring with Fastly

• Export access log to S3 using “Real Time Log Streaming” and trigger it on Elasticsearch with log file generation event

​

59

8.6 Spotify like

• [Short term] Mobile and web apps on FaaS (Firebase or Aws Amplify)
• Java / Kotlin and Swift, not relying on cross platform library
• [Long term] Several clients based on micro services
• BFF for each client
• Leverage many of public APIs
• Recommendation System
• Customized CMS
• K8s
• Data analytics infrastructure
• Marketing tech stack
• MLOps

​

​

9. Data Infrastructure

60

61

How Should We Build Data Analytics Infrastructure

Pypeline: collection, accumulation, processing, utilization

• We need architecture with several layers? 🤔
• Hadoop data lake for machine learning Engineers
• Marketing Data Warehouse or CDP (Customer Data Platform) - connecting marketing-tech stack
• DWH for data visualization for business or editors

62

DWH vs Data Lake

63

DWH vs Data Lake

A data warehouse is a database optimized to analyze relational data coming from transactional systems and line of business applications. The data structure, and schema are defined in advance to optimize for fast SQL queries, where the results are typically used for operational reporting and analysis. Data is cleaned, enriched, and transformed so it can act as the “single source of truth” that users can trust.

A data lake is different, because it stores relational data from line of business applications, and non-relational data from mobile apps, IoT devices, and social media. The structure of the data or schema is not defined when data is captured. This means you can store all of your data without careful design or the need to know what questions you might need answers for in the future. Different types of analytics on your data like SQL queries, big data analytics, full text search, real-time analytics, and machine learning can be used to uncover insights.

​

Data Infrastructure

64

Client Side

DWH

Data Analysis

Data Visualization

Machine Learning

Server Side

Elasticsearch

Database 3

Database 2

Database 1

Agregation

Application

65

Not DMP, But MDW

We chose marketing data warehouse (MDW) not CDP (Customer Data Platform) or DMP - connecting marketing-tech stack

​

• DWH for data visualization for business or editors

Flow from data capture to remarketing decision making

​

​

66

Log Analytics

• Centralized Logging: Fluentd, Elasticsearch, Kibana on aws…
• We analyze website, mobile device and server logs for digital marketing, application monitoring, fraud detection and advertising.
• Stackdriver Logging(GCP) or Amazon Kinesis + Elasticsearch Service(AWS): These solutions collect, ingest, process, and load both batch and streaming data so that the processed data is available to users in near real time on the analytics system you are already using.

​

Centralized logging on aws, made up of Fluentd, Elasticsearch, and Kibana. https://aws.amazon.com/jp/solutions/centralized-logging/

67

ETL (Extract, Transform, Load)

DWH essential workflow

• ETL is defined as a process that extracts the data from different RDBMS source systems, then transforms the data (like applying calculations, concatenations, etc.) and finally loads the data into the Data Warehouse system.
• ETL provides a method of moving the data from various sources into a data warehouse. In the first step extraction, data is extracted from the source system into the staging area. In the transformation step, the data extracted from source is cleansed and transformed.

Reference: Cookpad’s Data Infrastructure

68

69

MLops Workflow (1)

1. Fetch the data— You might have in-house example data repositories, or you might use datasets that are publicly available. Typically, you pull the dataset or datasets into a single repository.

​

2. Clean the data—To improve model training, inspect the data and clean it as needed. For example, if your data has a country name attribute with values United States and US, you might want to edit the data to be consistent.

​

3. Prepare or transform the data—To improve performance, you might perform additional data transformations. For example, you might choose to combine attributes. If your model predicts the conditions that require de-icing an aircraft, instead of using temperature and humidity attributes separately, you might combine those attributes into a new attribute to get a better model.

70

MLops Workflow (2)

4. Training the model— To train a model, you need an algorithm. The algorithm you choose depends on a number of factors. For a quick, out-of-the-box solution, you might be able to use one of the algorithms that Amazon SageMaker provides.

