Deconstructing Black-Boxes: Notes from the Field

This is the fourth and final research note from Anand Kumar Jha, one of the short-term social media research fellows at The Sarai Programme.


This post contains a detailed summary of the secondary research component and an early stage analysis of the three interviews that were completed as a part of the primary research component of this project. Two of these Interviews were conducted with data scientists working with large scale public data. The third Interview was with an Image processing expert now active in the field of data sciences.

Derivatives of the Secondary Research

Work done until now was cursorily captured in the two research notes published earlier. Substantiating the conceptual metaphoric convergence of the Camera and the Interface were key conclusions such as:

1. Building upon the ‘medium as a vector argument’ put forward during the Gun- Camera narrative, the Camera and the Interface behave the same.

2. The stakeholders [1] are both physically and conceptually situated at the two ends of the medium. They assume media is the vector and then they play the roles aligned to the power structures associated with their position. What is shared in the shot/tweet is a private moment. However the backend – which is the invisible face of artificial learning, algorithmically cognizant data eating machine – generates the private moment. It generates the practice of using the media. It creates the idea of the image, the need to capture the image and the need to pose for it. It, similarly constructs the idea of systematic utility of the behavioral information and the ritual of producing it [2] and consuming it [3]. The participants in the practice assume that they are in a setup where one of them or both of them have the agency over the media whereas the media (and largely the backend) is the one with the agency.

3. What makes the participant trust and hence engage with the Media? It is the Interface. The Interface which was constructed with the forces of user research (a combination of usability studies and ethnographic studies) and Visual design (aesthetic seduction informed by the rigour of color and form theories and validated by cognitive sciences) and now, supplemented by data sciences(with their statistically and programmatically defined user cohorts being used to mass customize the interface). This attribute is the ‘Mirror’ which depicts familiarity, telling participants that they are with themselves or with someone like them. In case of the Camera, these are the well curated mass produced images Norming (shaping the environment) and forming (shaping the dream environment) participant’s world in the same time, drawn from amongst social cohorts same or similar to the user’s on a consumer scale.

4. Another attribute that the Interface has is its Slave/Agency-less behavior which convinces the participants of its invisibility and hence the innocence in the entire act. When the shoot/tweet happens the media becomes invisible because it is capable of shifting focus away from itself and onto a subject.

5. Intermittently and interchangeably referred to as media, the Interface and the Camera are membranes, the spaces for constant contestation If we reduce it to being a spectator then it is witnessing the power struggles at the two end, the supposed vector, which essentially was its doing.

6. This point was largely derived out of my reading of ‘The Networked Society’ by Manuel Castells [4]. The Media, (I called it the Affordance in the last post) gets this position from the time and space it filters, coupled by the scale and automation axes it rides on. The co-existence of various types of affordances with various types of class (the sickle with the farmer, the lathe with the factory worker and the Futures and derivative trading interface with the banker) defines the historic trajectory the Media has taken, which reveals its politics.

The last point was not so much substantiation as an open thread that needs to be closed by another body of work. This summarized the ‘What’ of the enquiry, the first part of my research question [5]. The ‘How’ part of the research, which I defined as the Image and Interface Black-box, both of which function as Big Data Black-boxes in the Social Media, are largely detailed out through the Primary Research which have been a body of three interviews, two with data scientists and the third with an image processing expert turned data scientist. Highlights from this are made available below. A detailed analysis will follow in the study report.

Highlights from the Primary Research

The Data Science duo I spoke to, co-found a company called DataWeave. As their website reads, this company “provides actionable data by aggregating, parsing, organizing and visualizing millions of data points from the Web”. The enterprise largely operates in fashion trend prediction and catalogue management space and will soon expand into other spaces.

1. About the Trajectory of Big Data and Social Media

A big part ofthe application space in computing has been dedicated to management of various types of data-bases, be those medical record systems, electronic logs from an access card reader or a photo album. The Archive was considered to be a passive, legally required beast of burden which companies were willing to outsource. Data warehousing [6] with third party warehouses and access of data using Mainframe computers were the norm till late 90’s. The data archives were largely owned by a specific organization and so was the structured metadata around specific processes or activities. This data was centrally stored in warehouses and had multiple layers of protocols to access it thus defining ‘Who’ can access it and ‘How’ it can be accessed.

With the advent of Distributed Computing [7] and processing capacities of Client Side Machines [8] ramping up manifolds, efficient ways of creating- distributed storage, retrieval, processing and analysis of data- was arrived at. This created the ground for the arrival of social networks which were capable of generating massive amount of unstructured data that could be analyzed for consumer insights.

