Environmental impact aware product refurbishing

A computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computing device to cause the computing device to:

• generate, using a machine learning framework, product embeddings of a plurality of products based on a first set of digital images of the plurality of products, wherein the machine learning framework comprises an encoder component that processes each image in the first set to encode attributes of the corresponding product within a vector embedding space such that one or more new product designs are generatable via vector operations performed within the vector embedding space, and wherein the product embeddings are jointly trained with location-specific demand data and location metadata to enable attribute-based, location-specific demand forecasting for the one or more new product designs;
• perform one or more of the vector operations on the product embeddings to generate additional product embeddings within the vector embedding space, wherein the additional product embeddings represent the one or more new product designs;
• generate, by a decoder component of the machine learning framework, a second set of digital images, wherein the decoder component constructs the second set of digital images by processing the additional product embeddings, wherein the second set of digital images correspond to the one or more new product designs for refurbishing each given one of the plurality of products, and wherein the additional product embeddings are based on an initial image of the given product and one or more design constraints associated with the given product;
• calculate, by the machine learning framework, an environmental impact score and a demand impact score associated with each of the one or more new product designs, wherein the demand impact score is based at least in part on location-specific demand data, and wherein the environmental impact score is associated with at least one of the attributes of the corresponding product;
• generate, by the machine learning framework, a recommendation to refurbish at least one of the plurality of products in accordance with at least one of the one or more new product designs based on the environmental impact scores and the demand impact scores, wherein generating the recommendation comprises executing a constraint-based selection process performed in the vector embedding space, and wherein the constraint-based selection process comprises iteratively checking for design constraint violations by:
• (i) providing a candidate product embedding from the vector embedding space to the decoder component to generate a corresponding candidate digital image;
• (ii) processing the candidate digital image using a product attribution network to obtain a set of candidate image attributes, and processing the initial image of the given product using the product attribution network to obtain a set of original product attributes;
• and
• (iii) processing the original product attributes and the candidate image attributes using a rule engine that applies a set of design rules to output a penalty score representing a distance metric between the candidate digital image and the initial image, wherein the set of design rules is based at least in part on the one or more design constraints, and wherein an attribute change that violates the one or more design constraints results in a higher penalty score than an attribute change that fulfills the one or more design constraints;
• wherein the product attribution network is trained separately from the machine learning framework on a training objective directed to evaluating discrete product attributes of the candidate digital image, without consideration of the location-specific demand data or the location metadata;
• and
• output the recommendation and the generated digital image of the at least one of the one or more new product designs corresponding to the recommendation to a user.

A system comprising:

• a memory configured to store program instructions;
• a processor operatively coupled to the memory to execute the program instructions to:
• generate, using a machine learning framework, product embeddings of a plurality of products based on a first set of digital images of the plurality of products, wherein the machine learning framework comprises an encoder component that processes each image in the first set to encode attributes of the corresponding product within a vector embedding space such that one or more new product designs are generatable via vector operations performed within the vector embedding space, and wherein the product embeddings are jointly trained with location-specific demand data and location metadata to enable attribute-based, location-specific demand forecasting for the one or more new product designs;
• perform one or more of the vector operations on the product embeddings to generate additional product embeddings within the vector embedding space, wherein the additional product embeddings represent the one or more new product designs;
• generate, by a decoder component of the machine learning framework, a second set of digital images, wherein the decoder component constructs the second set of digital images by processing the additional product embeddings, wherein the second set of digital images correspond to the one or more new product designs for refurbishing each given one of the plurality of products, and wherein the additional product embeddings are based on an initial image of the given product and one or more design constraints associated with the given product;
• calculate, by the machine learning framework, an environmental impact score and a demand impact score associated with each of the one or more new product designs, wherein the demand impact score is based at least in part on location-specific demand data, and wherein the environmental impact score is associated with at least one of the attributes of the corresponding product;
• generate, by the machine learning framework, a recommendation to refurbish at least one of the plurality of products in accordance with at least one of the one or more new product designs based on the environmental impact scores and the demand impact scores, wherein generating the recommendation comprises executing a constraint-based selection process performed in the vector embedding space, and wherein the constraint-based selection process comprises iteratively checking for design constraint violations by:
• (i) providing a candidate product embedding from the vector embedding space to the decoder component to generate a corresponding candidate digital image;
• (ii) processing the candidate digital image using a product attribution network to obtain a set of candidate image attributes, and processing the initial image of the given product using the product attribution network to obtain a set of original product attributes;
• and
• (iii) processing the original product attributes and the candidate image attributes using a rule engine that applies a set of design rules to output a penalty score representing a distance metric between the candidate digital image and the initial image, wherein the set of design rules is based at least in part on the one or more design constraints, and wherein an attribute change that violates the one or more design constraints results in a higher penalty score than an attribute change that fulfills the one or more design constraints;
• wherein the product attribution network is trained separately from the machine learning framework on a training objective directed to evaluating discrete product attributes of the candidate digital image, without consideration of the location-specific demand data or the location metadata;
• and
• output the recommendation and the generated digital image of the at least one of the one or more new product designs corresponding to the recommendation to a user.