Empirical Discovery - Concept and Workflow Model
By Joe Lamantia on Jun 20, 2014
Concept models are a powerful tool for articulating the essential elements and relationships that define new or complex things we need to understand. We've previously defined empirical discovery as a new method, looking at antecedents, and also comparing and contrasting the distinctive characteristics of Empirical Discovery with other knowledge creation and insight seeking methods. I'm now sharing our concept model of Empirical Discovery, which identifies the most important actors, activities, and outcomes of empirical discovery efforts, to complement the written definition by illustrating how the method works in practice.
In this model, we illustrate the activities of the three kinds of people most central to discovery efforts: Insight Consumers, Data Scientists, and Data Engineers. We have robust definitions of all the major actors involved in discovery (used to drive product development), and may share some of these various personas, profiles, and snapshots subsequently. For reading this model, understand Insight Consumers as the people who rely on insights from discovery efforts to effect and manage the operations of the business. Data Scientists are the sensemakers who achieve insights, and create data products, and analytical models through discovery efforts. Data Engineers enable discovery efforts by building the enterprise data analysis infrastructure necessary for discovery, and often implement the outcomes of empirical discovery by building new tools based on the insights and models Data Scientists create.
A key assumption of this model is that discovery is by definition an iterative and serendipitous method, relying on frequent back-steps and unpredictable repetition of activities as a necessary aspect of how discovery efforts unfold. This model also assumes the data, methods, and tools shift during discovery efforts, in keeping with the evolution of motivating questions, and the achievement of interim outcomes. Similarly, discovery efforts do not always involve all of these elements.
To keep the essential structure and relationships between elements clear and in the foreground, we have not shown all of the possible iterative loops or repeated steps. Some closely related concepts are grouped together, to allow reading the model on two levels of detail.
For a simplified view, follow the links between named actors and groups of concepts shown with colored backgrounds and labels. In this reading, an Insight Consumer articulates questions to a Data Scientist, who combines domain knowledge with the Empirical Discovery Method (yellow) to direct the application of Analytical Tools (blue) and Models (salmon) to Data Sets (green) drawn from Data Sources (magenta). The Data Scientist shares Insights resulting from discovery efforts with the Insight Consumer, while Data Engineers may implement the models or data products created by the Data Scientist by turning them into tools and infrastructure for the rest of the business. For a more detailed view of the specific concepts and activities common to Empirical discovery efforts, follow the links between the individual concepts within these named groups. (Note: there are two kinds of connections; solid arrows indicating definite relationships, and for the Data Sets and Models groups, dashed arrows indicating possible paths of evolution. More on this to follow)
Another way to interpret the two levels of detail in this model is as descriptions of formal vs. informal implementations of the empirical discovery method. People and organizations who take a more formal approach to empirical discovery may require explicitly defined artifacts and activities that address each major concept, such as predictions and experimental results. In less formal approaches, Data Scientists may implicitly address each of the major concepts and activities, such as framing hypotheses, or tracking the states of data sets they are working with, without any formal artifact or decision gateway. This situational flexibility is follow-on of the applied nature of the empirical discovery method, which does not require scientific standards of proof and reproducibility to generate valued outcomes.
The story begins in the upper right corner, when an Insight Consumer articulates a belief or question to a Data Scientist, who then translates this motivating statement into a planned discovery effort that addresses the business goal. The Data Scientist applies the Empirical Discovery Method (concepts in yellow); possibly generating a hypothesis and accompanying predictions which will be tested by experiments, choosing data from the range of available data sources (grouped in magenta), and selecting initial analytical methods consistent with the domain, the data sets (green), and the analytical or reference models (salmon) they will work with. Given the particulars of the data and the analytical methods, the Data Scientist employs specific analytical tools (blue) such as algorithms and statistical or other measures, based on factors such as expected accuracy, and speed or ease of use. As the effort progresses through iterations, or insights emerge, experiments may be added or revised, based on the conclusions the Data Scientist draws from the results and their impact on starting predictions or hypotheses.
For example, an Insight Consumer who works in a product management capacity for an on-line social network with a business goal of increasing users’ level of engagement with the service wishes to identify opportunities to recommend users establish new connections with other similar and possibly known users based on unrecognized affinities in their posted profiles. The data scientist translates this business goal into a series of experiments investigating predictions about which aspects of user profiles more effectively predict the likelihood of creating new connections in response to system-generated recommendations for similarity. The Data Scientist frames experiments that rely on data from the accumulated logs of user activities within the network that have been anonymized to comply with privacy policies, selecting specific working sets of data to analyze based on awareness of the shoe and nature of the attributes that appear directly in users’ profiles both across the entire network, and among pools of similar but unconnected users. The Data Scientist plans to begin with analytical methods useful for predictive modeling of the effectiveness of recommender systems in network contexts, such as measurements of the affinity of users’ interests based on semantic analysis of social objects shared by users within this network and also publicly in other online media, and also structural or topological measures of relative position and distance from the field of network science. The Data Scientist chooses a set of standard social network analysisalgorithms and measures, combined with custom models for interpreting user activity and interest unique to this network. The Data Scientist has predefined scripts and open source libraries available for ready application to data (MLlib, Gephi, Weka, Pandas, etc.) in the form of Analytical tools, which she will combine in sequences according to the desired analytical flow for each experiment.
