System and method for machine learning-based delivery tagging

1. A system comprising:

one or more processors; and

one or more non-transitory computer-readable media storing computing instructions configured to run on the one or more processors and perform:

training a first submodel of a machine learning model by at least:

creating a cumulative addition of light gradient boosting models using features;

training each of the light gradient boosting models to reduce binary log loss, increase precision scores, and increase recall scores; and

determining, using a Bayesian Model Combination, (i) weights for aggregation with probabilities from the light gradient boosting models and (ii) a tuned threshold;

receiving historical delivery records over a predetermined time period from partners associated with items offered to subregions through an online platform;

generating nodes for combinations each comprising a respective one of the partners, a respective one of the items offered by the partners, and a respective one of the subregions;

generating, using the machine learning model, as trained, a respective classification for each respective node on whether to tag each respective node as deliverable in a predetermined time window, wherein the respective classification comprises a union of a respective output of the first submodel of the machine learning model and a respective output of a second submodel of the machine learning model; and

based on the respective classification for each respective node, automatically tagging a portion of the nodes as deliverable in the predetermined time window in the online platform.

2. The system of claim 1 , wherein

the features are from a first portion of the historical delivery records and responses from a second portion of the historical delivery records.

3. The system of claim 2 , wherein the features comprise partner features, item features, region features, and group features.

4. The system of claim 1 , wherein

determining the weights and the tuned threshold is performed at a per-seller level and at a full-data level.

5. The system of claim 1 , wherein:

training the first submodel further comprises:

using a second instance of the Bayesian Model Combination on the weights and the tuned threshold to combine a per-seller level and a full-data level.

6. The system of claim 4 , wherein generating the respective classification further comprises:

generating a respective probability score for each of the light gradient boosting models for each respective node; and

generating the respective output of the first submodel by applying the weights, as determined, to the respective probability scores to generate an aggregate probability that is compared against the tuned threshold.

7. The system of claim 1 , wherein the computing instructions are further configured to perform, before generating the respective classification for each respective node:

training the second submodel of the machine learning model using numerical features from the historical delivery records across a first rolling time period and a second rolling time period.

8. The system of claim 7 , wherein:

training the second submodel further comprises:

training a respective CatBoost model for each of the partners using the numerical features.

9. The system of claim 8 , wherein:

training the second submodel further comprises:

determining a respective threshold for the respective CatBoost model corresponding to each partner based on a respective second probability score for each node of the nodes; and

generating the respective classification further comprises:

generating the respective output of the second submodel based upon the respective threshold.

10. The system of claim 1 , wherein the computing instructions are further configured to perform:

monitoring a respective on-time-delivery (OTD) performance of the nodes over a subsequent time period after the portion of the nodes were tagged as deliverable in the predetermined time window;

automatically un-tagging a first node of the nodes when the respective OTD performance for the first node falls below one or more predetermined un-tagging thresholds; and

automatically re-tagging a second node of the nodes as the deliverable in the predetermined time window when the OTD performance exceeds one or more predetermined re-tagging thresholds.

11. A method being implemented via execution of computing instructions configured to run on one or more processors and stored at one or more non-transitory computer-readable media, the method comprising:

training a first submodel of a machine learning model by at least:

creating a cumulative addition of light gradient boosting models using features;

training each of the light gradient boosting models to reduce binary log loss, increase precision scores, and increase recall scores; and

determining, using a Bayesian Model Combination, (i) weights for aggregation with probabilities from the light gradient boosting models and (ii) a tuned threshold;

receiving historical delivery records over a predetermined time period from partners associated with items offered to subregions through an online platform;

generating nodes for combinations each comprising a respective one of the partners, a respective one of the items offered by the partners, and a respective one of the subregions;

generating, using the machine learning model, as trained, a respective classification for each respective node on whether to tag each respective node as deliverable in a predetermined time window, wherein the respective classification comprises a union of a respective output of the first submodel of the machine learning model and a respective output of a second submodel of the machine learning model; and

based on the respective classification for each respective node, automatically tagging a portion of the nodes as deliverable in the predetermined time window in the online platform.

12. The method of claim 11 , wherein:

the features are from a first portion of the historical delivery records and responses from a second portion of the historical delivery records.

13. The method of claim 12 , wherein the features comprise partner features, item features, region features, and group features.

14. The method of claim 11 , wherein

determining the weights and the tuned threshold is performed at a per-seller level and at a full-data level.

15. The method of claim 11 , wherein:

training the first submodel further comprises:

using a second instance of the Bayesian Model Combination on the weights and the tuned threshold to combine a per-seller level and a full-data level.

16. The method of claim 14 , wherein generating the respective classification further comprises:

generating a respective probability score for each of the light gradient boosting models for each respective node; and

generating the respective output of the first submodel by applying the weights, as determined, to the respective probability scores to generate an aggregate probability that is compared against the tuned threshold.

17. The method of claim 11 , further comprising, before generating the respective classification for each respective node:

training the second submodel of the machine learning model using numerical features from the historical delivery records across a first rolling time period and a second rolling time period.

18. The method of claim 17 , wherein training the second submodel further comprises:

training a respective CatBoost model for each of the partners using the numerical features.

19. The method of claim 18 , wherein:

training the second submodel further comprises:

determining a respective threshold for the respective CatBoost model corresponding to each partner based on a respective second probability score for each node of the nodes; and

generating the respective classification further comprises:

generating the respective output of the second submodel based upon the respective threshold.

20. The method of claim 11 , further comprising:

monitoring a respective on-time-delivery (OTD) performance of the nodes over a subsequent time period after the portion of the nodes were tagged as deliverable in the predetermined time window;

automatically un-tagging a first node of the nodes when the respective OTD performance for the first node falls below one or more predetermined un-tagging thresholds; and

automatically re-tagging a second node of the nodes deliverable in the predetermined time window when the OTD performance exceeds one or more predetermined re-tagging thresholds.