Google Trend - Big Data
A Google Trend showing the rise of searches on Big Data against Dip on searches for Data Warehousing.
Credit: Accessed by the Author on Google Trends.

The graph above shows this switch in the mindset of data being a source of liability to being a profit fetcher as more people looking for Data Warehousing as a solution dips while Big Data as a sought solution-space picksup.

One of the DataWeave founders, Dr. Sanket Patil explaines this below:

Data storage and processing has historically been a challenge. File systems, pay rolls, administration frameworks churn out a lot of data. Data storage was thought to be a solved problem with Oracles and Microsoftsof the world. However the rate of production of data is significantly higher now. So storing it becomes the primary problem.Social media applications such as twitter produce huge amount of data in a very small timeframe, so does industry with sensors and trackers or pharma with truckloads of documentation. Managing all this is a huge problem. Parallel to this, technology is also evolving. From Huge servers in Ac server rooms we have graduated to large number of small systems under the paradigm of Distributed Computing. Since the processing and storage challenges were growing at a much higher rate than processor speed and hard disk capacities, technology and platform had to get better so data is now stored and processed in a distributed manner to serve it to applications.

Another co-founder, Dr. Mandar Mutallik Desai supplements:

You see in older days there were accounts people and now data is being produced by people and machines who are technologically enabled to do so. Due to this, the data has scaled enormously. This has lead to problems in consumption and sanitization of the data. Earlier the scale was lesser and the technological assistance was also lesser. (Pause) See, data is always big for what it can tell us, its worth lies in what it answers.

2. About the Big Data Framework (The Interface Black-box)

The Interface as discussed collects the user data actively(through forms/text boxes, affordances) and passively (tags tracing user footprint on the site). What user chooses to spend time on decides what user gets to see next time they log in. Data is far from being structured; it contains a lot of noise. It gets generated in high volumes and does not leant itself easily to analysis. Big data is coming together of a lot of technological events to produce consistently fast data points for decision-making. Dr. Patil breaks down the black-box as a series of activities:

Ok this I will explain in a sequence of steps. Data Aggregation i.e. how I get the data is the first step. Next one is cleaning the data, removing noise to make public data ready to get insights from (in our case). The next stage is storing data which should be according to how that data should be processed (the multi-dimensionality of such data). This is followed by representation, which is serving the data. A question at this stage is not just how one takes queries/usage patterns, but also how does one scale to accommodate a huge level of traffic of queries. Each stage has a different kind of technology. The first stage, which is Aggregation, has Crawlers [9] and Scrapers [10], which in our business help us aggregate the data. Data clean-up is done by machine learning algorithms [11] which are trained over domain specific knowledge base. Storage and processing is done via the Hadoop [12] and MapReduce [13] frameworks. Serving is done through ElephantDB [14] and Memcashed [15] (light and fast). For reporting and analytics Tableu [16] is used. For visualization we show a lot of aesthetic infographics via libraries that are Javascript and Python based. D3.js [17] is one of them . Google and IBM have similar libraries. For dashboard we use a combination of HTML and JS. twitter library called Bootstrap is also quite well known. Since we are largely in retail which has traditionally not been a technology space we cannot just dump data on them, it has to be shown as insights that are classifiable into actionables by showing visualizations and reports.

He also gives a historical perspective on the development of technologies related to this space.

Google was the pioneer in this field when they introduced this framework called Mapreduce with google file system which is very similar to HDFS which is open source. Yahoo followed it up with its open source framework Hadoop. You also have to understand that in case of big data analytics, Data is not structured. By this i mean that the dataset is not fixed and more data you have more possible number of questions it can answer. When one wishes to log this data, they key variable they are looking at what questions do they want this data to answer, which will dictate what would be the best way to save this data (the data structure) and which, in turn will dictate what kind of infrastructure is required. An example which explains how the question dictates the approach is say searching info about an individual in a social network is a point query, while looking up a friend’s friend is a network query and will make use of network graphs (which is a model of data network). A query like this is very difficult in a traditional network and needs new models of data network. We apply big data framework largely to retail problems like fashion analysis, scenario planning, trending colors and palettes and Runway analysis. A lot of people are also using it in real estate tracking. Largely businesses with massive information and transcation online, with searchable data adopt big data frameworks to make sense of what is happening on their website. People are using Big data in healthcare even though its tricky due to privacy issues, but the step forward is to still adopt and use it.Archeologists are using it for pattern recognition of older motifs and machine learning has revealed patterns missed by the experts. Art, architecture and healthcare are few early adapters where production of data is extremely high.

