Profile

Cover photo
David Lanz
Works at 年代網際事業股份有限公司
Attended National Cheng Kung University
Lives in San Francisco, CA
809 followers|670,994 views
AboutPostsCollectionsPhotosYouTube

Stream

David Lanz

Shared publicly  - 
 
想要嗎?辦不到,太貴啦!
 ·  Translate
近年在台灣越來越受到歡迎的掃地機器人,一直以來由美國的 iRobot 與
1
Add a comment...

David Lanz

Shared publicly  - 
1
Add a comment...

David Lanz

Shared publicly  - 
 
After being revealed as 'Android N' in March, and officially receiving the title of Nougat in June, Google's Android 7.0 is officially beginning to roll out to users following a lengthy beta period.
1
Add a comment...

David Lanz

Shared publicly  - 
 
Farewell, Chrome apps
New apps will be available only on Chromebooks by the end of this year, and will stop loading on non-Chrome OS machines in 2018.
2
Add a comment...

David Lanz

Shared publicly  - 
 
Google Cloud Natural Language API
Using the Natural Language API, you can take unstructured text and add structure to it then analyze the imputed structure and derive insights.
1
Add a comment...

David Lanz

Shared publicly  - 
 
 
The linux bash environment that just arrived on Windows 10 with the mega anniversary update, might just be the solution for a lot of problems you could run into while developing on a windows machine.

No searching around to get all the setup files you need for git with a cygwin or mingw terminal, node and npm, python, and whatever other tools you might need, and then struggling to make them all work in the Windows environment.

Just enable the shell, install everything you need with some simple apt-get and npm commands, and you are all set to go in a couple of minutes.


34 comments on original post
1
Add a comment...
In his circles
1,795 people
Have him in circles
809 people
Ming Xing's profile photo
Tim Wells's profile photo
Andy Lee's profile photo
King Orange's profile photo
Hisard Hsu's profile photo
吳易倡's profile photo
Quinn Hung's profile photo
maggic ju's profile photo
alex lu's profile photo

David Lanz

Shared publicly  - 
 
In the last few years we have seen a rise in programmatic ad buying space where marketers can bid for impressions in real-time through RTB (real-time bidding) protocol. This industry shift in ad-tech space has increased the demand for exploiting machine learning algorithms for optimizing ad buying & maximize conversion rates (e.g. sign ups, purchase. game installs etc).

At #LnData core competency is developing machine learning algorithms to optimally bid for inventories which maximize the conversion rate while minimizing the spend (i.e. CPA).

There are several challenges which an AI algorithm faces while tackling the ad buying problem. A few listed below:

The Dimensionality of feature space is high. There are a number of media features such as: exchange, publisher, app/website placement features, time of day, day of week, ad size etc. We also have several users features including: geo (city, state, country) and system-based features such as: device, connection, browser, OS etc. You can also grab extra user demographics from 3rd party data providers. The dimensionality of the feature-value space for advertising can be high.
The well-known Rarity problem - usually there are few clicks and much less conversions when running a branding campaign. To put this in perspective you might have a CTR around 0.08% to 0.11% in the best case. Now multiply this with 10% conversion rate which gives you a conversion rate in the range of 0.008%.

The ML algorithm needs to calculate and spits out conversion rate in real-time (~100ms) so that we can use that to compute the optimal bid value and incorporate the market competition rate for bid calculation. This becomes specially important if you are implementing your own customized bidder in-house.

The ML algorithm should support Online-learning feature by which it can keep learning as it bids/wins impressions. In other words, the ML algorithm has to continously analyze the history data to learn and adapt its bidding strategies according to the dynamic of the market and the human behaviour change over the period of the campaign.

The "Dimensionality" issue should be addressed otherwise there is no way for the ML algorithm to use the history data, train itself & extract useful knowledge from it. There are well-known ways to address this problem through using feature engineering techniques to reduce the dimensionality of the feature space. See the following article on the feature selection topic:

http://www.jmlr.org/papers/volume3/guyon03a/guyon03a.pdf

For addressing the "Rarity" issue, there are different techniques. The main thing is to pick the RIGHT Learning Algorithm for the problem in hand. For instance, one can prove (theoretically & empirically) that Discriminative Algorithms such as Logistic Regression are better choices over generative algorithms such as Naive Bays Classifiers. Read the following article for more details on the issue:

http://ai.stanford.edu/~ang/papers/nips01-discriminativegenerative.pdf

Gary Weiss has written a comprehensive survey on the Rarity issues. Some of the proposed technique in Weiss’s paper can be applied to designing learning ML algorithms for optimal ad buying to maximize for rare events such as clicks/conversions. See the following article for more details:

http://archive.kdd.org/sites/default/files/issues/6-1-2004-06/weiss.pdf

Microsoft research lab has also developed a scalable algorithm for Click Prediction for Bing a few years ago where they needed to rank the ads based on their click probabilities to increase their advertising revenue. They have used a Bayesian Learning model which serves ads and collects users responses to the ads (click/no-click) over time and feeds the data back to the learning algorithm to optimize the weights for used features. See below for the Microsoft paper:

http://research.microsoft.com/pubs/122779/AdPredictor%20ICML%202010%20-%20final.pdf

There is also a recent work by University of College London in RTB space where they trained and compared the performance of Logistic Regression with Gradient Boosting Regression Tree. See the full paper below:

http://arxiv.org/pdf/1407.7073v3.pdf

Here, at #LnData we design machine learning algorithms for customized RTB bidders for optimal traffic buying & maximizing our clients campaigns conversions. Most of the campaigns we run are performance marketing where we need to spend the ad budget to achieve high LTV costumers (e.g. mobile games) or generate qualified leads for tech companies with B2B SaaS products.

