Method, system, and computer program product for representational machine learning for product formulation
A method, system, and computer program product for representational learning of product formulas are provided. The method accesses a set of product formulas. Each product formula includes a set of ingredient tuples. A directed graph is generated from the set of product formulas. The directed graph including a node for each ingredient of the sets of ingredient tuples of the set of formulas. The method generates a weighted graph from the directed graph. The weighted graph has a weight assigned to each edge in the directed graph. The method generates an embedding model based on the directed graph. A set of embeddings is determined for the weighted graph where each node is represented with low-dimensional numerical vectors.
1. A computer-implemented method, comprising:
accessing a set of product formulas, each product formula including a set of ingredient tuples;
generating a directed graph from the set of product formulas, the directed graph including a node for each ingredient of the sets of ingredient tuples of the set of product formulas;
generating a weighted graph from the directed graph, the weighted graph having a weight assigned to each edge in the directed graph;
modifying random walk models by incorporating the weight assigned to each edge into the random walk model;
vectorizing the set of product formulas based on ingredients within the product formulas;
feeding the random walks from the random walk models into a deep learning approach to determine a set of embeddings comprising ingredient embeddings and a set of ingredient vectors from the vectorizing, wherein the deep learning approach is run on data including ingredient recommendations; and substitute ingredient recommendation;
generating, based on the weighted graph, an embedding module;
deploying the random walk models on the weighted graph;
building neural models for two or more positive formulas in the weighted graph by: extracting, using the random walk models, a subgraph with depth d; extracting n biased walks in the subgraph; and building a neural model for each entity in the weighted graph, wherein positive formulas are used to generate a weight of edges to provide the bias for the biased walks on the weighted graph, wherein positive examples are formulas that have a threshold level of market success;
training, using a bias of the biased walks from the weight of the edges, a sequence-to-sequence neural network generated using a pairwise ranking loss functions derived from the neural models;
augmenting the sequence-to-sequence neural network with the set of embeddings for the weighted graph as a sigmoid layer in the sequence-to-sequence neural network,
wherein each node is represented with low-dimensional numerical formula vectors;
deploying the embedding module to generate a new product formula by passing the set of ingredient embeddings through the sequence-to-sequence neural network using one or more of the neural models; and
identifying, using the sequence-to-sequence neural network, a new ingredient tuple for the new product formula.
2. The method of claim 1 , wherein generating the weighted graph further comprises:
removing a subset of edges of the directed graph, the subset of edges having a weight below a first threshold.
3. The method of claim 1 , wherein each ingredient tuple identifies an ingredient, an amount of the ingredient, and a set of categorical features of the ingredient.
4. The method of claim 3 , wherein the set of embeddings include a set of formula embeddings and determining the set of embeddings further comprises:
determining a set of formula vectors based on the set of ingredient vectors, the set of formula vectors including a formula vector for each product formula of the set of product formulas.
5. The method of claim 1 , further comprising:
generating a new formula based on the sequence-to-sequence neural network.
6. The method of claim 5 , wherein generating the new formula further comprises:
generating a set of proposed formulas based on the sequence-to-sequence neural network;
eliminating a first subset of proposed formulas from the set of proposed formulas based on a similarity threshold to a formula of the set of product formulas; and
based on eliminating the first subset of formulas, selecting the new formula from a second subset of formulas remaining in the set of proposed formulas.
7. A system, comprising:
one or more processors; and
a computer-readable storage medium, coupled to the one or more processors, storing program instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
accessing a set of product formulas, each product formula including a set of ingredient tuples;
generating a directed graph from the set of product formulas, the directed graph including a node for each ingredient of the sets of ingredient tuples of the set of product formulas;
generating a weighted graph from the directed graph, the weighted graph having a weight assigned to each edge in the directed graph;
modifying random walk models by incorporating the weight assigned to each edge into the random walk model;
vectorizing the set of product formulas based on ingredients within the product formulas;
feeding the random walks from the random walk models into a deep learning approach to determine a set of embeddings comprising ingredient embeddings and a set of ingredient vectors from the vectorizing, wherein the deep learning approach is run on data including ingredient recommendations, and substitute ingredient recommendation;
generating, based on the weighted graph, an embedding module;
deploying the random walk models on the weighted graph;
building neural models for two or more positive formulas in the weighted graph by: extracting, using the random walk models, a subgraph with depth d; extracting n biased walks in the subgraph; and building a neural model for each entity in the weighted graph, wherein positive formulas are used to generate a weight of edges to provide the bias for the biased walks on the weighted graph, wherein positive examples are formulas that have a threshold level of market success;
training, using a bias of the biased walks from the weight of the edges, a sequence-to-sequence neural network generated using a pairwise ranking loss functions derived from the neural models;
augmenting the sequence-to-sequence neural network with the set of embeddings for the weighted graph as a sigmoid layer in the sequence-to-sequence neural network, wherein each node is represented with low-dimensional numerical formula vectors;
deploying the embedding module to generate a new product formula by passing the set of ingredient embeddings through the sequence-to-sequence neural network using one or more of the neural models; and
identifying, using the sequence-to-sequence neural network, a new ingredient tuple for the new product formula.