5. Evaluating the model—After you've trained your model, you evaluate it to determine whether the accuracy of the inferences is acceptable. In Amazon SageMaker, you use either the AWS SDK for Python (Boto) or the high-level Python library that Amazon SageMaker provides to send requests to the model for inferences.�You use a Jupyter notebook in your Amazon SageMaker notebook instance to train and evaluate your model.

6. Deploy the model— You traditionally re-engineer a model before you integrate it with your application and deploy it. With Amazon SageMaker hosting services, you can deploy your model independently, decoupling it from your application code.

​

ML Workflow (3)

71

https://www.slideshare.net/JulienSIMON5/mlops-with-serverless-architectures-october-2018

10. Marketing Tech

72

73

10.1 Our “Marketing” Definition

CRM: Engagement with existing customers and a continuous approach to prospective customers

​

Digital Marketing: Acquisition of customers in a wide area, discovery and nurturing of prospective customers, expansion of awareness

​

​

​

​

​

​

CRM

Engagement with existing customers and a continuous approach to prospective customers

Digital

Marketing

Mail, social, and closed event ..

Acquisition of customers in a wide area, discovery and nurturing of prospective customers, expansion of awareness

Adtech, Marketing tech and open event...

74

10.21 Analytical CRM

We use CRM partially, since our business is “SaaS without sales person”. We use "Analytical CRM" , mail marketing(including newsletter), and closed event.

​

"CRM" refers to concept that has direct contact with customers towards back office system.

​

"Operational CRM" consists of SFA (Sales Force Automation) that performs as a customer contact, customer support and field service. "Analytical CRM" consists of analyzing customer data and recommend the best offer for the customer.

​

75

10.22 Marketing Tech of Netflix

Campaign Management and Advertising platform tools

Ad Budget Optimization System

​

This system work with data scientist or economist analyzing advertising effect

The Campaign Management Service relies on a variety of technologies to achieve its goals.

Engineering to Improve Marketing Effectiveness (Part 3) — Scaling Paid Media campaigns

https://medium.com/netflix-techblog/engineering-to-scale-paid-media-campaigns-84ba018fb3fa

​

​

11. Recommender System

76

77

11.1 Recommendation System in Axion Two-Sided Marketplace

• In the two-sided marketplace, it is important to be able to maintain engagement between the demand side and the supply side. One solution is to improve the matching of supply and demand. The recommendation system is an effective means
• The scale, convenience, and speed of such marketplaces are enabled by recommendation systems and search engines that use predicted relevance to match buyers to suppliers.

​

* Takushi Yoshida. 「Spotifyの両面市場に働くネットワーク効果とそれを促す機械学習を検討する」Axion https://www.axion.zone/spotify-two-sided-marketplace/

78

11.1.1 Objective of Recommender System in Axion

• Welfare or wellbeing. Since Axion does not adopt advertising model, it seeks the welfare of users rather than the click through rate closely related to advertising revenue. The goal of our recommendation system is not to be spinal reflex drawn by sensational information, but to bring knowledge and optimize total happiness in life.
• Users are suffering from information overload outside Axion. Axion's recommender system is a way to help such people meet the information they need to meet.

79

11.1.2 What do we set for KPIs?

Spending a certain amount of time in a space full of quality products

Time spent. We appreciate that the increased residence time of high-quality articles extends readers' knowledge and increases welfare.

Good contents aggregation is the basis for this KPI, time spent. Axion assumes that staying time in a space limited only by high-quality content generally expresses high user interest and is likely to have a high learning effect.

​

80

11.1.3 Considering User’s Intent

Various intents when making a recommendation

• Get the information you need in a concise manner
• I want to read one solid article
• I want to take my time and use it like reading a book
• I want to use it on the train or in the car like a podcast or audiobook
• I want to search for information to research my work

​

Intent to recommend

• I want you to recommend only what I like
• I want you to recommend things I don't know yet but will surely like
• I want to be surprised by serendipity (exploratory)
• Don't want to be continuously disappointed

​

​

​

​

​

81

11.1.4 News Recommendation Difficulties and Countermeasures

Difficulty

• News value has time decay tendency
• User interest has changed very quickly
• Application of existing algorithms is not possible

