Asymmetrically hierarchical networks with attentive interactions for interpretable review-based recommendation

Patent metadata

Field	Value
Publication number	US-11521255-B2
Application number	US-202016995052-A
Country	US
Kind code	B2
Filing date	Aug 17, 2020
Priority date	Aug 27, 2019
Publication date	Dec 6, 2022
Grant date	Dec 6, 2022

Abstract

A method for implementing a recommendation system using an asymmetrically hierarchical network includes, for a user and an item corresponding to a user-item pair, aggregating, using asymmetrically designed sentence aggregators, respective ones of a set of item sentence embeddings and a set of user sentence embeddings to generate a set of item review embeddings based on first item attention weights and a set of user review embeddings based on first user attention weights, respectively, aggregating, using asymmetrically designed review aggregators, respective ones of the set of item review embeddings and the set of user review embeddings to generate an item embedding based on a second item attention weights and a user embedding based on second user attention weights, respectively, and predicting a rating of the user-item pair based on the item embedding and the user embedding.

First claim

What is claimed is:

1. A computer-implemented method for implementing a recommendation system using an asymmetrically hierarchical network, comprising:
encoding sentences, using at least one hardware processor operatively coupled to a non-transitory computer-readable storage medium and a bidirectional recurrent neural network (BiRNN), of a set of user historical reviews associated with a user and a set of item historical reviews associated with an item to generate a set of user sentence embeddings and a set of item sentence embeddings, respectively, the user and the item corresponding to a user-item pair, the encoding representing at least one sentence by a sequence of word vectors and comprising learning a vector embedding for the at least one sentence based on the sequence of word vectors by training the BiRNN to learn the vector embedding by max-pooling hidden states of the BiRNN on the at least one sequence of word vectors;
aggregating, using the at least one hardware processor, asymmetrically designed sentence aggregators, respective ones of the set of item sentence embeddings and the set of user sentence embeddings to generate a set of item review embeddings based on first item attention weights and a set of user review embeddings based on first user attention weights, respectively;
aggregating, using the at least one hardware processor and asymmetrically designed review aggregators, respective ones of the set of item review embeddings and the set of user review embeddings to generate an item embedding based on a second item attention weights and a user embedding based on second user attention weights, respectively; and
predicting, using the at least one hardware processor, a rating of the user-item pair based on the item embedding and the user embedding.

2. The method as recited in claim 1 , wherein encoding the sentence of the set of user historical reviews and the set of item historical review further includes:
encoding context information for the at least one sentence based on the vector embedding.

3. The method as recited in claim 1 , wherein aggregating the set of item sentence embeddings and the set of item review embeddings further includes using respective gating mechanisms during calculation of the first item attention weights and the second item attention weights to improve model performance.

4. The method as recited in claim 1 , wherein:
aggregating the set of user sentence embeddings further includes:
learning and normalizing a sentence affinity matrix between user sentences and item sentences; and
obtaining the first user attention weights based on the sentence affinity matrix and the first item attention weights; and
aggregating the set of user review embeddings further includes:
learning and normalizing a review affinity matrix between user sentences and item sentences; and
obtaining the second user attention weights based on the review affinity matrix and the second item attention weights.

5. The method as recited in claim 4 , wherein:
obtaining the first user attention weights further includes performing row-wise max pooling on the sentence affinity matrix and the first item attention weight for obtaining maximum affinity; and
obtaining the second user attention weights further includes performing row-wise max pooling on the review affinity matrix and the second item attention weight for obtaining maximum affinity.

6. The method as recited in claim 5 , wherein:
obtaining the first user attention weights further includes calculating the first user attention weights as α i u =softmax(max row (G i ⊙ row α v )), where α i u corresponds to attention weights in a user review matrix including n entries for each i ∈[1,n], α v corresponds to a concatenation of the first item attention weights, max row refers to row-wise max-pooling for obtaining the maximum affinity and ⊙ row refers to the Hadamard product between each row; and
obtaining the second user attention weights further includes calculating the second user attention weights as β u =softmax(max row (G⊙ row β v )), where β u corresponds to the second user attention weights, and β v corresponds to a concatenation of the second item attention weights.

7. The method as recited in claim 1 , wherein predicting the rating of the user-item pair further includes:
receiving a concatenated vector of the item embedding and the user embedding; and
generating a predicted rating based on the concatenated vector; and
calculating an error between a real rating and the predicted rating as a loss function for model training.

8. A computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method for implementing a recommendation system using an asymmetrically hierarchical network, the method performed by the computer comprising:
encoding sentences, using a bidirectional recurrent neural network (BiRNN), of a set of user historical reviews associated with a user and a set of item historical reviews associated with an item to generate a set of user sentence embeddings and a set of item sentence embeddings, respectively, the user and the item corresponding to a user-item pair, the encoding representing at least one sentence by a sequence of word vectors and comprising learning a vector embedding for the at least one sentence based on the sequence of word vectors by training the BiRNN to learn the vector embedding by max-pooling hidden states of the BiRNN on the at least one sequence of word vectors;
aggregating, using asymmetrically designed sentence aggregators, respective ones of the set of item sentence embeddings and the set of user sentence embeddings to generate a set of item review embeddings based on first item attention weights and a set of user review embeddings based on first user attention weights, respectively;
aggregating, using asymmetrically designed review aggregators, respective ones of the set of item review embeddings and the set of user review embeddings to generate an item embedding based on a second item attention weights and a user embedding based on second user attention weights, respectively; and
predicting a rating of the user-item pair based on the item embedding and the user embedding.

9. The computer program product as recited in claim 8 , wherein encoding the sentence of the set of user historical reviews and the set of item historical review further includes:
encoding context information for the at least one sentence based on the vector embedding.

10. The computer program product as recited in claim 8 , wherein aggregating the set of item sentence embeddings and the set of item review embeddings further includes using respective gating mechanisms during calculation of the first item attention weights and the second item attention weights to improve model performance.

11. The computer program product as recited in claim 8 , wherein:
aggregating the set of user sentence embeddings further includes:
learning and normalizing a sentence affinity matrix between user sentences and item sentences; and
obtaining the first user attention weights based on the sentence affinity matrix and the first item attention weights; and

aggregating the set of user review embeddings further includes:
learning and normalizing a review affinity matrix between user sentences and item sentences; and
obtaining the second user attention weights based on the review affinity matrix and the second item attention weights.

12. The computer program product as recited in claim 11 , wherein:
obtaining the first user attention w

Assignees

• Nec Lab America Inc
• Nec Corp

Inventors

• Ni Jingchao
• Chen Haifeng
• Zong Bo
• Dong Xin
• Cheng Wei

Classifications

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N20/00

  Machine learning · CPC title

• G06N3/08

  Learning methods · CPC title

• G06F40/30

  Semantic analysis · CPC title

• G06Q30/0631 Primary

  Recommending goods or services · CPC title