Model and Recommendation Generation
LS Recommend Batch is calculated using the Calculate Recommend Batch function after the master data and sales history data are synchronized and uploaded to the LS Recommend engine.
The calculated LS Recommend Batches are then loaded to the Model Reader using the Load Recommend Batch function. Item recommendations can then be retrieved from LS Recommend's Model Reader via POS terminals, eCommerce sites, or other integrated systems.
This topic describes how a model is created in the cloud, and how recommendations are then made with the calculated model.
Model Generation
LS Recommend Batch models are generated by running an association rule mining algorithm on prior transaction history. The algorithm generates a model with items that have a strong association between themselves. Items are considered to have a high association if the items were often sold together in the past.
For example, a store sells the following items: smoke detectors, batteries, knives, forks, and spoons with the following transaction history.
Transaction No. | Items in basket |
1 | “Batteries”, “Smoke detector” |
2 | “Batteries”, “Smoke detector”, “Knife” |
3 | “Batteries”, “Smoke detector”, “Fork” |
4 | “Batteries”, “Knife” |
5 | “Batteries” |
6 | “Batteries”, “Fork”, “Knife” |
7 | “Batteries”, “Fork” |
8 | “Batteries”, “Spoon” |
9 | “Batteries”, “Fork”, “Spoon” |
The LS Recommend engine then looks at the transaction history and picks a minimum support value for the dataset. Let’s say the engine picked the value 0.3, which means that item combinations that appear in less than 30% of transactions are not considered for recommendations. For example, the item “Spoon” only appears in two transactions (<30%) and is therefore not considered as a rule. The LS Recommend engine then proceeds to calculate lift values for the remaining (common) item set combinations. The formula for calculating lift is shown below. Lift is a measurement of how strong an association is between two items or item sets.
Once the lift values have been calculated for all common item sets, the engine selects association rules above a lift threshold, usually 1.1, and creates a model based on those rules. The following rules would be generated by the dataset above.
Item set | Recommendation | Confidence | Lift Value |
“Smoke detector” | “Batteries” | 1.0 | 1.8 |
“Fork” | “Batteries” | 0.75 | 0.84 |
“Batteries” | “Fork” | 0.375 | 0.33 |
“Batteries” | “Smoke detector” | 0.375 | 1.125 |
From these rules, you can see that only two rules have a lift value above 1.0. Out of those two rules, the rule “Smoke detector” then “Batteries” has a confidence value of 1.0, which means that it applies for 100% of historical transactions, indicating a strong rule. Note: The “Batteries” then “Smoke detector” rule has a positive lift value, indicating association, but the confidence value is not as high. The confidence value is not as high because “Batteries” are included in every transaction making the rule “Batteries” then “Smoke detector” less useful than “Smoke detector” then “Batteries”. The rules generated with lift value less than 1.0 will never be returned in the final model.
Recommendation Generation
A model reader component can interpret a model and provide recommendations based on items submitted in a basket. When a model reader receives a basket, it will look for rules that match the basket and return recommendations with the highest lift value.
For example, a basket consisting of the items “Fork”, “Smoke detector”, and “Batteries, checks if there is a match for the rules below.
“Fork”, “Smoke detector”, “Batteries” |
“Fork”, “Smoke detector” |
“Fork”, “Batteries” |
“Fork” |
“Smoke detector”, “Batteries” |
“Smoke detector” |
“Batteries” |