Measures

• Remove straight news consumed instantly from products. We increase slow content such as magazines and books in the marketplace
• Even when dealing with news, a human-edited small newspaper

​

The Economist:

Slow Content and Valuable Content

82

11.1.5 Insights into User Behavior

​

• "The Dynamics of Onlookers (Yoneda, 2018). People want to flock to news that others are gathering around.
• This is modeled mathematically, similar to approaches in behavioral economics.
• To what extent can this predict user behavior?
• The rate at which human interests shift is high (hence, information related to professions and the like is not swayed by these changes in interest).
• Micro-moments: Instantaneous, emerging desires.
• Self-improvement desires: The urge to outdo others.
• Self-expression: A desire to acquire knowledge for self-expression."a

​

​

​

​

83

11.2 Discovery

• Customers need “discovery”
• The recommendation system should present the best thing among things that would never be bought unless recommended, and this cannot be measured by "accuracy" in the machine learning term.
• Between serendipity and recommendation

​

​

​

84

11.3 Classical Method

• Top N Recommendation
• Model-based
• Demographic Filtering
• Content-based filtering
• Collaborative Filtering
• Hybrid filtering
• Combined content-based and hybrid filtering

​

J. Bodadila et al. Recommender systems survey. Knowledge-Based Systems, Volume 46,2013, Pages 109-132, ISSN 0950-7051, https://doi.org/10.1016/j.knosys.2013.03.012.

​

​

85

11.3.1 Content Based Filtering

• Recommendations based on similarity of item characteristics
• Domain knowledge required, but recommendations can be made even with a cold start

User preference profile. Cosine similarity widely used, although there are many types of similarity functions. How well does Euclidean distance work?

​

86

11.3.2 Collaborative Filtered

• "Customers Who Bought This Also Bought" feature.
• No need for domain knowledge.
• Addresses the cold start problem.
• In collaborative filtering, the issue is that unless there's a significant shift in users' historical ratings, it tends to repeatedly identify similar groups, leading to similar item recommendations each time.

​

​

​

​

87

11.3 Transition of Recommendation

• GroupLens
• Item-based Collaborative Filtering
• Matrix Factorization
• Factorization Machines
• Recurrent Neural Network

​

• GroupLens: It is a recommendation system primarily based on collaborative filtering. This system analyzes a user's past preferences and behaviors to make recommendations based on data from other users with similar tastes.
• Item-based Collaborative Filtering: Find items that are similar to that item and use the rating values attached to similar items to make predictions.

​

​

​

88

11.3.1 Transition of Recommendation

3. Matrix Factorization: Use implicit ratings, which users do not explicitly rate but which can be retrieved by the system.

4. Factorization Machines: A general model that can be used for a variety of problems, including the identification problem, which was introduced in 2010. User and item attributes.

5. Recurrent Neural Network: Recommendation of information and items that users really want "right now" by learning patterns of user preferences based on time-series data.

​

​

89

11.4 Application of Contextual Bandit to News Recommendation

In the setting where context x is given when choosing each arm, there are the following problems with collaborative filtering and content filtering, which are the main methods of traditional news recommendation:

• User information is required.
• It's challenging to keep up with rapid changes in content.
• Learning and computation need to be quick.
• New content lacks user information, so collaborative filtering cannot be used.

​

90

11.5.1 Bandit Algorythm

ε-greedy, UCB, and Thompson Sampling involve:

• Experimenting with various pieces of information (exploration).
• Recommending items that users click on more frequently (exploitation).
• Iterating between exploration and exploitation.
• The aim is to recommend optimal items under limited conditions.

Lihong et al. 2010. A Contextual-Bandit Approach to Personalized News Article Recommendation. Yahoo Labs.

​

91

11.5.2 ε-greedy

The ε-greedy algorithm is a simple yet effective approach commonly used in reinforcement learning, particularly in the context of multi-armed bandit problems. The basic principle of this algorithm is to balance exploration (trying out new things) with exploitation (leveraging what is already known to be effective). Here's how it works:

• Random Exploration: With a probability of ε (where ε is a small value, typically between 0.01 and 0.1), the algorithm randomly selects an action. This random selection is the exploration part, allowing the algorithm to try different options and not just stick to what it already knows.
• Exploitation: With a probability of 1-ε, the algorithm chooses the action that has the highest estimated reward based on past experiences. This is the exploitation part, where the algorithm leverages its accumulated knowledge to make the best decision.

