A retail chain’s general manager asks the sales director: ‘Why are some of our branches getting different results from the same campaign?’ The question moves to IT, and three different answers come back — one about a big data platform, one about a data science project, one about a machine learning model. The meeting ends without resolution. This scene plays out regularly in mid-sized and larger businesses. The problem is not technical; it is conceptual. These three terms overlap, feed into each other, and are often treated as synonyms. They are not.
Big data refers to datasets that exceed the processing capacity of conventional database tools in terms of volume, velocity, and variety. Daily clickstream logs from an e-commerce site, location data from a telecom operator, sensor output from a manufacturing floor — these fit the definition. Big data is fundamentally an infrastructure and storage question: where data is kept, how it moves, and how it can be accessed in raw form. Distributed file systems and data warehouse architectures operate at this layer. Big data alone produces no decisions; it only supplies the raw material.
Data science is the discipline that processes that raw material. Combining statistics, programming, and domain knowledge, data scientists extract meaningful patterns from raw data, test hypotheses, and produce findings that managers can act on. Data science does not require a big data infrastructure to be useful. A regression analysis run on two years of sales records from a mid-sized manufacturer is data science just as much as working on petabytes of log data. In Turkey, this distinction has not yet settled clearly in practice; the title ‘data scientist’ is often applied to anyone who knows SQL, which makes expectation management genuinely difficult.
Machine learning is one of the most powerful tools in the data science toolkit. Algorithms learn patterns from historical data and generate predictions against new inputs. Bank credit scoring, email spam filtering, inventory demand forecasting — all of these are machine learning applications. But machine learning is not magic. Without clean, labeled, and sufficiently large training data, no model performs reliably. Many companies stumble at exactly this point: years of accumulated data sit in inconsistent formats, ERP records are disconnected from CRM entries, and structured data from e-Invoice systems cannot be exported in a usable form. A data quality audit before starting a machine learning project is not optional; it directly determines whether the project succeeds or stalls within six months.
To place the three concepts in a working hierarchy: big data is the infrastructure layer, data science is the analysis layer, and machine learning is the automation and prediction layer. Which layer a business needs depends entirely on the question being asked. ‘Which product sold most last month?’ is a reporting question that requires no big data platform. ‘Which product will sell most next month?’ is a forecasting question that calls for data science and machine learning. ‘Which product should we recommend to which customer, and when?’ is a personalization question that demands robust infrastructure, a working model, and a continuously updated data pipeline. Setting up this mapping at the start of a project keeps scope and budget realistic.
The most common practical failure is ambiguity about which layer a project should begin from. A company announces it wants to build a machine learning model, but its databases are siloed, invoice data from the e-Invoice system has never been joined with CRM records, and nobody has defined what a clean training dataset would look like. In this situation, data integration and cleansing must come before any modeling work. Allocating a meaningful share of the project budget to infrastructure and data preparation — rather than concentrating it on model development — is the more defensible choice. This sequencing is not a workaround; it is the correct order of operations.
Cloud-based data platforms are making entry-level data science more accessible for smaller businesses in 2015, lowering the infrastructure cost that once made these projects the exclusive domain of large enterprises. Even so, the underlying logic does not change with the platform. A company that has not yet built consistent data collection habits will not benefit from a sophisticated cloud analytics stack. The tooling is not the bottleneck; the data discipline is.
For a manager deciding where to start: if your company’s data is not being collected consistently in a single system, big data infrastructure and machine learning projects are premature. The practical sequence is to build data management discipline first, develop analytical capacity with data science tools second, and automate repetitive forecasting tasks with machine learning third. This order reduces risk, produces measurable returns at each stage, and makes the total cost of ownership easier to justify to stakeholders. Distinguishing between these three concepts is not an academic exercise; it is the starting point for asking the right question with the right tool.
This article was originally written in Turkish by Gökhan MERCANOĞLU on July 6, 2015 and has been automatically translated into English and other languages using machine translation.