CRISP-DM for Data Science Projects

• CRISP-DM originated in the second half of the nineties.
• It is still considered the de facto standard for data mining and knowledge discovery projects.
• Data science is now favored over data mining.
• Factors investigated include CRISP-DM's fitness for data science projects.
• The process model view still holds if the project is goal-directed and process-driven.
• A more flexible model is called for when data science projects become more exploratory.
• Seven real-life exemplars are examined where exploratory activities play an important role.
• They are compared against 51 use cases extracted from the NIST Big Data Public Working Group.
• The categorization can help project planning in terms of time and cost characteristics.

Introduction

• Toward the end of the previous century, companies and institutions joined forces to identify good practices and common mistakes.
• With funding from the European Union, a team of data mining engineers developed a broadly applicable data mining methodology.
• In 1999 the first version of CRISP-DM was introduced, designed to catalogue and guide common steps in data mining projects.
• In the last two decades, electronic devices and sensors, social networks, and data storage have increased.
• They have increased opportunities for extracting knowledge through data mining projects.
• The diversity of data has increased in origin, format, and modalities.
• The variety of techniques coming from machine learning, data management, visualization, and causal inference has increased.
• Data can be monetized in many more ways, through new applications, interfaces, and business models.
• The area of deriving value from data has grown exponentially in size and complexity.
• In contrast to traditional data mining, data-driven and knowledge-driven stages interact.
• Traditional data mining starts from business goals that translate into a clear data mining task.
• New methodologies have been proposed to accommodate some of the changes.
• IBM introduced ASUM-DM [3], and SAS introduced SEMMA [4].
• The original CRISP-DM model can still be recognized in more recent proposals.
• It remains focused on the traditional paradigm of a sequential list of stages from data to knowledge.
• After twenty years, to what extent do the original CRISP-DM and the underlying data mining paradigm remain applicable.
• It is applicable for the much wider range of data science projects seen today.

Limitations of the original CRISP-DM

• Recognition of the limitations of the original CRISP-DM and other related methodologies is key.
• The diversity of data science projects today is considered.
• Exploratory activities that are common in data science but not covered by CRISP-DM are identified.
• Recognition of popular trajectories in this space describes practices in data science.
• Popular trajectories can be used as templates, making the DST model exemplary rather than prescriptive.
• Some general suggestions are made on how the DST model can be coupled with project management methodologies.
• The DST model can be customised to different organisations and contexts.

Organization of the Paper

• CRISP-DM and other related variations are revisited.
• The identification of new activities and the formulation of the DST map are included.
• These trajectories are illustated on real cases of data science projects.
• A precise notation on trajectory charts is used.
• Data science project management is discussed.
• Three kinds of activities are considered for these seven real cases plus 51 use cases from the NIST. Comparison of the model with software methodologies and the scientific method
• It suggests how organisations can couple this with new methodologies and ethical issues and automation.

The Goal standard

• CRISP-DM is overhauled
• It is cruicial to not discard CRISP-DM
• It represents common trajectories in data science
• It goes from data to knowledge when there is a goal that translates data mining
• the DST is "backwards compatible" with the need for flexibility
• Some other trajectories that capture the common routes of data science projects are identiified

CRISP-DM and Related Process Models

• Section 2 gives a succinct description of the most used and cited data mining.
• It provides an overview of its evolution, basis and primary characteristics.
• Knowledge Discovery in Databases (KDD)overall process of knowledge discovery from data.
• It includes how the data is stored and accessed, how algorithms can be scaled to massive data-sets and still run efficiently.

Fayyad, Piatetsky-Shapiro and Smyth

• How results can be interpreted and visualised, how the overall human-machine interaction can be modeled and supported"
• Data mining is a step in this process, turning pre-processed data into patterns that can be turned into knowledge [7]. Data mining is often used as a synonym for KDD, and it will not distinguish between the two meanings in this paper.
• It elaborates and extends the steps in the original KDD proposal into six steps.
• Business understanding, Data understanding, Data preparation, Modeling, Evaluation, and Deployment are key.
• The way they are sequenced in a data mining application.

Process models and methodologies:

• Human-Centered Approach to Data Mining: Involves a holistic understanding of the entire Knowledge Discovery Process and interpretations.
• SEMMA stands for Sample, Explore, Modify, Model and Assess.
• Cabena's model: used in the marketing and sales domain.
• Buchner's model is adapted to the development of web mining projects focused on an online customer.
• Two Crows: takes insights from (first versions of) CRISP-DM (before release) and proposes a non-linear list of steps.
• D3M, a domain-driven data mining approach.
• The 5 A's Process originally developed by SPSS, already included an “Automate” step. Motorola developed the 60 approach stressing statistical control to excellence in management.
• KDD Roadmap is an iterative data mining methodology.
• The arrows in the figure indicate that CRISP-DM incorporates principles and ideas from most of the aforementioned methodologies.
• CRISP-DM is still considered the most complete data mining methodology in terms of meeting the needs of industrial projects.