​

92

11.5.3 UCB

• The Upper Confidence Bound (UCB) algorithm, used in reinforcement learning and multi-armed bandit problems, offers a balance between exploring new options and exploiting known ones. It has several advantages, including dynamic adjustment of exploration rates, incorporation of uncertainty in decision-making, and often superior performance in non-stationary environments without the need for a predefined randomness parameter.
• However, UCB has its drawbacks: it's more computationally intensive than simpler algorithms like ε-greedy, requires tuning of its parameters, can be less adaptive in highly dynamic environments, and assumes relatively regular reward distributions. Overall, UCB is a sophisticated choice for exploration-exploitation balance, but it may require careful implementation and parameter adjustment.

​

​

93

11.5.4 User Clustering

By applying bandit algorithms with different tunings for each user cluster, accuracy can be improved. A famous example is from Netflix, as highlighted in “How Netflix Wants to Rule the World: A Behind-the-Scenes Look at a Global TV Network”

Netflix divides its 93 million users around the world into 1,300 “taste communities”

“Netflix is now dividing up its subscriber base into 1,300 taste communities, which are solely based on past viewing behavior. Each and every user can belong to multiple such communities, and all of these communities spread across the globe. Sure, Yellin admitted, German comedians may be more popular in Germany, but there’s also plenty of users in the U.S. who turn into their shows.”�

​

94

11.5.5 Spotify Home

The method uses Contextual Bandits for personalized recommendations, learning if the explanatory text and music match the user's preferences. This approach not only aligns content with user preferences but also offers relevant explanations, enhancing recommendation personalization and relevance.

• The bandit approach selects both items and their explanations specifically for the Spotify homepage.
• It employs a Factorization Machine integrated with ε-greedy exploration to navigate through a personalized set of options.
• Counterfactual minimization is utilized to effectively train the bandit model.

​

95

11.5.6 Artwork Personalization at Netflix

Requirements:

• Pre-arrange the composition and creative aspects of thumbnails.
• Display based on the user's context.

Contextual Bandit

• In Contextual Bandits, optimizing rewards becomes more efficient by dynamically adjusting the probability distribution parameters of arm rewards based on context. This approach, differing from classical bandits that rely only on arm reward information, involves continuously estimating and updating the linear model parameters of reward distributions, thus allowing for more informed arm selection tailored to the current context.

​

96

11.6 Multi-Stakeholder Recommendation

• Traditional recommender systems focus solely on optimizing the utility of the end-user, the recipient of the recommendations.
• Recommendations considering the utilities of multiple stakeholders add complexity, due to factors like stakeholder relationships, dynamic elements, and the interplay of utilities among different stakeholders.
• In Axion, the main stakeholders are the operators, content providers, and users.

​

Yong Zheng. 2019. Multi-stakeholder recommendations: case studies, methods and challenges. In Proceedings of the 13th ACM Conference on Recommender Systems (RecSys '19). Association for Computing Machinery, New York, NY, USA, 578–579. https://doi.org/10.1145/3298689.3346951

Himan Abdollahpouri, Robin Burke. Multi-stakeholder Recommendation and its Connection to Multi-sided Fairness. arXiv:1907.13158 [cs.IR]

​

​

97

11.7 Dynamic Ensemble of Contextual Bandits to Satisfy User's Changing Interests

​

• Recommender systems must navigate highly non-stationary environments due to rapidly changing user interests.
• Traditional methods require periodic model rebuilding, incurring high computational costs, yet they fail to adapt to sudden trend changes from timely information.
• Changes in reward distributions in non-stationary environments can also be context-dependent, and models should be reused for better prediction when changes are orthogonal to the given context.
• This work focuses on contextual bandit algorithms for adaptive recommendations, leveraging context-dependent reward changes to address the non-stationary environment.