The nature of analytical engagement with data sets varies during the course of discovery efforts, with different types of data sets playing different roles at specific stages of the discovery workflow. Our concept map simplifies the lifecycle of data for purposes of description, identifying five distinct and recognizable ways data are used by the Data Scientist, with five corresponding types of data sets. In some cases, formal criteria on data quality, completeness, accuracy, and content govern which stage of the data lifecycle any given data set is at. In most discovery efforts, however, Data Scientists themselves make a series of judgements about when and how the data in hand is suitable for use. The dashed arrows linking the five types of data sets capture the approximate and conditional nature of these different stages of evolution. In practice, discovery efforts begin with exploration of data that may or may not be relevant for focused analysis, but which requires some direct engagement to and attention to rule in or out of consideration. Focused analytical investigation of the relevant data follows, made possible by the iterative addition, refinement and transformation (wrangling - more on this in later posts) of the exploratory data in hand. At this stage, the Data Scientist applies analytical tools identified by their chosen analytical method. The model building stage seeks to create explicit, formal, and reusable models that articulate the patterns and structures found during investigation. When validation of newly created analytical models is necessary, the Data Scientist uses appropriate data - typically data that was not part of explicit model creation. Finally, training data is sometimes necessary to put models into production - either using them for further steps in analytical workflows (which can be very complex), or in business operations outside the analytical context.
Because so much discovery activity requires transformation of the data before or during analysis, there is great interest in the Data Science and business analytics industries in how Data Scientists and sensemakers work with data at these various stages. Much of this attention focuses on the need for better tools for transforming data in order to make analysis possible. This model does not explicitly represent wrangling as an activity, because it is not directly a part of the empirical discovery method; transformation is done only as and when needed to make analysis possible. However, understanding the nature of wrangling and transformation activities is a very important topic for grasping discovery, so I’ll address in later postings. (We have a good model for this too…)
Empirical discovery efforts aim to create one or more of the three types of outcomes shown in orange: insights, models, and data products. Insights, as we’ve defined them previously, are discoveries that change people’s perspective or understanding, not simply the results of analytical activity, such as the end values of analytical calculations, the generation of reports, or the retrieval and aggregation of stored information.
One of the most valuable outcomes of discovery efforts is the creation of externalized models that describe behavior, structure or relationships in clear and quantified terms. The models that result from empirical discovery efforts can take many forms — google 'predictive model’ for a sense of the tremendous variation in what people active in business analytics consider to be a useful model — but their defining characteristic is that a model always describes aspects of a subject of discovery and analysis that are not directly present in the data itself. For example, if given the node and edge data identifying all of the connections between people in the social network above, one possible model resulting from analysis of the network structure is a descriptive readout of the topology of the network as scale-free, with some set of subgraphs, a range of node centrality values', a matrix of possible shortest paths between nodes or subgraphs, etc. It is possible to make sense of, interpret, or circulate a model independently of the data it describes and is derived from.
Data Scientists also engage with models in distinct and recognizable ways during discovery efforts. Reference models, determined by the domain of investigation, often guide exploratory analysis of discovery subjects by providing Data Scientists with general explanations and quantifications for processes and relationships common to the domain. And the models generated as insight and understanding accumulate during discovery evolve in stages from initial articulation through validation to readiness for production implementation; which means being put into effect directly on the operations of the business.
Data products are best understood as 'packages' of data which have utility for other analytical or business purposes, such as a list of users in the social network who will form new connections in response to system-generated suggestions of other similar users. Data products are not literally finished products that the business offers for external sale or consumption. And as background, we assume operationalization or ‘implementation’ of the outcomes of empirical discovery efforts to change the functioning of the business is the goal of different business processes, such as product development. While empirical discovery focuses on achieving understanding, rather than making things, this is not the only thing Data Scientists do for the business. The classic definition of Data Science as aimed at creating new products based on data which impact the business, is a broad mandate, and many of the position descriptions for data science jobs require participation in product development efforts.
Two or more kinds of outcomes are often bundled together as the results of a genuinely successful discovery effort; for example, an insight that two apparently unconnected business processes are in fact related through mutual feedback loops, and a model explicitly describing and quantifying the nature of the relationships as discovered through analysis.
There’s more to the story, but as one trip through the essential elements of empirical discovery, this is a logical point to pause and ask what might be missing from this model? And how can it be improved?