Predictive Analytical Tools for Store Managers (HackGT 2019)

A predictive analytics tool that helps brick-and-mortar store managers decide what to stock together. It mines historical retail transactions with market-basket / association-rule analysis to surface the items most frequently bought together, and exposes those recommendations through a web app with an interactive tree-map visualization.

Built at HackGT 2019.

Store managers often need to know which products complement each other so they can plan placement, bundling, and inventory. This project answers the question "if a customer buys item X, what else are they likely to buy?" by:

1. Running an Apriori association-rule analysis over a real online-retail transaction dataset.
2. Persisting the resulting rules (antecedent → consequent with confidence) to a spreadsheet.
3. Serving those rules through a Node/Express web app, where a manager selects an item and sees the frequently-bought-together items rendered as a tree map sized and colored by confidence.

The analysis uses the Online Retail dataset (Online Retail.xlsx), a transactional dataset of purchases from a UK-based online retailer. Each row is a line item on an invoice (invoice number, stock description, quantity, country, etc.).

Preprocessing in the analysis:

• Trims item descriptions and drops rows without an invoice number.
• Removes cancelled transactions (invoice numbers containing C).
• Filters to the United Kingdom market.
• Pivots invoices into a one-hot basket matrix (one row per invoice, one column per item), and drops the POSTAGE pseudo-item.

The mined rules are written to assoc_rules.xlsx, which the web app reads at request time.

1. Basket construction – Transactions are grouped by invoice and item into a binary basket matrix where each cell indicates whether an item appeared on that invoice.
2. Frequent itemsets – mlxtend's apriori finds itemsets with min_support = 0.015.
3. Association rules – association_rules derives rules using lift (min_threshold = 1); antecedents and consequents are flattened to single item names.
4. Related-items lookup – For a selected item, the app filters the rules to that antecedent, groups consequents, and returns the average confidence (as a percentage) per related item.
5. Visualization – The related items and their confidence scores are rendered as a tree map, where larger / greener tiles indicate higher confidence.

The project has three layers:

• Python analytics – Generates association rules from the raw dataset (Get_other_frequent_Items.*) and serves per-item lookups (itemsList.py, getRelatedItems.py) against the precomputed assoc_rules.xlsx.
• Node/Express server (index.js) – Renders the home page and exposes a /getItem endpoint. It shells out to the Python scripts via python-shell, parses their stdout into JSON, and returns it to the browser.
• Frontend (views/, public/) – An EJS-rendered page with an item dropdown. On selection it calls /getItem, then draws an AnyChart/d3-style tree map of the related items (public/js/custom.js). A standalone d3 tree-map prototype lives in treeMap.html.

Request flow:

Online Retail.xlsx
   │  (apriori / association_rules, mlxtend)
   ▼
assoc_rules.xlsx
   │  itemsList.py  ───────────────►  populate item dropdown
   │  getRelatedItems.py <item> ────►  {related item: avg confidence %}
   ▼
index.js (Express + python-shell)
   ▼
home.ejs + custom.js  ──►  interactive tree-map of related items

• Analytics: Python, pandas, mlxtend (Apriori + association rules)
• Backend: Node.js, Express, EJS, python-shell
• Frontend: HTML/CSS/SASS, JavaScript, d3.js / AnyChart tree map, Tableau Public embed
• Data: Excel (.xlsx) via xlrd/pandas
• Deployment: Google App Engine (flexible environment)

• Python 3.7+
• Node.js and npm

pip install -r requirements.txt

Regenerate the association rules by running the analysis (e.g. open Get_other_frequent_Items.ipynb in Jupyter, or run python Get_other_frequent_Items.py). The web app reads the precomputed assoc_rules.xlsx.

Look up related items directly:

python getRelatedItems.py "JUMBO BAG PINK VINTAGE PAISLEY"

npm install
npm start

The server listens on http://localhost:8080 (or $PORT). Open it, pick an item from the dropdown, and click Frequently Together Bought Items to see the related-items tree map. The Express endpoints invoke the Python scripts, so Python and its dependencies must be available on the same machine.

Released under the MIT License.