He explains the constrains around the ownership of data and running analytics over it.

So a traditional data analytics is called an Intra firewall, meaning whatever data is available within the organization. Then there is something called Inter-firewall where organizations in similar domain make a data syndicate and run analytics over it. The third thing is called trans-firewall which is the analytics done beyond one’s firewall i.e. Analytics on public data. A lot of companies, like our clients, have intra firewall data that they augment with trans-firewall data for better decision-making.We are in this space largely looking at this problem industry by industry, currently focusing on retail, mainly on the products being shown online.

A Processing based code running a basic image processing algorithm over an existing image.
Credit: Code written by the author.

3. About the Image being the Data

Settling the conflict of disciplinary distance between Data Sciences and Image Processing, conversations with Image Processing Specialist Rahul Thota highlight that Image is yet another signal that is being processed by the Social Media Funnel. In his words:

Image processing is a subset of data sciences. Huge image datasets like Flickr, geotagging of pictures, recreating a scene using geotagging of pictures are typically overlapping areas. You must have heard of photosynth, an augmented reality generator using a lot of images. Similar applications are happening in healthcare with kidney CT towards understanding stage by stage kidney deformation.

He also goes on to explain the difference between applications which do Image Curation vis-à-vis applications that do Image Manipulation:

Pinterest or similar apps manage albums which are image datasets. Various pattern recognition algos try to define regions of interest in the image and also work on occlusion of redundant content (such as sky in the backgrocund etc). Face detection and number of faces is also a key parameter for sorting and ranking images. Contrast and details in an image could also be an important parameter. Curatorial sites also have machine learning algos which learn user behavior on album management. Algos also learn how data is labled and try to find what rules are satisfied by this labeling and classification. Rule engines thus manage the new ‘unlabeled’ image.

4. About the Image Blackbox

Rahul explains a basic image processing Blackbox through an Image posting or a filter camera:

Signal processing, Image processing and Graphics are related domains. Signal processing is largely measuring and drawing inference from any signal. Things like periodicity or frequencies are analyzed and inferences are taken. If you consider the image as a 2D signal that will you a fair idea of what image processing does. Graphics combined with image processing looks at rendering and recreating a scene, a popular example being the Raytracing algorithm which places a light source at infinity and calculates how the light gets reflects off various objects and thus renders the scene. Talking about the role image processing plays in social media channels, it’s largely that subjective preferences are standardized, with more details, like standard brightness and contrast enhancement. Instagram for example was a big hit because it makes the image look vintage. It notches up the red a little bit and then dulls the contrast. Image looks nostalgic. It took off because of the network effect.



[1] Refers to the human agents in the act, for example the shooter and the shot in case of the camera and the data producer and the data consumer in case of the Camera.

[2] Users produce such information through online forms, likes and comments on social media, mere movement of mouse from one section of page to another and offline movements in sensor environments such as access controlled rooms, public spaces with IP Cameras etc.

[3] Users consume the analytics output through dashboards, machine logs and automated to do lists or task lists. Auto switches such as action rules that machine or a user does after an alert has been generated come under this category.

[4] Castells, Manuel. 2011. The Rise of the Network Society: The Information Age – Economy, Society, and Culture. Vol. 1. John Wiley & Sons.

[5] See the first post in the series:

[6] For a quick scan see:

[7] For a quick scan see:

[8] Client side machines are the Computers used to access data stored on Servers using the internet. For more information see:

[9] For a quick scan see:

[10] For a quick scan see:

[11] Details on various machine learning algorithms:

[12] Information on Hadoop:

[13] Information on MapReduce:

[14] Elephant DB is a very minimalist DB used to serve MapReduce results from Hadoop. See:

[15] MemCached stores arbitrary data in-memory. See:

[16] Tableau augments data visualization workflows. See:

[17] D3 is a Javascript Library for visualization. See:

Share this:


Print Print

Categorised in: ,
Published on: October 9, 2014

No comments yet. Please leave a comment using the form below.

Leave a Reply

Your email address will not be published. Required fields are marked *

Comments Policy: We look forward to your comments on the posts published on this website. The comments should be constructive and conversational, as opposed to being offensive or merely critical. Personal attacks and rudeness are absolutely not tolerated. As the Editors of the website approve comments before they are published, there can be a slight delay in their appearance on the pages, especially during weekends. The comments are published under Creative Commons Attribution 2.5 India license. For any clarification, write to us at dak[at]sarai[dot]net.