從大數據看社會影響力 - AVPN會議分享Big Data for Social Impacts - the insights aggregated from 2018 AVPN

English | 正體中文

⚠️ 警示：文章內容僅為筆者個人觀點，歡迎交流，但請勿有惡意攻擊的言論與行為.

從大學畢業後到工作，一直有在想該如何做一些對永續世界有貢獻的事情，而不是每月定期捐款、捐年菜...總覺得幫助有限，甚至不確定這些錢投入是不是真的有預期的效果...常常在思考一個切入點，如何能善用自己的工作經驗/技能，能夠帶來社會本質上的改變⁉️

因緣際會，在我思考的過程遇到了活水社投 (B Current Impact Investment)的總經理Ray，很大方的介紹我AVPN (Asian Venture Philanthropy Network)這個會議、還願意贊助我一張票讓我去新加坡參加，覺得備受命運之神眷顧 😚～ 也讓我能先到新加坡考察一下，為之後的學生生活做準備‼️

今年AVPN會議為期4天，每天都是sessions滿滿，加上身為在社企投資/影響力投資的新進者，這些議題內容實在沒辦法聽聽就消化完畢...😓 所以會議結束後，想了又想，決定來當個部落客，將這些收穫文字化，除了幫助自己內化外、也希望能和大家分享，讓這些知識散布到台灣


首先，從第一個會議主題開始：⟪Big Data for Social Impacts⟫ 大數據看社會影響力
這算是我看到會議主題就很感興趣的一個，畢竟之前工作就是和資料分析相關，並且是從資料分析結果 來看客戶對服務/產品體驗，也是服務/產品影響力衡量的一環；這個主題，也許能幫助思考，如何將我在資料分析這塊的經驗、應用在社會影響力的領域🙂

這個會議我想可以分成兩部份分享：實際案例＆應用上挑戰
實際案例分享3個我印象特別深刻的，而這3個我自己認為 各自和大數據分析過程中，數據搜集、數據遷移、與數據挖掘有關，以下就一個個說明，希望能簡單明瞭，到每個人都能了解他背後的運作方式...

Part I: 實際案例

 ➊ 中和農信在中國鄉村的應用 (數據搜集的應用):

                (picture downloaded from google)

中和農信 (CFPA Microfinance) 屬於中國扶貧基金會 (China Foundation for Poverty Alleviation, CFPA) 旗下專營微型信貸的子公司，簡單說，是中國政府支持的NGO。主要提供微型信貸給鄉村農民，因為這些鄉村農民一般來說識字率低、收入低、沒有抵押品，無法通過銀行信貸審核。中和農信透過這種做法，期望幫助農村脫離貧困，掌握經濟自主，進而讓整個經濟體能成長。

但其實這些做法和當今很夯的那些科技(行動支付、AL、ML...)並沒太大的關聯，反倒是很需要人力執行；因為鄉村農民多半是用現金、沒有複雜的金融交易，因此沒有信評歷史紀錄(簡單說，不像我們常常刷卡，透過網路就可以到聯合徵信中心上面，下載自己的信用報告，裡面就會有刷卡與還款紀錄～要是有貸款也會有紀錄喔！有興趣的人可以去下載看看😏) ，也就是說，中和農信需要設立分支駐點在不同區域，由駐點人員深入當地瞭解農民的經濟狀況，評估還款能力、或再發放信貸後，追蹤還款進度

這些做法開始科技化，是從2016年底 螞蟻金服 (Ant Financial)注資開始，而現在螞蟻金服也是中和農信的第二大股東。這一項注資行動，標示中和農信將進行轉型，透過螞蟻金服身為金融科技公司巨擘，導入其在科技基礎建設的經驗與技術，例如應用大數據分析螞蟻金服的信貸平台，預期將可使中和農信提升在農民信用評估的效率、以及結合螞蟻金服理財服務，可望將理財帶入農村 (ex: 增加農民投保率)，提升農民金錢控管能力