8. The system of claim 7 , wherein generating the weighted graph further comprises:
removing a subset of edges of the directed graph, the subset of edges having a weight below a first threshold.
9. The system of claim 7 , wherein each ingredient tuple identifies an ingredient, an amount of the ingredient, and a set of categorical features of the ingredient.
10. The system of claim 9 , wherein the set of embeddings include a set of formula embeddings and determining the set of embeddings further comprises:
determining a set of formula vectors based on the set of ingredient vectors, the set of formula vectors including a formula vector for each product formula of the set of product formulas.
11. The system of claim 7 , wherein the operations further comprise:
generating a new formula based on the sequence-to-sequence neural network.
12. The system of claim 11 , wherein generating the new formula further comprises:
generating a set of proposed formulas based on the sequence-to-sequence neural network;
eliminating a first subset of proposed formulas from the set of proposed formulas based on a similarity threshold to a formula of the set of product formulas; and
based on eliminating the first subset of formulas, selecting the new formula from a second subset of formulas remaining in the set of proposed formulas.
13. A computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions being executable by one or more processors to cause the one or more processors to perform operations comprising:
accessing a set of product formulas, each product formula including a set of ingredient tuples;
generating a directed graph from the set of product formulas, the directed graph including a node for each ingredient of the sets of ingredient tuples of the set of product formulas;
generating a weighted graph from the directed graph, the weighted graph having a weight assigned to each edge in the directed graph;
modifying random walk models by incorporating the weight assigned to each edge into the random walk model;
vectorizing the set of product formulas based on ingredients within the product formulas;
feeding the random walks from the random walk models into a deep learning approach to determine a set of embeddings comprising ingredient embeddings and a set of ingredient vectors from the vectorizing, wherein the deep learning approach is run on data including ingredient recommendations, and substitute ingredient recommendation;
generating, based on the weighted graph, an embedding module;
deploying the random walk models on the weighted graph;
building neural models for two or more positive formulas in the weighted graph by: extracting, using the random walk models, a subgraph with depth d; extracting n biased walks in the subgraph; and building a neural model for each entity in the weighted graph, wherein positive formulas are used to generate a weight of edges to provide the bias for the biased walks on the weighted graph, wherein positive examples are formulas that have a threshold level of market success;
training, using a bias of the biased walks from the weight of the edges, a sequence-to-sequence neural network generated using a pairwise ranking loss functions derived from the neural models;
augmenting the sequence-to-sequence neural network with the set of embeddings for the weighted graph as a sigmoid layer in the sequence-to-sequence neural network, wherein each node is represented with low-dimensional numerical formula vectors;
deploying the embedding module to generate a new product formula by passing the set of ingredient embeddings through the sequence-to-sequence neural network using one or more of the neural models; and
identifying, using the sequence-to-sequence neural network, a new ingredient tuple for the new product formula.
14. The computer program product of claim 13 , wherein generating the weighted graph further comprises:
removing a subset of edges of the directed graph, the subset of edges having a weight below a first threshold.
15. The computer program product of claim 14 , wherein the set of embeddings include a set of formula embeddings and determining the set of embeddings further comprises:
determining a set of formula vectors based on the set of ingredient vectors, the set of formula vectors including a formula vector for each product formula of the set of product formulas.
16. The computer program product of claim 13 , wherein the operations further comprise:
generating a new formula based on the sequence-to-sequence neural network.
17. The computer program product of claim 16 , wherein generating the new formula further comprises:
generating a set of proposed formulas based on the sequence-to-sequence neural network;
eliminating a first subset of proposed formulas from the set of proposed formulas based on a similarity threshold to a formula of the set of product formulas; and
based on eliminating the first subset of formulas, selecting the new formula from a second subset of formulas remaining in the set of proposed formulas.