MACHINE LEARNING TECHNIQUES FOR GENERATING PREDICTIONS FOR A TRANSACTION
Techniques for training a price prediction model are disclosed. An example method includes receiving historical transaction data comprising a plurality of transactions, each transaction comprising a plurality of attributes and a transaction price. The method also includes processing the historical transaction data to generate training data comprising features extracted from the plurality of attributes and price indices generated from the transaction price. The method also includes training, by a processing device, a price prediction model using the training data, wherein training the price prediction model comprises training a neural network to generate a mapping between the features and the price indices, and wherein the features used to train the neural network are not segmented and correspond with a plurality of products, product types, geographies, and customer sizes.
1 . A method comprising:
receiving historical transaction data comprising a plurality of transactions, each transaction comprising a plurality of attributes and a transaction price;
processing the historical transaction data to generate training data comprising features extracted from the plurality of attributes and price indices generated from the transaction price;
training, by a processing device, a price prediction model using the training data, wherein training the price prediction model comprises training a neural network to generate a mapping between the features and the price indices, and wherein the features used to train the neural network are not segmented and correspond with a plurality of products, product types, geographies, and customer sizes;
receiving, from a client device, a pricing request that describes a potential future transaction;
generating a price prediction for the potential future transaction using the trained price prediction model; and
sending, to the client device, a report comprising the price prediction.
2 . The method of claim 1 , wherein at least some of the features are categorical features extracted from categorical attributes of the plurality of attributes using a vector embedding technique that converts the categorical attributes to vector representations that capture semantic relationships and similarities between the categorical attributes.
3 . The method of claim 1 , wherein generating the price indices comprises, for each transaction:
converting the transaction price to a unit price that describes a price per unit of a product identified in the transaction; and
scaling the unit price by a normalizing value.
4 . The method of claim 3 , wherein the normalizing value is an average unit price of the product across the plurality of transactions.
5 . The method of claim 1 , wherein the features comprise cross features and side features, wherein the cross features are features that exhibit more significant feature interactions compared to the side features, and wherein the price prediction model comprises:
a deep neural network trained using the cross features and the side features; and
a cross network trained using only cross features.
6 . The method of claim 1 , wherein the price prediction model comprises:
a deep neural network trained using the features; and
a time network trained using time features, wherein the time features are a subset of the features that exhibit a greater time-dependent effect on prices due to trend or seasonality effects.
7 . The method of claim 6 , further comprising multiplying an output of the time network by one or more time functions to capture the trend and seasonality effects of the subset of the features.
8 . The method of claim 7 , further comprising passing the time functions directly to the last layer of the price prediction model to capture the seasonality and trend across all of the features.
9 . The method of claim 1 , wherein the price prediction model comprises three parallel subnetworks that each provide an output to a last layer of the price prediction model, the method further comprising passing the last layer of the price prediction model to an activation function to generate the price prediction.
10 . The method of claim 1 , wherein the price prediction is a predicted market price or a predicted customer-specific price.
11 . A system comprising:
a memory; and
a processing device, operatively coupled to the memory, the processing device to:
receive historical transaction data comprising a plurality of transactions, each transaction comprising a plurality of attributes and a transaction price;
process the historical transaction data to generate training data comprising features extracted from the plurality of attributes and price indices generated from the transaction price;
train a price prediction model using the training data, wherein to train the price prediction model comprises to train a neural network to generate a mapping between the features and the price indices, and wherein the features used to train the neural network are not segmented and correspond with a plurality of products, product types, geographies, and customer sizes.
12 . The system of claim 11 , wherein at least some of the features are categorical features extracted from categorical attributes of the plurality of attributes using a vector embedding technique that converts the categorical attributes to vector representations that capture semantic relationships and similarities between the categorical attributes.
13 . The system of claim 11 , wherein to generate the price indices comprises, for each transaction, to:
convert the transaction price to a unit price that describes a price per unit of a product identified in the transaction; and
scale the unit price by a normalizing value, wherein the normalizing value is an average unit price of the product across the plurality of transactions.
14 . The system of claim 11 , wherein the features comprise cross features and side features, wherein the cross features are features that exhibit more significant feature interactions compared to the side features, and wherein the price prediction model comprises:
a deep neural network trained using the cross features and the side features; and
a cross network trained using only the cross features.
15 . The system of claim 11 , wherein the price prediction model comprises:
a deep neural network trained using the features; and
a time network trained using time features, wherein the time features are a subset of the features that exhibit a greater time-dependent effect on prices due to trend or seasonality effects.
16 . The system of claim 15 , wherein the processing device is to multiply an output of the time network by one or more time functions to capture the trend and seasonality effects of the subset of the features.
17 . The system of claim 16 , wherein the processing device is to pass the time functions directly to the last layer of the price prediction model to capture the seasonality and trend across all of the features.
18 . The system of claim 11 , wherein the price prediction model comprises three parallel subnetworks that each provide an output to a last layer of the price prediction model, and wherein the processing device is to pass the last layer of the price prediction model to an activation function to generate the price prediction.
19 . A non-transitory computer-readable storage medium including instructions that, when executed by a processing device, cause the processing device to:
receive historical transaction data comprising a plurality of transactions, each transaction comprising a plurality of attributes and a transaction price;
process the historical transaction data to generate training data comprising features extracted from the plurality of attributes and price indices generated from the transaction price;
train, by the processing device, a price prediction model using the training data, wherein to train the price prediction model comprises to train a neural network to generate a mapping between the features and the price indices, and wherein the features used to train the neural network are not segmented and correspond with a plurality of products, product types, geographies, and customer sizes.
20 . The non-transitory computer-readable storage medium of claim 19 , wherein the features comprise cross features, side features, and time features, wherein the cross features are features that exhibit more significant feature interactions compared to the side features, wherein the time features are a subset of the cross features and time features that exhibit a greater time-dependent effect on prices due to trend or seasonality effects, and wherein the price prediction model comprises:
a deep neural network trained using the cross features and the side features; and
a cross network trained using only the cross features; and
a time network trained using the time features.