Multi-model based account/product sequence recommender
To automatically identify a sequence of recommended account/product pairs with highest likelihood of becoming a realized opportunity, an account/product sequence recommender uses an account propensity (AP) model and a reinforcement learning (RL) model and target engagement sequence generators trained on historical time series data, firmographic data, and product data. The trained AP model assigns propensity values to each product corresponding to received account characteristics. The trained RL model generates an optimal sequence of products that maximizes the reward over future realized opportunities. The target engagement sequence generators create target engagement sequences corresponding to the optimal sequence of products. The recommender prunes the optimal sequence of products based on the propensity values from the trained AP model, the completeness of these target engagement sequences, and a desired product sequence length. The recommender uses the remaining products, validated on three models, for account/product recommendations.
1. A method comprising:
generating a first plurality of product propensity values with a trained propensity model that has been trained with first training data corresponding to a first account, wherein the first plurality of product propensity values generated for a first plurality of product identifiers;
generating a first product path with a trained reinforcement learning model, wherein the first product path comprises a second plurality of product identifiers and wherein the trained reinforcement learning model was trained with the first training data and wherein the first plurality of product identifiers includes the second plurality of product identifiers;
determining that a first subset of the first plurality of product propensity values satisfies a propensity threshold for the first account;
identifying a first subset of the second plurality of product identifiers that corresponds to the first subset of product propensity values; and
indicating the first subset of the second plurality of product identifiers in a sequence according to the first product path.
2. The method of claim 1 further comprising:
detecting account characteristic values corresponding to the first account; and
selecting the trained propensity model from a plurality of trained propensity models and the trained reinforcement learning model from a plurality of trained reinforcement learning models, based, at least in part, on the detected account characteristic values.
3. The method of claim 2 , wherein each of the plurality of trained propensity models and each of the plurality of trained reinforcement learning models were trained with data corresponding to different accounts.
4. The method of claim 1 further comprising determining, for each product identifier in the first subset of the second plurality of product identifiers, a sequence of contacts to engage about a product identified by the product identifier.
5. The method of claim 4 , wherein determining, for each product identifier in the first subset of the second plurality of product identifiers, a sequence of contacts to engage comprises:
for each product identifier in the first subset of the second plurality of product identifiers, retrieving a trained artificial recurrent neural network from a plurality of artificial recurrent neural networks based, at least in part, on product characteristic values of a product identified by the product identifier and account characteristic values for the first account;
generating a sequence of organizational classifications with each of the retrieved artificial recurrent neural networks; and
identifying a contact for each of the sequence organizational classifications.
6. The method of claim 5 further comprising determining that each of the first subset of the second plurality of product identifiers corresponds to a sequence of organizational classifications with a sequence completeness above a threshold sequence completeness value.
7. The method of claim 1 further comprising determining that each of the first subset of the second plurality of product identifiers is within a threshold distance of the first product path.
8. The method of claim 1 , wherein the trained propensity model is a random forest based model.
9. The method of claim 1 , wherein the trained reinforcement learning model is trained with Q-learning.
10. A non-transitory, machine-readable medium having program code stored thereon that is executable by a machine, the program code comprising instructions to:
invoke a first trained propensity model to generate a first plurality of propensity values for a first plurality of products a first account;
invoke a first trained reinforcement learning model to generate a first product path for a first product of the first plurality of products that corresponds to a maximal of the first plurality of propensity values, wherein the first product path identifies a first set of products from the first plurality of products in a particular order to present to the first account for a highest likelihood of a successful outcome for the first product; and
associate contacts of the first account for at least a first subset of the first set of products.
11. The non-transitory, machine-readable medium of claim 10 , wherein the program code further comprises instructions to prune the first product path based, at least in part, on a propensity threshold.
12. The non-transitory, machine-readable medium of claim 11 , wherein the instructions to associate contacts comprise instructions to associate contacts of the first account for those of the first set of products remaining after the pruning.
13. The non-transitory, machine-readable medium of claim 11 , wherein the propensity threshold corresponds to the first account.
14. The non-transitory, machine-readable medium of claim 10 , wherein the instructions to invoke the first trained propensity model comprise instructions to invoke the first trained propensity model with first input comprising characteristics of the first account and product features of the first plurality of products.
15. The non-transitory, machine-readable medium of claim 10 , wherein the program code comprises instructions to:
invoke a second trained reinforcement learning model to generate a second product path for a second product of the first plurality of products that corresponds to the maximal or a second maximal of the first plurality of propensity values, wherein the second product path identifies a second set from the first plurality of products in a particular order to present to the first account for a highest likelihood of a successful outcome for the second product; and
associate contacts of the first account for at least a first subset of the second set of the first plurality of products.
16. The non-transitory, machine-readable medium of claim 10 , wherein the program code further comprises instructions to:
for each product of at least the first subset of the first set of the first plurality of products, generate a sequence of organizational classifications with a trained artificial recurrent neural network retrieved based, at least in part, on the first product,
wherein the contacts are determined from the sequence of organizational classifications.
17. An apparatus comprising:
a processor; and
a machine-readable medium having instructions stored thereon that are executable by the processor to cause the apparatus to,
detect account characteristic values corresponding to a first account;
retrieve a trained propensity model from a plurality of trained propensity models and a trained reinforcement learning model from a plurality of trained reinforcement learning models, based, at least in part, on the detected account characteristic values;
generate a first plurality of product propensity values with the trained propensity model based, at least in part, on the detected account characteristic values and product features of a first plurality of products;
generate a first product path for a first product of the first plurality of products with the trained reinforcement learning model, wherein the first product corresponds to a maximal one of the first plurality of product propensity values and the first product path identifies a first set of products of the first plurality of products including the first product;
prune the first product path based on a propensity threshold for the first account and those of the first plurality of product propensity values corresponding to the first set of products; and
indicate the pruned first product path.
18. The apparatus of claim 17 , wherein the machine-readable medium further comprises instructions executable by the processor to cause the apparatus to:
determine, for each product identified in the pruned first product path, a sequence of contacts to engage about the product.
19. The apparatus of claim 18 , wherein the instructions to determine, for each product identified in the pruned first product path, a sequence of contacts to engage comprise instructions to:
for each product identified in the pruned first product path, retrieve a trained artificial recurrent neural network from a plurality of artificial recurrent neural networks based, at least in part, on product characteristic values of the identified product and account characteristic values for the first account;
generate a sequence of organizational classifications with each of the retrieved artificial recurrent neural networks; and
identify a contact for each of the sequence of organizational classifications.
20. The apparatus of claim 17 , wherein the machine-readable medium further comprises instructions executable by the processor to cause the apparatus to:
determine whether each of the first set of products is within a threshold distance of the first product path.