IEEE Technology and Society Magazine - September 2019 - 87

soon figured out that they could make money by farming cobras and killing more of them. The government,
on learning about this, scrapped the system but ended
up with a situation in which there were more cobras
after the intervention than before.
For our second example, consider the Paper Town
effect. In the 20th century, a famous map of New York
was created by cartographers Lindberg and Alpers, who
cleverly embedded a fake city Agloe into their map.
Agloe, NY, was not a real town. It was a paper town - a
booby trap to catch plagiarizers. A few years after Lindberg and Alpers set their map trap, the fake town
appeared on another map by the cartographer Rand
McNally, prompting the two mapmakers to sue for copyright infringement. Eventually it was discovered that a
real town called Agloe had in fact emerged in New York2
since users of the Linberg Alpers map thought that the
Agloe that once existed must have gone missing and
rebuilt it, and that Rand McNally may not have after all
have plagiarized Lindberg and Alpers.

Missing Data Problem
Missing data is a big problem for development statistics. It has been reported by Jerven that around half
of the 82 low-income countries have had one or partial poverty surveys within the past decade [21]. Kaiser Fung, in his book Numbersense [15], discourages
the assumption that we have everything and says: "N =
All is often an assumption rather than a fact about
the data." Due to missing data problem, any ranking
of such countries based on GDP only will be misleading because of the uneven use of methods and
access to data. Handling missing data is very important for data mining processes as missing observations can significantly affect the performance of the
model [32]. Therefore, analysts should handle the
missing patterns by employing appropriate methods
to cope with it and to avoid the "streetlight effect."
T he s t ree t l i g ht ef fec t i s t he t rend adopted by
researchers to study what is easy to study. The streetlight effect is a major issue that keeps Big Data findings from being realistically useful for human
development - especially when findings are yielded
using user-generated and easily available data.

Some Remedies
Interpretable AI and Big Data Analysis
With the wide adoption of deep learning and ensemble
methods, modern AI systems have become complex
and opaque and increasingly operate as black boxes.
Although these black-box models are producing
2

Understanding the internal workings
of modern AI systems can help
ensure fairness, bias identification,
robustness, and generalization of
Big-Data-based AI models.
Interpretable AI can help ensure algorithmic fairness, bias identification, robustness, and generalization
of Big-Data-based AI models. Interpretation of AI-based
decision-making algorithms is also necessary to ensure
smooth deployment of real-world intelligent systems.
But the development of interpretable AI requires that
the following questions be answered:
1) How to ensure accountability of a model?
2) How to ensure the transparency of the model output?
3) How to ensure the fairness of the model predictions?
Since BD4D is directly related to human development, the Big-Data-based AI model used in these cases
(i.e., for healthcare systems, judicial systems, etc.) must
ensure high accuracy and interpretability. A possible
remedy is to insist on using interpretable ML for highstakes BD4D decisions and to utilize explanation methods for justifying the decisions, where explanation
means the provision of visual or textual evidence of a
certain features related to an AI model's decision. As an
example of work in this space, Bach et al. [6] proposed
the layer-wise relevance propagation (LRP) method
which provides a visual contribution for each input feature in decision making. In another work, Ribeiro et al.
[40] proposed two explanation methods namely locally
interpretable model-agnostics (LIME) and submodular
pick locally interpretable model-agnostics (SP-LIME).
Interpretable AI for BD4D is still an open research avenue to explore for Big Data and the AI community.

Better Generalization
In Big Data research, researchers make inferences by
training models on a larger subset of data with the goal

https://en.wikipedia.org/wiki/Agloe, New York

SEPTEMBER 2019

outstanding results, it is very difficult to trust their predictions due to their opaque and hard-to-interpret
nature. This has been dubbed by some as the AI interpretability problem. We can define interpretability as
the ability to describe the internal processes of a system (i.e., complex AI and ML techniques) in such a way
that they are understandable to humans [17].

∕

IEEE TECHNOLOGY AND SOCIETY MAGAZINE

IEEE Technology and Society Magazine - September 2019

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - September 2019