Data-Driven Foundations for Building Recommendation Systems

Although data is not yet listed on corporate balance sheets, it is only a matter of time before it becomes a vital asset, especially for recommendation systems where data preparation is a prerequisite for effective modeling.

Data drives business by enabling comprehensive product and user understanding through collection, mining, and decision-making, allowing targeted actions for the right users at the right time.

1. Understand the problem: Merely gathering massive data does not guarantee business impact; data quality and relevance are essential, and data analysis must be tightly coupled with business modeling to avoid noisy, ineffective outcomes.

2. Data-driven approaches for recommendation systems: Accurate data, appropriate methods, and correct interpretation are the core principles. Engineers should define analysis goals, identify necessary metrics, source data, process it, and derive actionable insights that guide recommendation iterations.

The pre‑recommendation stage requires dissecting user behavior into explicit actions (purchase, click, collect) and implicit signals (duration, skip), and analyzing who buys, what is bought, why they buy, user journey paths, individual tracking, and fine‑grained segmentation.

3. Data assessment dimensions: Completeness, timeliness, standardization, consistency, accuracy, and relevance must be evaluated to ensure data supports business needs.

4. Data types: Structured data (e.g., user profiles, click logs) and unstructured data (e.g., reviews, images) are both important for recommendation models.

5. Data collection (tracking): Define tracking points, plan metrics, and decide on data reporting frequency (real‑time, daily, weekly) to capture actions such as clicks, adds to cart, purchases, search queries, and recommendation exposures.

6. Post‑tracking processes: 6.1 ETL & data cleaning – cleanse and store data with attention to quality and scalability; 6.2 Data integration – combine data from multiple systems for a unified view; 6.3 Reporting – design visualizations and dashboards, ensuring data validity for operational decisions.

7. Recommendation‑specific data: User dimensions (basic info, explicit feedback, implicit behavior), item dimensions (title, tags, category), and auxiliary data (weather, location, holidays) are essential for building robust recommendation models.

Explicit feedback includes user ratings, while implicit feedback covers clicks, adds to cart, purchases, and dwell time, each offering different trade‑offs between cost and volume.

The article concludes with author information, a job posting for algorithm and development engineers, and an introduction to the DataFun community, emphasizing the practical sharing of data intelligence expertise.