而回到一開始提到，我認為這個case算是大數據分析中 數據搜集 (data collecting) 的應用，因為就這整個case來看，會和大數據分析有關主要是因為螞蟻金服的注資，使得螞蟻平台上大數據資料得以導入應用，但要與螞蟻數據串接，關鍵還是在農民載具使用➡︎簡單說，就是手機 (或者說行動支付 可能比較恰當)。一但行動支付能在鄉村普及，中和農信可以透過手機搜集農民貸款資訊 (貸款額度、還款記錄...)，和螞蟻金服數據一起進行分析，不僅在數據搜集上比現在有效率好幾倍，在貸款信用評估過程也更精確、更能降低呆帳風險

但要做到這點想必要一番時間精力...因為就目前中國鄉村現況，多數農民對數位服務是陌生的。整體農民中僅有不到一半(記得會議好像是說40%)有手機，但有手機的這些人，活躍用戶 (active users) 也不多...上網找到2016/12月中國官方 網民比例數據，可以看出城鄉差距，而且2016較2015 城鄉差距是有些許擴大的狀況

                         (picture downloaded from CNNIC Report)

這個問題想必不易解決，因為涉及鄉村數位素養 (Digital Literacy)的普及，也就是農民能了解並使用網路資源的能力。因此，儘管有農民擁有手機，卻不是活躍用戶，因為他們不了解網路的強大、進而不會想理解如何操作。因此，在這個case中，我想大數據要進一步應用，光是在數據搜集階段，就需要投入資源，想辦法提升鄉村的數位素養

☛ 更多關於中國微型信貸發展，可參考亞洲發展銀行報告(請點我)
☛ 更多關於中國互聯網發展，可參考第39次《中国互联网络发展状况统计报告》(請點我)


➋ 肯亞Jaza Duka 微型信貸專案 (數據遷移的應用):

    (picture downloaded from google)

這個case我認為目的其實與上一個中和農信大同小異，都是透過提供微型信貸給社會弱勢，以期幫助他們經濟獨立，進而對社會經濟有貢獻。只是上個case的目標族群是中國鄉村農民，這個是針對肯亞當地商店自營者 (類似台灣早期的雜貨店)，他們也是因為收入不高、沒有抵押品擔保，以致無法向銀行申請貸款；這些商店自營者，常常會有現金短缺問題 (ex:繳小孩學費)，以致於沒辦法有多餘資金向製造商購買更多的商品來販售...🙁 為了幫助這些商店自營者，發揮他們應有的銷售能力，萬事達卡 (Mastercard) 和聯合利華 (Unilever) 合作，試做這個專案，當地語言叫Jaza Duka，意思就是要填滿你的雜貨店～希望商店自營者們能有能力購買貨品、填滿自己的雜貨店😎

這個專案的兩大支柱是萬事達卡和聯合利華，萬事達卡是國際發卡組織，擁有海量刷卡資料，能夠進行信用評等分析與模型設計；聯合利華則是國際消費品公司、販賣貨品到世界各地，包括肯亞，因此擁有與肯亞這些商店自營商的進貨紀錄。這兩間公司合作，透過聯合利華的進貨資料，結合萬事達卡的信評分析，算出每位商店自營商的借貸能力、與建議的微型信貸額度，並將資料提供肯亞商業銀行 (Kenya Commercial Bank, KCB)，由KCB依此核發信貸

☛ 點我看當地商店自營商的訪談

在這個專案中，我會聚焦在大數據分析中 數據遷移的應用，主因為進貨資料和信評資料落在兩間公司各自資料庫，而這個案子聽起來是將聯合利華這邊的進貨資料拋到萬事達卡資料庫，進行資料串接，但從拋到另一個資料庫開始就會有資料傳輸安全性/隱私問題、拋過去後要怎麼串接也是問題 (i.e. 取哪個時間區間、和哪些變數一起分析...)，而這些都關係到最後產出的借貸能力指數。簡單說，如果在一開始拋入資料本身就有偏差(ex: 資料都是常進貨的商家，但常進貨的在當地並不多，故資料不足以代表全體)，那後端的分析結果是否可用就是問題...不過目前這個案子，網路上還沒找到後端方法論的討論，可能也是因爲還在試做階段吧...只能之後再觀察觀察了～