In the last few years, we have also implemented machine learning algorithms for a few performance AdTech companies. Some of these algorithms are based on the area of Multi-armed bandit where we have designed customized algorithms for each AdTech client.

In a Multi-Armed Bandit algorithm, the machine switches between two states: exploration & exploitation. During the exploration phase, the ML algorithm explores the ad/user/publisher space and collects performance data around ads/users/placements etc. As we collect performance data, the algorithm starts exploiting the obtained knowledge on what ads/offers performed best on what publishers (i.e. apps/websites). This class of algorithms performs well for some of our previous AdTech clients where we customized an online learning algorithm to match the specific client business needs.

You can read the "Using Machine Learning Algorithm for Serving Ads" article from #LnData website.

This is a short summary of the state of art for the programmatic ad buying but it should give the interested reader a high level overview of the space.
1
Add a comment...

David Lanz

Shared publicly  - 
 
Andrew Ng 教的 Stanford CS229 Machine Learning
[8/1/16] Autumn Quarter 2016-17 students: This website will be updated in mid September. [6/12/16] Project reports and posters have been posted here. [6/9/16] Letter grades have been posted on Axess. [6/3/16] The Friday final exam has been posted under materials. [6/2/16] If you are taking the ...
1
Add a comment...

David Lanz

Shared publicly  - 
 
越來越貴不等於越來越有效呀!! 而且封閉不給第三方監測,結案只有一張Excel, 還不如用LnData更能讓廣告組得到投放最佳效益 #lndata #3tdparty #bigdata #igrp
 ·  Translate
(圖片來自:Facebook 官方網站)作者呂元鐘,台大資工系肄業,創辦經營國內外電子商務已十餘年並具成功出場經驗。 擅長電子商務、網路行銷、創新營運模式,曾任數位時代網站電子商務專欄作家,現任痞客邦PIXNET社群商務部副總經理,負責新商務開發,商務合作。原文出自呂元鐘 Facebook 貼文,INSIDE 獲授權刊登。好友張瑋容(Wei-Rong Chang)提到,電商老闆覺得 FB 廣告越來越貴,紛紛尋求其他出路。「最近,跟很多電商老闆聊天,幾乎每個老闆都有共同需求:就是脫離廣告費越來越貴的 FB,另尋出口...。」有關 Facebook 廣告越來越貴,這件事情是這樣的。電商網站基本上是
2
Add a comment...

David Lanz

Shared publicly  - 
 
Google Duo上架囉
 ·  Translate
Duo 是適用於所有人的一對一視訊通話應用程式,設計簡單、實用又有趣,讓您絕不錯過任何重要時刻。 功能如下: 簡單易用的介面 只要輕觸一下就能撥打電話...
1
Add a comment...

David Lanz

Shared publicly  - 
As per a new Google project over on Github, the company is working on a brand new operating system. The new OS, currently dubbed Fuchsia, would add to Google’s current operating system offeri…
1
Add a comment...

David Lanz

Shared publicly  - 
1
Add a comment...
David's Collections
People
In his circles
1,795 people
Have him in circles
809 people
Ming Xing's profile photo
Tim Wells's profile photo
Andy Lee's profile photo
King Orange's profile photo
Hisard Hsu's profile photo
吳易倡's profile photo
Quinn Hung's profile photo
maggic ju's profile photo
alex lu's profile photo
Work
Occupation
專案經理
Skills
Programming, XCode, Android, Mobile Graphic Library Dev., Google Glass, Mirror API, Chromecast API, JAVA, PHP, Python, Perl, HTML5, Chrome APP
Employment
  • 年代網際事業股份有限公司
    專案經理, 2006 - present
  • 蕃薯藤數位科技股份有限公司
    專案經理, 2004 - 2006
  • 威盛電子
    資深韌體工程師, 1998 - 2004
Places
Map of the places this user has livedMap of the places this user has livedMap of the places this user has lived
Currently
San Francisco, CA
Story
Tagline
對我,鋼琴演奏與程式設計是相同的原理,在執行或演奏的瞬間,達到完美的script~
Introduction
對我,鋼琴演奏與程式設計是相同的原理,在執行或演奏的瞬間,達到完美的script~

Just as a computer follows a script to run a program, so do we direct our lives according to some script written by others for us.
Collections David is following
View all
Education
  • National Cheng Kung University
    CSIE
  • Tamkang University
    CSIE
Basic Information
Gender
Male
Other names
David Lanz