Organization administrators and workspace administrators can configure an algorithm model for User-Product Matching. The model optimizes the sorting of rule-based matching results and generates baseline predictions for user groups not covered by any rules.
The model trains on data from the associated user tag datasets, product tag datasets, and behavioral datasets. Model accuracy improves with higher data quality and larger data volumes.
Training data must meet the following requirements:
Behavioral dataset: The behavior object must have product granularity — the object attribute must be "product" and the object attribute value must be the product name. The dataset must contain at least one year of purchase behavior data (data from the last two years is optimal), with a minimum of 1 million records.
User tag datasets and product tag datasets: These datasets must contain only static attribute tags, such as gender, age, and product price. Derived statistical tags — such as purchase amount in the last 90 days — interfere with model training and must not be included.
The users and products in the user tag and product tag datasets must match the users and products in the behavioral dataset.
Configure the algorithm model
Go to User Insights > Marketing Models > Model Hub > User-Product Analysis > Model List. The configuration page appears.
-
Note
If an algorithm model is already configured, click Edit to open the configuration page. Only one high-potential model per workspace can be in a running state at a time. Running states include Not Started, Pending Training, Training, and Training Successful.
Configure the required parameters, including the order details table, order purchase time, positive behaviors (such as purchases), and the product tag dataset.
Has statistical tags: Statistical tags — such as the sales volume of a product in the last N days — can interfere with model training. Enable this option and select all statistical tags in the current dataset. The system filters out these tag features during training.
Prediction period: Enter a value between 15 and 90 days. Set this value based on your business needs, because the prediction period varies by brand. For example, if the average prediction period for a brand is 60 days, enter 60.
Updating the prediction period reduces the number of available models.
Avoid using statistical tags with a prediction period range of 15 to 90 days in your datasets, as they fall within the model's prediction window and will be filtered out.
Click Save and Execute. In the dialog box that appears, click Confirm to start model training.
Manage algorithm models
A workspace can have only one algorithm model — always the most recently submitted one. View recent training runs in the model training list.
Note the following constraints:
Only one product recommendation model per workspace can be in a running state at a time.
Running states include Not Started, Pending Training, Training, and Training Successful.
A successfully trained model cannot be restored after it is unpublished.
A model in a failed training state can be edited and retrained.
The following operations are available on a model:
-
Click Model Details to open the model validation page and view detailed model data and prediction results.
The model validation page shows the analysis conditions used for the model.
The page also shows the audience predicted by the model and the applied filtering rules.
The evaluation metrics show the top N product recommendation records for the user group.
-
Click the Product Association Prediction tab to view the product association distribution graph based on the model's predictions.
Product association
-
Single product association
Select a specific product to view other products associated with it.
In Product Association Distribution view, the detailed data shows the top 10 products by sales revenue.
In Single Product Association view, the detailed data shows the distribution of the top 10 associations for that product.
Update: Updating the model uses one of your available model slots. Failed updates do not consume a slot. When an update starts, the previously successful model is unpublished.
Unpublish: A successfully trained model cannot be restored after you unpublish it.
-
If training fails, edit the model, adjust the conditions, and start a new training run.
NoteUpdate the model when the volume of your training data changes significantly.
End Training: For models in the Pending Training or Training state, click End Training to stop the process. This returns you to the model editing page.
