These instructions may help you determine if your model is suitable for providing product recommendations.
The questions to be addressed are:
- How much data do you need?
- How to select the appropriate parameters for the model?
- How to evaluate the model?
Note: These questions do not provide a definitive answer and the information provided here should be considered a guideline only.
In addition to tuning model settings to improve model quality that affects the performance of LS Recommend, you should consider the human factor as well if you are using staff facing recommendation. In the case of customer facing recommendation, it might depend more on where you display the recommendation rather than the quality of the recommendation.
Model data
The models provided by LS Recommend learn from usage data to predict which items are more likely to be of interest to the user. The quality of the model depends on the usage data itself, which means that you will need to experiment with your data.
A good rule of thumb is to have most items in 20 transactions or more. This means that if you have 10,000 items in your catalog, the total number of transactions should be around 200,000 transactions. The same rule can be applied to personalized recommendations, where the model takes member or customer ID as input. The difference is that here you must ensure that each member or customer has a sufficient history of transactions.
It is possible to have different time periods for usage history and catalog. There are two reasons for this:
- Catalog period shorter than usage history time period: This can be useful if you want to restrict the recommendation to items that have been recently created or sold, but you want to have a stronger signal for those items.
- Catalog period longer than usage history time period: This can be useful if you want to be able to give recommendation for a large set of items but only base your recommendation on recent sales. In this case you might be increasing the number of cold items and therefore, you should consider including features.
Model selection
The values of build parameters should be selected based on the quality of the provided data. The table below lists common cases and explains how to act on them.
| If your data includes | Then |
|---|---|
| Many cold items, that is many new items or many items with little sales history. | Consider selecting the Include Features In Catalog check box. |
| Large proportion of slow items, that is items that have steady sales but are low in volume. | Increase the Decay Period In Days build parameter. |
| If a large portion of the purchase history includes member or customer data. | Set true for the model. |
You can set the above parameters in Departments - LS Retail - LS Recommend - Model Template or for every model on the Model and Build card pages.
Model evaluation
The best way to choose between different models or builds is to test them in the real world.
If the recommendations are showing unexpected results, try the following if it describes your scenario.
| Scenario | Try |
|---|---|
| The model recommends items that are not frequently bought together. | Increase the Minimum Number of Cooccurrence Units build parameter. |
| The model recommends only warm items for cold items. | Select the Cold to Cold Item Recommendation check box to enable cold-to-cold item recommendations. |
You can set the above parameters in Departments - LS Retail - LS Recommend - Models - Builds.
Another way to evaluate a model's performance is to track the sales of recommended items for each model or build. See Reporting.