講到這個案子，我也想特別點出我認為和上一個case最大的不同點：數位化程度的差異。
相較於前面提到，中國鄉村網路使用仍不普及，肯亞當地的數位化程度卻相對高很多‼️主要體現在行動支付的普及➡︎ M-PESA是當地常用的行動支付工具 (也是fintech常見的案例之一～可見非常成功)，這是由電信商Safaricom和KCB共同發展的，當地人儘管只有低階手機，也能透過M-PESA進行查帳、匯款、繳貸款...並且從事這些金融行為時，也能累積點數使用電信服務，因此在當地已經成為人們生活的一部分

因此，以M-PESA在當地如此成功，也打下了當地資訊搜集的基礎。以目前來看，KCB在Jaza Duka中似乎僅是被動資訊接收角色，但若之後KCB能將M-PESA數據結合，感覺很有可看性🤔 (至少已經得到信貸的商店自營商會用M-PESA還款，這樣就有還款記錄可以作為信評參考)～不過，當然又回到資料要怎麼串接這個問題了...

☛ 想知道更多關於M-PESA怎麼用 請看這裡


➌ 菲律賓在天災預防/管理的方式 (數據挖掘的應用)

這個case是由菲律賓當地的發展基金會Gawad Kalinga (是”提供照護“的意思)和微軟科技合作的實例，主要防範颱風的災情。看這個案子特別有感，想到台灣也是一個天災多多的地方，而且也身受颱風侵略之苦...想到之前2009 莫XX颱風，我們全村都淹水，放眼望去一片汪洋...😰(但小林村更慘被滅村..) 2016 梅X颱風，風大到我家窗戶玻璃被震了一地，整張床上都是玻璃...😭 但在菲律賓，是整個村落都被輾過...地理位置位在熱帶環海、加上當地建設不比台灣先進，災情似乎比這裡嚴重好幾倍...

於是為了預防嚴重災情，Gawad Kalinga與微軟合作，應用Microsoft Azure雲端計算平台，期望透過雲端計算，預測颱風路徑與規模，進而做到事前糧食配發、人員避難等措施，將傷害降到最低。但要做到這一點，需要當地村落的詳細資訊，包括村民戶數人數、可避難場所...，因此Gawad Kalinga志工深入當地社區進行田野考察，使用微軟APP進行資料登入，登入後資訊會上傳雲端，導入Microsoft Azure

這個case中，也有數據搜集的過程，但我會更著重他在數據挖掘的功夫上，因為就這個案子本身，重點是如何有效預測颱風路徑與規模、以及預測後快速有效的決策行動。颱風預測最困難點，在於每個颱風都不一樣，因此才會需要大數據分析；就這個case，是透過Microsoft Azure，因為累積夠多數據、透過機器學習 (Machine Learning) 技術，平台可以透過過去上千萬筆颱風路徑，找出相關性並作預測；我認為，在機器學習之前，數據挖掘決定機器學習的成果。數據挖掘是從海量資料中，挖出有相關的那個部分進行分析，簡言之，要是挖到不相關的、或沒挖到相關的，都會導致之後模型預測/演算法有偏誤 (ex: a+b+c=y, 但用a+b=y來解釋)。所以，儘管這個case本身較focus在執行面上，我會對他背後的資料處理邏輯更有興趣；另外，我想在做災害因應行動上，為了讓颱風對當地影響更貼近真實 (即讓模型更準確)，當地資訊 (村落家戶分佈、家庭財產狀況...)是關鍵影響因子，故需要Gawad Kalinga深入當