Qingyun Wu, Huazheng Wang, Yanen Li, and Hongning Wang. 2019. Dynamic Ensemble of Contextual Bandits to Satisfy Users' Changing Interests. In The World Wide Web Conference (WWW '19). Association for Computing Machinery, New York, NY, USA, 2080–2090. https://doi.org/10.1145/3308558.3313727

​

98

11.7.1 Dynamic Ensemble of Contextual Bandits

• Bandit Experts (Collection of Contextual Bandit Models):
• These models are developed for a stationary state and are instrumental in estimating the underlying reward distributions.
• Bandit Auditor (Evaluator of Expert Performance):
• This component is responsible for monitoring the quality of each expert's reward estimation and determining their suitability (admissibility) for specific arms under the given conditions.
• Arm Selection by Admissible Experts:
• The selection of arms is based on the consensus from the set of experts that the auditor has deemed admissible for accurately estimating rewards for each arm.

The process involves the bandit auditor evaluating the bandit experts, with the selected, suitable experts then making arm selections for the environment in question. This method leverages a comprehensive suite of bandit models to enable efficient estimation in stable conditions, with model reuse for consistent arms and reconstruction for arms experiencing fluctuating rewards.

​

99

11.8 Evaluation of User Behavior

Evaluating Unrated and Undiscovered Items: It's unclear whether a user's lack of rating or engagement with an item is due to unawareness or preference. When a user has left a review, even a low rating can be considered a negative example, indicating some level of interest in that item. However, identifying items that have garnered no interest is challenging.

​

​

​

This is a research study utilizing datasets from Spotify and ByteDance. The study focuses on using post-click feedback, like skips and completions, to improve content recommendations, addressing the limitation of clicks that only capture initial user interest. developed probabilistic framework, which merges click and post-click data, enhances recommendation models and outperforms traditional methods by significant margins in AUC for short videos and music. The approach's effectiveness across various content types and the balance in user feedback signals are also discussed.

Hongyi Wen, Longqi Yang, and Deborah Estrin. 2019. Leveraging post-click feedback for content recommendations. In Proceedings of the 13th ACM Conference on Recommender Systems (RecSys '19). Association for Computing Machinery, New York, NY, USA, 278–286. https://doi.org/10.1145/3298689.3347037

​

​

100

11.8.1 Features employed by Netflix

• “Your interactions with our service (such as your viewing history and how you rated other titles)”
• “Other members with similar tastes and preferences on our service, and information about the titles, such as their genre, categories, actors, release year, etc.“

​

In addition to knowing what you have watched on Netflix, to best personalize the recommendations we also look at things like:

​

• the time of day you watch,
• the devices you are watching Netflix on, and
• how long you watch.

​

How Netflix’s Recommendations System Works

https://help.netflix.com/en/node/100639

​

​

101

11.8.2 Features employed by Axion

Article Metadata

• Author, Tags, Description.

Things that could be made into features.

• Page view, time spent, heatmap, eyeball tracking, social shares, social responses, behavior after reading the article (if leaving without leaving vs. browsing several other articles) 🤔

​

102

11.9 Opt-Out from Recommendations

• Implementing an Opt-Out Option for the Recommendation System: It's necessary to allow users the choice to opt out of the recommendation system.

​

Features for Exploring Non-Recommended Content: Providing functionalities to explore content beyond what is recommended:

• Customization Feature: Enabling a tailored experience based on user follows is crucial.
• Advanced Search Functionality: Highly important for facilitating user-driven content discovery.

103

11.9.1 Filter Bubble in Recsys

The problem of filter bubbles arise in recommender systems when the user preference model is trained on data that is confounded by the recommender.

Cycling within the same area of interest may be acceptable for some, but for those seeking 'discovery,' it can become a source of frustration after a while

​

104

11.9.2 Difference from the Others

Axion differs from existing news services such as Yahoo News in its business model, type of content supplied, and user intent. The two can be considered as separate genres of disciplines.