Modernizing CRISP-DM

• The data mining methodology has evolved in the business application of data mining
• Several new methodologies have appeared as extensions of CRISP-DM
• The CRISP-DM 2.0 Special Interest Group (SIG) was established.
• SIG’s Goal: Modernize without fundamental changes
• Cios et al.'s Six-step discovery process adapts the CRISP-DM model to academia.
• RAMSYS (RApid collaborative data Mining SYStem) is another way to develop collaborative DM and KD projects with geographically diverse groups

Crisp-DM and extensions/personalizations

• ASUM-DM refines and extends CRISP-DM.
• It adds infrastructure, operations, deployment and project management.
• It has templates and guidelines, personalised for IBM's practices.
• CASP-DM addresses specific challenges of machine learning and data mining.
• HACE is a Big Data processing framework based on a three tier structure
• Mining platform (Tier I), challenges on information sharing and pri-vacy, and Big Data application domains (Tier II), and Big Data mining algorithms (Tier III).

Business Understanding

• Spend time with the business.
• Methodologies have a great deal that aims at gathering information before starting a data mining.
• Current data deluge as well as the experimental / exploratory nature of data requires lightweight and flexible methods.
• Similar lifecycles / methodologies have bee introduced by IT companies for data science
• IBM released the Foundational Methodology for Data Science (FMDS).
• Microsoft released TDSP a data science methodology to deliver predictive analytics solutions and applications effeciently
• FMDS and TDSP = CRISP DM
• FMDS and TDSP = data mining methodologies assuming an identifiable Goal, unlike data science requires an exploratory mindset

Processes and Data

• Mining is like starting mining at place known for minerals and metals
• Data is the valuable knowledge
• Whenever this metaphor is applicable, it's suggested that CRISP-DM is methodology
• Data science is more commong now than Data mining in context of discovery a Google Trends query showed this
• Term used a lot now
• Two Broad senses: the science OF data applying scientific method = testing machine Learning

Data Science is Data Oriented and Exploratory

• CRISP is about the processes and tasks = process takes centre stage.
• In Dat Science = Data takes Centre stage.
• What happens to the Data? What possible operations can apply? While move away = this becomes more inquisitive & prescriptive, things you can do to data.
• Continue with mining metaphor, data mining = previous data Data Science Prospecting searching precious metals that can be located

Data prospecting

• Find business goals
• discover and new value
• Data value explanation
• Relating data science results
• Narrative exploration - extracting valuable stories
• Product exploration: turn extracted data value that comes from consumers

Data Science and CRIP Example

• it is possible two see (Weak) Links business data
• Data Depends on decisions of scientist
• unstatifactory result
• need to further source
• No data = source before value

Managing stages of projects

• exploration with parts / steps of DM
• activities from start to deplyment are there
• partial tracers happen
• there is work but no activity beyond preperation
• CRISP interrupted by explorer, new data sets goals
• successful plan and project is one like the one depicted

What a data Project Requires

• Contrasted to CRISP
• No Arrows. No Predetermination.
• Leader decides the next
• Contatins these phases
• exploratory = these can set more information
• Take in science and
• modelling will depend on carrying out over

Technical data

• “data” is not fully capturing work
• The variety
• Applications now Especially under the “business intelligence” extraction
• data from other source For Self own the following data acquisition simulation

The variety of models

• acquisition
• Simulation: simulate to produce
• architecture
• Release: to database

Trajectories

• Activities have been introduced trajectory.
• The graph over sequence is forking
• Give in the data
• Way the exploration
• knowledge value
• Activities cleaning, model transformation
• presented
• Examples in next
• Remove this activities Trajectorey

DST (data science trajectorey)

• dst chart = graph w activities conections .
• arrows are labelled from sequence activities
• cannot be loops
• the types are circles
• If two they take place as well
• The trajectory can go through the same activity many more times
• annotate as label - to show more transition
• chart is charted
• the charts make model

Types of data science trajectories

• It's no way of complete = demonstrate data
• Data Acquisition is more than expensive + complex
• analyzing different scenarios
• estimating alternates
• weather, fuel, byproducts
• Also stimulating the of certain parts of city
• the city to predict the pollution
• Estimated using volume. Integrate a model

Insurance /refininf policies

• Using history to Clients data can be can get
• Data for example The info can be to pay the data
• Data where has the product for the customer enrich other data to what needs by means needed