最後，題外話提一下Microsoft Azure，目前微軟有在網路學習平台EdX上開相關免費課程，叫Introduction to Artificial Intelligence，裡面有介紹這個平台 (似乎含有實際操作，但我也還沒上到那邊..進度緩慢哈)，以及機器學習的一些背後方法，有興趣的可以去上上看囉！

☛ 點我瞭解更多Microsoft Azure課程資訊


Part II: 應用上的挑戰
其實關於挑戰，是源自於會議中發問的問題：”隨著AI的發展，人類是否被機器人取代？“ 答案當然是NO啦～🤣 原因很簡單，而且其實前面也有多少提到➡︎如何確保搜集的資料是可用的？如何確保演算法沒有偏誤？當然都要透過人類囉‼️ 機器當然可以很快學習並應用，但若他沒有一個好的老師、好的教材，學到的東西也只是垃圾😅

就用個實例來向大家解釋，這是一個微軟和NGO寶貝回家 (Baby Come Home) 合作的案子，透過人臉辨識科技來協尋失蹤兒童

                 ⬇︎ 請看以下影片，瞭解這個案子進行方式 ⬇︎

這項人臉辨識技術在不同區域多次測試後，微軟發現測試結果存有偏差：機器對女孩臉孔的辨識能力不如男孩，並且機器辨識能力在不同區域的結果存在落差(ex: 美洲vs非洲)，因此修正背後演算法勢在必行，但機器自己不會知道有這個問題存在，而是人類必須透過多次檢驗、發現問題、修正問題、重新教機器學習。

因此，在這個會議的結論，是在未來機器不可能取代人類，相反的，更多的資料解讀者 (data translator) 是需要的。什麼是資料解讀者？簡單說，就是一個領域的專家、知道搜集哪些資料來用，舉例來說：在菲律賓防颱的案例中，Gawad Kalinga就是這樣角色，因為Gawad Kalinga瞭解當地社區的狀況，也有颱風防範的經驗，因此能知道若要有效防颱，哪些當地資料需要搜集，才能快速避難並進行物資分配


References:
➤ CFPA Microfinance Received Strategic Investment from Three Investors to Help Accelerate Technology Adoption to Enhance Efficiency of Financial Inclusion and Fight Against Poverty
➤ Mastercard & Unilever launch Jaza Duka, inventory-backed loans to empower micro entrepreneurs in Africa
➤ Mastercard and Unilever Break Down Barriers to Growth for Micro Entrepreneurs with First-of-its-Kind Digital Lending Platform
➤ Mapping hope: how community data is driving cloud-based disaster management in the Philippines

 ⚠️ Alert: All articles in the blog are the personal perspectives of the blogger, interactions are gratefully welcome, but please do not voice any hatred words or behave in harassing way.

Formerly in a retail bank as a marketing researcher, who dealt with data every weekday to extract business insights from strings of customer voices, I'm with certain to recognize the impacts derived from data analysis. In fact, I was always thinking the possibility to apply this data analysis knowledge to deliver more impacts to my society, or even the whole world, instead of being fixed to the daily job only to find out the ways to improve the service quality and customer relations for the company. Thus, when finding the AVPN session combining Big Data with social impacts, I was too excited to miss it ☺️.

The session was proceeded with experts/managers from Microsoft, Mastercard, CFPA Microfinance, and Yishan. On the whole, I think I could have the session divided into 2 parts: real practices and challenges.

Part I: real practices
In this part, there were 3 cases impressed me most, and each of them coincidentally took a role in Big Data analysis process: data collecting, data migration, data mining

➊ Microfinance to the farmers of the rural areas in China (data collecting):

                (picture downloaded from google)

this is the project of CFPA Microfinance (a government sponsored NGO), to lend loans to the people in rural areas, who are low-income, without collateral, and normally have difficulties in getting loan permission from banks. The initial actions taken were in fact demanding and had little to do with technology- the staff needed to do much field work to contact the borrowers in person or by phone, to do credit management and also track their repay capacity.