​

Similar example: Netflix vs Youtube

​

Competitive Recommendation System

Yahoo News

• Manpower

Smartnews

• demographic filtering

Gunosy

• content based filtering

Ads

Subscription

High Quality

Straight News

105

11.9.4 Ads News Aggregators

Multi-stakeholder, multi-objective, difficult to meet each objective.

• Operator: maximize revenue
• Suppliers: Maximize revenue from supplied content
• Advertisers + ad agencies: Optimize advertising effectiveness
• Users: enjoy articles that match their interests for free
• Operator's KP: longer dwell time and PVs per hour (a factor that maximizes ad revenue)
• As a result, users are forced into the "inherently unfortunate" situation of being recommended clickbait, sensational, low-quality articles
• The nature of the industry makes it easy for advertisers + ad agencies to
• ​

​

​

The different objectives of the four parties can create losers within the stakeholder and distorted incentive designs. An unsolvable problem.

106

11.9.5 Subscription News Aggregators

Demand

Supply

Subscription model + two-sided market

• Operator: Maximize revenue
• Suppliers: maximize revenue from supplied content
• Users: enjoy articles that match their interests
• KPI: Increase user subscriptions = optimize customer satisfaction

​

As a result, there is no conflict between optimizing customer happiness and welfare and maximizing platform revenue. Users will enjoy these benefits for the cost of the monthly subscription.

​

Intermediate

Paid Subscription

Loyalty

To promote subscription retention, the bulk of the subscription fee is allocated to the supplier.

​

107

11.9.6 Concept of Axion Recommendation

• Reduce the share of straight news, which quickly loses value, shifting towards a usage model more similar to magazines and books.
• Enhance the quality of the article inventory, removing lower-quality content.
• Secure login to consolidate user behavior under unique IDs, allowing for deeper insights into users.

Advantages: This enables a recommendation approach different from ad-based systems, similar to the in-depth recommendations of Spotify or Netflix. Their inventories consist of 'stock-type' content with slower value depreciation. Deep user insights based on IDs, along with rigorous content tagging and feature extraction, make such recommendations achievable.

​

108

11.9.7 No competition

• Yahoo! News, SmartNews, Gunosy, and similar platforms are not direct competitors. Their user intent is to utilize spare moments, with businesses primarily interested in targeting these moments with digital ads. Their competition lies with casual games, social media, messengers, and video apps.
• Axion's target audience is business professionals seeking knowledge for work and self-improvement, collecting high-quality information. It's assumed that their disposable income for paid subscriptions also extends to books, magazines, seminars, and newspapers like Nikkei.
• Our competitors are more likely to be seminars, sports gyms, new books, and financial newspapers like Nikkei. Axion is unlikely to directly compete with ad-based models, and while there

109

??? What we need to achieve in RS

• Quick Response
• When user make request at client, the system push back contents to client in 100 millisecond.
• The server accepting the request is indexing and updating the search asynchronously with the request.
• User Behaviour Modeling
• Regarding articles as embedded continuous vectors, by using common techniques of distributed word representation
• Mathematical modeling of typical psychology
• Realtime reflection. The model instantly reflects the state of user behavior

110

??? What we need to achieve in RS

• Approximate Nearest Neighbor Search
• Spotify / Annoy https://github.com/spotify/annoy
• Gann = Go-Approximate-Nearest-Neighbor
• https://github.com/mathetake/gann
• Almost same algorithm as spotify / Annoy

​

Reference

Mathetake “ニュース推薦システムにおける 機械学習の活用事例

人工知能学会全国大会 2018 ランチョンセミナー”

https://speakerdeck.com/mathetake/niyusutui-jian-sisutemuniokeru-ji-jie-xue-xi-falsehuo-yong-shi-li?slide=44

​

​

12. Paywall

111

12. Paywall Types and Basic Tactics

News Product subscriptions have a "paywall" billing structure

• Hard
• Metered
• Freemium
• Hybrid of meter and freemium
• Donation
• Dynamic

12.1 Dynamic Paywalls

• Predicting the likelihood of users converting to paid subscriptions based on estimated attributes, behaviors, location data, and referral information, and deciding actions based on these predictions.
• By presenting subscriptions based on user behavior analysis, there's a tendency for higher conversion rates.

