System and method for automatically computing and updating advertisement biddings for cold items
Systems and methods for automatically computing biddings for cold items are disclosed. A first bid value is computed and submitted for a first bidding for a first set of items each having a predicted sales metric larger than a first threshold in a first time period. From the first set of items, a second set of items are selected that have a traffic metric smaller than a second threshold in a second time period. Each item in the second set is assigned to one of a plurality of clusters. A second bid value is computed for each cluster based on cluster-level metrics and a click distribution among the plurality of clusters. An updated bid value is computed, for each item in the second set, to be submitted for a second bidding, based on the second bid value for a cluster to which the item is assigned.
1 . A system, comprising:
a non-transitory memory having instructions stored thereon; and
at least one processor operatively coupled to the non-transitory memory, and configured to read the instructions to:
receive, from a database, user transaction data and attribute data for a plurality of items in a first time period;
execute a trained prediction model to operate on the user transaction data and the attribute data and generate a predicted sales metric for each of the plurality of items;
select, from the plurality of items, a first set of items each having the predicted sales metric larger than a first threshold in the first time period;
execute, for each item in the first set, a first bidding model to generate a first bid value to be submitted for a first advertisement bidding;
select, from the first set of items, a second set of items each having a traffic metric smaller than a second threshold in a second time period;
execute a trained clustering model to operate on metrics and features of each of the second set of items and generate a plurality of clusters by assigning each item in the second set to one of the plurality of clusters;
execute a second bidding model to operate on cluster-level metrics of each cluster and a click distribution among the plurality of clusters and generate a second bid value for each of the plurality of clusters;
execute, for each item in the second set, the second bidding model to operate on the second bid value for a cluster to which the item is assigned and generate an updated bid value to be submitted for a second advertisement bidding; and
store the updated bud value for each item in the second set in a database.
2 . The system of claim 1 , wherein the trained clustering model is an asymmetrical clustering model.
3 . The system of claim 1 , wherein the trained clustering model is a k-means clustering model.
4 . The system of claim 1 , wherein the second bidding model generates the second bid value at least based on Bidirectional Encoder Representations from Transformers (BERT) embeddings.
5 . The system of claim 4 , wherein the BERT embeddings are provided to a trained model, and wherein the trained model is generated based on a training data structure constructed by emulating a process of transitioning from a cold state to a warm state.
6 . The system of claim 5 , wherein the training data structure includes a plurality of points, and wherein each point is labeled with a corresponding conversion rate.
7 . A computer-implemented method, comprising:
receiving, from a database, user transaction data and attribute data for a plurality of items in a first time period;
executing a trained prediction model to operate on the user transaction data and the attribute data and generating a predicted sales metric for each of the plurality of items;
selecting, from the plurality of items, a first set of items each having the predicted sales metric larger than a first threshold in the first time period;
executing, for each item in the first set, a first bidding model to generate a first bid value to be submitted for a first advertisement bidding;
selecting, from the first set of items, a second set of items each having a traffic metric smaller than a second threshold in a second time period;
executing a trained clustering model to operate on metrics and features of each of the second set of items and generating a plurality of clusters by assigning each item in the second set to one of the plurality of clusters;
executing a second bidding model to operate on cluster-level metrics of each cluster and a click distribution among the plurality of clusters and generating a second bid value for each of the plurality of clusters; and
executing, for each item in the second set, the second bidding model to operate on the second bid value for a cluster to which the item is assigned and generating an updated bid value to be submitted for a second advertisement bidding; and
storing the updated bid value for each item in the second set in a database.
8 . The computer-implemented method of claim 7 , wherein the trained clustering model is an asymmetrical clustering model.
9 . The computer-implemented method of claim 7 , wherein the trained clustering model is a k-means clustering model.
10 . The computer-implemented method of claim 7 , wherein the second bidding model generates the second bid value at least based on Bidirectional Encoder Representations from Transformers (BERT) embeddings.
11 . The computer-implemented method of claim 10 , wherein the BERT embeddings are provided to a trained model, and wherein the trained model is generated based on a training data structure constructed by emulating a process of transitioning from a cold state to a warm state.
12 . The computer-implemented method of claim 11 , wherein the training data structure includes a plurality of points, and wherein each point is labeled with a corresponding conversion rate.
13 . A non-transitory computer readable medium having instructions stored thereon, wherein the instructions, when executed by at least one processor, cause at least one device to perform operations comprising:
receiving, from a database, user transaction data and attribute data for a plurality of items in a first time period;
executing a trained prediction model to operate on the user transaction data and the attribute data and generating a predicted sales metric for each of the plurality of items;
selecting, from the plurality of items, a first set of items each having the predicted sales metric larger than a first threshold in the first time period;
executing, for each item in the first set, a first bidding model to generate a first bid value to be submitted for a first advertisement bidding;
selecting, from the first set of items, a second set of items each having a traffic metric smaller than a second threshold in a second time period;
executing a trained clustering model to operate on metrics and features of each of the second set of items and generating a plurality of clusters by assigning each item in the second set to one of the plurality of clusters;
executing a second bidding model to operate on cluster-level metrics of each cluster and a click distribution among the plurality of clusters and generating a second bid value for each of the plurality of clusters; and
executing, for each item in the second set, the second bidding model to operate on the second bid value for a cluster to which the item is assigned and generating an updated bid value to be submitted for a second advertisement bidding; and
storing the updated bid value for each item in the second set in a database.
14 . The non-transitory computer readable medium of claim 13 , wherein the trained clustering model is an asymmetrical clustering model.
15 . The non-transitory computer readable medium of claim 13 , wherein the trained clustering model is a k-means clustering model.
16 . The non-transitory computer readable medium of claim 13 , wherein the second bid value is determined at least based on Bidirectional Encoder Representations from Transformers (BERT) embeddings.
17 . The non-transitory computer readable medium of claim 16 , wherein the BERT embeddings are provided to a trained model, and wherein the trained model is generated based on a training data structure constructed by emulating a process of transitioning from a cold state to a warm state, and wherein the training data structure includes a plurality of points, and wherein each point is labeled with a corresponding conversion rate.
18 . The system of claim 1 , wherein the at least one processor comprises a first processor and a plurality of processing units of a plurality of virtual machines, wherein the first processor is configured to read the instructions to assign each of the trained prediction model, the first bidding model, the trained clustering model, and the second bidding model to at least one of the plurality of virtual machines for execution.
19 . The system of claim 1 , wherein the at least one processor is configured to read the instructions to:
receive, over a network, a bid request from a remote device; and
transmit, over the network, the updated bid value for each item in the second set to the remote device.
20 . The computer-implemented method of claim 7 , comprising assigning each of the trained prediction model, the first bidding model, the trained clustering model, and the second bidding model to at least one of a plurality of virtual machines for execution.