Things began changing and becoming more technological when Ant Financial instilled its investment and became the 2nd largest shareholder of CFPA Microfinance. As a big Internet finance supplier, Ant Financial is surely to share its infrastructure with CFPA Microfinance, making CFPA Microfinance more efficient in credit management by utilizing its database of lending platform (with 622 million users and 600 billion yuans into consumer loans), and more diverse in service offered by combining its strength in not merely lending but wealth management as well (e.g. to take rural farmers into insurance coverage).

But why's it about data collecting? To make it simple, it's about devices, or more specifically, mobiles. Once mobiles could prevail among the rural farmers, the transaction records of the farmers would be collected and written into database through the financial app on mobiles, and based on the database while mapping with that of Ant Financial (which seems to be another application of Ali Cloud...but not the topic today), CFPA Microfinance could decide the credit and loan amount more efficiently.

The application sounds wonderful, but the fact isn't always running in the ideal path. For the model to be successfully fulfilled, digital literacy among the rural farmers is the key driver behind the scene. And the truth is, among the farmers, only some had mobile banking account, among whom, only a few are active users. In this regard, the further use of Big Data is only germinating, and demands more efforts on infrastructure for data collecting.

☛ For more info, you can refer to ADB report about microfinance innovations in China


➋ Jaza Duka, a lending platform in Kenya (data migration):

    (picture downloaded from google)

The purpose and vision of this project is quite similar to that of CFPA Microfinance. This project is also with aim of microcredit disbursement to Kenya entrepreneurs, who are normally low-income and without collateral to acquire loans from banks. The project mainly to solve the problem that such entrepreneurs usually with cash constraints to expand their business by buying and stocking more inventory, and thus are insufficient with money to better of their living standard. Therefore, the project is called Jaza Duka, which means in local "Fill up your stores", attached with wish to grow the local entrepreneurs ability to finance themselves to buy inventory.

The project was initiated by Mastercard and Unilever. As the locals buy inventory from Unilever, Unilever then keeps the distribution data and cooperates with Mastercard to do credit analysis (supposedly Mastercard as a leading payment institution has its credit modeling skills for this...), and later the two come out with microcredit eligibility recommendations to Kenya Commercial Bank (KCB). In this way, the locals could gain more possibility to acquire credit from banks, since in the past, they could hardly have formal credit history to request for bank loans.

☛ There's an interview video of the locals to let you better catch the situation there

Because in this case, it seems the key motivators are Unilever and Mastercard, the critical action in the project is how to generate the microcredit eligibility recommendations from the distribution data. And since the data acquired is separated between the 2 companies, how to transfer one to the other, in what format, are key issues to be considered. As someone dealing with data, we all understand that when combining different sorts of data, how to do data structuring and cleaning before migration, and how to combine and read the data against potential bias, are critical to make the result useful and insightful. However, there seems no articles or reports about how project is handling with the data...so could only keep the topic in mind for future check out.

Another thing I'd like to specially mention, is the digitalization in Kenya. Although as earlier said, the project goal/target is similar to that of CFPA Microfinance, the big difference is the digitalization of people behavior. While in rural China people are still stuck by low digital literacy, in Kenya people are very familiar with mobile payment using M-PESA as a tool. That is, in Kenya, M-PESA has become a part of life for locals. It's a payment tool launched by Safaricom and KCB, and it not only offers financial service (e.g. remit money, pay loan...), but also telecom service (and more financial service used, more bonus gained for telecom service), which means, people only with a low-end phone can do everything they need for both finance and telecom.