This example utilizes a user model that balances utility and cost to optimally determine when to display a paywall.

Heidar Davoudi et al. "Adaptive Paywall Mechanism for Digital News Media" http://www.cs.yorku.ca/~aan/research/paper/KDD18.pdf

12.2 Markov decision process

To model the dynamic environment of users, we use Markov processes. A Markov Decision Process (MDP) is a probabilistic control process in discrete time. At each time step, the process occupies a certain state, and the decision-maker can choose any available action within that state.

​

​

A simple example of a Markov decision process used in dynamic programming and other applications. Via Wikimedia Commons

Heidar Davoudi et al. "Adaptive Paywall Mechanism for Digital News Media" http://www.cs.yorku.ca/~aan/research/paper/KDD18.pdf

12.3 Dynamic paywalls

Dynamic paywalls can be viewed as sequential decision-making problems. At each timestep when a user requests an article during a session, it's necessary to decide whether to allow the user to read the requested article or to present a paywall.

This is similar to how a professional chess player makes decisions at each turn.

​

Model the environment using a Navigation Graph based on the Markov Decision Process. Develop a model that strategically determines the optimal timing to present a paywall, considering the estimated utility and cost.

Heidar Davoudi et al. "Adaptive Paywall Mechanism for Digital News Media" http://www.cs.yorku.ca/~aan/research/paper/KDD18.pdf

12.4 Dynamic paywalls by WSJ

• The Wall Street Journal has developed a 'Dynamic Paywall' using machine learning to identify where visitors are in the 'purchase journey' and take actions based on the predicted likelihood of subscription. The dynamic paywall is activated up to 15 million times a week.
• At WSJ, a model is built based on 60 variables, creating a 'Subscription Likelihood Index' scored from 0 to 100. WSJ makes proactive proposals to readers with higher indices, but all visitors can access up to five free articles, after which the conversion rate drops.

WSJ can flex based on audience but as far as the consumer sees, WSJ are a freemium paywall.”

12.5 Paywalls and Marketing Funnel

• For high-performing publishers in our study, optimizing each step of the subscription process is essential for success. In addition to this optimization, initiatives like Facebook Instant Articles, Facebook Analytics, Google's 'Subscribe to Content with Google', and Apple's iOS subscription program provide publishers with opportunities to streamline these subscriptions.

12.6 Factors Impacting Conversion Rates

1. Audience Engagement: the propensity of their audience to subscribe (readers’ level of engagement and their perception of value in the product)

2. Targeted Offers: the effectiveness of a publisher’s marketing messages and offers

3. Frictionless Purchase: the ease of actually completing a subscription purchase once a reader has decided to do so.

Nicco Mele, Matthew Skibinski, Matthew Spector, Harvard Kennedy School ”Digital Pay-Meter Playbook"

12.7 Pricing Variation

• Deconstruct bundles and create price diversity by offering individual parts for readers with partial product subscription intentions

13. Contents

120

121

​

News Consumption Cycle

So many conversations on the topic. People read just title of the article and talk about that...

People lost curiosity and move to the other trending topic

time

Traffic

Sticky people still talk about the topic. They have meaningful discussion on it

We think people who talk on the topic at this period can be loyal customer

122

Not Quick...

Book

Weekly Magazine

News

Monthly Magazine

Consumption Speed

“Stock”

Our Position

“Flow”

Japan

US & EU

123

​

​

Qurated and Personalised

Curated by world-class editors.

Axion editors handpick the best stories and deliver them right to you. These must reads include everything from the latest headlines to in-depth special-interest pieces. Axion subscribers can also unlock premium article and magazine selections.

​

Personalized to your interests. As you read, Axion gets a better read on your interests, then suggests stories relevant to you. Quickly scan recommendations throughout your Today feed.

Like spotify..

