The prototype runs on 4 billion records of POS data (~400GB in memory, 2TB raw data). The 4 billion records are partitioned into 100 partitions over the transaction ID. Additionally we have master data of products and stores, which include for example the names of products.

The basket analysis on the second page is based on an apriori algorithm. It calculates products that are most commonly sold together with the selected product. These products are shown in the table on the right. The support and confidence value indicates how strong the correlation between the two products is. The support value illustrates how many transactions contained both products. The confidence value shows the percentage of the transactions, which contained product A, also contain product B.

For each partition, we count the number of transactions that contain product A & B and the number of transactions that contain only product A. Afterwards, the quantity is added and divided after the apriori algorithm rule.

For selected products out of the table, you can see further information in the circle on the left. The size of the blue area represents the average basket value where both products were sold together.

At the top of the last page, the matrix shows the average revenue and basket value generated by two selected products over the days of the week and times of the day.
To increase the significance of this matrix, a next step could be to calculate the relative amount of revenue compared to the total revenue. Thereby outliers would be easier to detect. Additionally, one could consider adding filters to only analyze a subset of data, e.g. only a specific period of time.

At first, we get all transactions where both products were sold together. For these transactions, we calculate the basket value and select the weekday and hour. Afterwards, we add the basket values and count the number of transactions for each weekday and hour. This calculation happens on each partition. Following, the average basket value is computed.