Based on this circumstance, I think the project Jaza Duka has potential to enlarge its capacity. As now the role of KCB seems to be merely loan lender, to make decision depending on the external data, KCB may have ability to involve more once it knows how to utilize transaction data through M-PESA. But of course, the critical issue is still how to combine this data with others to make it useful, so, let's wait and see...

☛ You can find more about how M-PESA is working here!


➌ Disaster management in Philippines (data mining)

In this case, Philippines-based development foundation Gawad Kalinga cooperates with Microsoft, to react rapidly to natural disasters. This case actually is the one among all making me resonate most, because my country Taiwan is also an island suffering from natural disasters like typhoons and earthquakes.

In the case to ease the large catastrophe from typhoons in Philippines, Gawad Kalinga tried to make good use of cloud computing to predict the path of incoming typhoons, settle resources among the community, and mobilize residents to shelters beforehand.  To do this, by using Microsoft data entry app, volunteers of Gawad Kalinga step into the community to collect data of households (e.g. household composition, income, livelihood) and to tag important locations (e.g. government offices, schools, evacuation centers). Later with the data uploaded to the cloud, Gawad Kalinga could utilize Microsoft Azure, a could computing platform, to do calculation and predict typhoons path and scale based on Big Data and Machine Learning algorithm.

Although the case also involves data collecting process, I value more its data mining part, because I view it as the fundamental of the whole algorithm (i.e. the model used to predict typhoon path and scale), based on which, the disaster management can work successfully. And before go deeper, I think it's necessary to clarify the difference between data mining and machine learning, which to some extent look similar. After google for some information about the 2 concepts, I suppose data mining goes before machine learning. While data mining is about mining among data to extract the useful one, machine learning is based on large data to find the correlations/patterns and apply such correlations/patterns to predict. That is, if during data mining process we can't select the data right (i.e. data with relevance and quality), we probably can't go smooth in machine learning process because the data may have low correlation or no specific pattern to follow.

Therefore, in this case, I think the critical process is data mining, and besides data already stored in Microsoft Azure, the information from local is also key driver to manage disaster effects, so that's why the data collecting here also matters. However, as users of Microsoft Azure, we'll focus more on the results generated from data to make decisions, instead of the complex equation and algorithm for data analysis. Despite the fact, I still want to take this chance to introduce the courses launched by Microsoft on EdX called Introduction to Artificial Intelligence, which is free and you could gain more knowledge about machine learning and the statistical methods used behind.

☛ To know more about the Microsoft course


Part II: challenges
This part is actually derived from the question that "Will most of us be replaced by robots?", considering AI, ML, Big Data are thriving. And as you probably guessed, the answer is NO.

The reason is quite simple: it's human to pick up the data for analysis and set up algorithm for machine to learn. Just people are too fascinated with what AL/ML/Big Data are going on, to identify the essential role to qualify the data used. Here came a real example in the session. In the case, Microsoft cooperates with NGO-Baby Come Home to apply facial recognition technology to find the missing children.

⬇︎ You can find more about how it works via the video below ⬇︎

And after testing for many times in different regions, the Microsoft team found there's bias of the results. The machine is not smart enough to recognize girls than boys, and its facial recognition ability differs from regions. So obviously, someone must correct the learning algorithm and reteach the machine. Namely, those technologies are not taking over the role of human, but instead, making our role more important than ever.

I believe in the future, more data translators, who have proficiency in domain expertise to bridge the gap between data and the problems met, will be needed to facilitate synergy between high-techs and social participants.



References:
➤ CFPA Microfinance Received Strategic Investment from Three Investors to Help Accelerate Technology Adoption to Enhance Efficiency of Financial Inclusion and Fight Against Poverty
➤ Mastercard & Unilever launch Jaza Duka, inventory-backed loans to empower micro entrepreneurs in Africa
➤ Mastercard and Unilever Break Down Barriers to Growth for Micro Entrepreneurs with First-of-its-Kind Digital Lending Platform
➤ Mapping hope: how community data is driving cloud-based disaster management in the Philippines