124

Translation

Article in English

Editorial Check

Google Speech API

Article in Asian Languages

• Basically we create articles in English and translate into Asian languages
• Machine translation from English to the other languages is the most accurate currently

Netflix-nise

125

Procurement

long-tail

In-house

head

​

​

​

30%

70%

Recommendation

/ Customization

Users

Expand the number of contents by external procurement

Old Economy Category

Digital Economy Category

Industry newspaper(162), industry magazine(700), local newspaper, news wholesaler, international paper

Procure

For busy business person

https://axion.zone/

https://www.youtube.com/watch?v=-c3LAppxAeI

https://axion.blog/entry/amazon-go-and-tech-economy/

Business person's main smartphone usage time

​

​

Research in the office

Learn in your free time

​

​

​

​

Listen to the podcast while walking

Text

Video

Podcast

Super-Packed Train

Workplace

Private

When you are moving

Appendix 1 : Target User Interview

127

128

​

Target Users Interview

They feel commonly...

• Anxiety about competition outside Japanese traditional company’s lifetime employment
• Anxiety about long-term decline of Japanese economy
• Anxiety on that they don’t have general skills
• Anxiety on overload of child's education expenses and mortgage

​

I Interviewed 6 business person working for Japanese traditional big companies with excellent educational background.

​

129

​

A, sales, major electronics, early 30's, married, no child

- Responsible for the business flow of enterprise computer equipment by electric company.

- The department has a large number of people from college baseball. Organization has similarity with college baseball team.

- low motivation for work

- The parent company is firing a lot of people. He never have accumulation of skills. It is uneasy for him to hunt job for the other company.

130

​

B, sales, major general insurance, early 30s, married, no child

- He realized that we won’t promote to executive class 2 years after he joined the company as a new graduate.

- Sell insurance products to owners of local small businesses, and he knows it contain very severe condition for his customer

- Marriage between co-workers

- am unwilling to work and pursuit romance even during working hours

- no accumulation of skills, but he believe that the company is stable and will protect him for lifetime

​

131

C, MR, major pharmaceutical, early 30s, married, 1 child

- Demand for MR in Japan will decline significantly as the firm focus on oversea market

- it is rumered that the firm is preparing large scale restructuring. He feel hostility towards foreign CEO from abroad, Japanese MRs are enhancing their unity.

- MR skill does not have general purpose

- Working for long time

- Anxiety about whether life-long employment will be maintained 10 years later after having a child and expecting another one.

​

132

D, MR, major pharmaceutical, early 30s, married, 2 children

- Demand for MR in Japan will decline significantly as the firm focus on oversea market

- The firm is firing MRs

- He want to move to the foreign competitor, so he is learning english

- Working for long time

- He wants to have a third child even though the balance sheet looks aggravated by mogage

133

E, trading company, married, late 20's, no child

- He thought he was an elite, but now he doubt it.

- The company is like sport team, not in knowledge-intensity industry.

- He want to change job, but the company’ salary is one of the best level in Japanese companies. He hesitate to take risk.

- Skills for internal process have improved very much, but does he have general skills ?

​

​

134

F, major bank, early 30s, married, two children

- Experiencing leader of club at high school and campus, he have worked for Japanese largest bank as executive candidate.

- But bank business is getting difficult after Fintech came out

- The bank layed off around 10,000 employee last year.

- Although the annual income exceeds 10 million yen, the game is not easy considering education expense for two children, mogage and potential income decline.

​

​

​

135

Japanese Companies Take all and don’t share return to employee

• Allocation of added value is biased towards companies. The labor share rate is at a historically low level in conjunction with the fact that wages have not increased much. This means that corporate profits are at historically high levels conversely.
• Long-term decline in “quality of employment”. Wages have not rise. Informal employment has been expanding for a long time.

Labor distribution rate dropped to historic levels: via 三菱UFJリサーチ＆コンサルティング

Appendix 2 : Economic Situation Surrounding Target Users

136

137

Crisis of Japanese business person

• rofession

​

• Skill gap. local optimization to the company, they have less general skills
• Lack of internationality. “Japanese Only” makes Japanese decline

​

138

Low Labor Productivity

• Low labor productivity. According to the OECD, labor productivity is lower than the other major developed countries.
• Japan's high quality service is supported by unfair labor of workers.
• The participation rate of women after marriage is low. Many highly educated women never work again after marriage.

​

Source: OECD