Item recommendation method based on user intention in a conversation session

Patent metadata

Field	Value
Publication number	US-11257140-B2
Application number	US-202117325094-A
Country	US
Kind code	B2
Filing date	May 19, 2021
Priority date	May 25, 2020
Publication date	Feb 22, 2022
Grant date	Feb 22, 2022

Abstract

The present disclosure provides an item recommendation method and system based on a user intention in a conversation session. In this method, a current conversation session is encoded by use of items of prior time sequence and the last item so as to generate an accurate conversation session representation which is then used to perform initial item prediction as a user intention. Afterwards, an intention-guided neighbor detector is used to locate an accurate neighbor conversation session. Finally, the current conversation session and the neighbor conversation session representations are combined adaptively by a gated mixed layer to generate the final item recommendation, so as to effectively improve the accuracy of item recommendation.

First claim

What is claimed is:

1. An item recommendation method based on a user intention in a conversation session, configured to predict an item that a user is likely to interact at a next moment from an item set as a target item to be recommended to the user, comprising:
encoding a current conversation session S t , yielding an encoded current conversation session, wherein the current conversation session S t is a session S t ={s 1 ,s 2 , . . . ,s n } formed by n items at a time stamp t, where s i (1<i<n) refers to an i-th item at a time stamp t in the session S t , and wherein the current session S t is encoded in the following manner: embedding each item s i in the current conversation session S t to a d-dimension vector and converting the d-dimension vector into an item embedding vector x i , x i ∈Rd, x i is an i-th component of X t , X t ={x 1 ,x 2 , . . . ,x n } being an item embedding vector set formed by item embedding vectors;
obtaining a current conversation session representation by the encoded current conversation session based on a trained recommendation model, wherein obtaining the current conversation session representation by the encoded current conversation session based on the trained recommendation model comprises:
obtaining an item embedding vector by embedding each item in the current conversation session to one d-dimension vector, and taking an item embedding vector corresponding to a last item in the current conversation session as a current interest of the user in the current conversation session, wherein the current interest can be expressed in the following formula: Z t recent =x n (1);
obtaining a hidden state vector of the last item in the current conversation session according to the item embedding vector to capture a time sequence signal in the current conversation session, wherein the hidden state vector represents a long term preference of the user in the current conversation session, wherein the formula for obtaining a hidden state h i of the item s i is as follows: h i =GRU(x i ,hi −1 ) (2); and the long term preference of the user is expressed as: Z t globd =h n =GRU(x n ,h n−1 ) (3), where GRU is a gate-recurrent unit (GRU); and
obtaining the current conversation session representation by connecting the current interest and the long-term preference, wherein the current conversation session representation is expressed as: Z t crrent =W 0 [Z t globd ·Z t recent ] (4);
obtaining a user intention in the current conversation session S t according to the current conversation session representation, wherein a prediction score of each item to be recommended to the user in an item set V as a next target item in which the user is likely to be interested is calculated according to the current session representation in the following calculation formula (5): ŷ t 1 =softmax(z t currentT X) (5), wherein X is an embedding vector formed after all candidate items in the item set V are embedded into one d-dimension vector respectively to be converted into embedded items, ŷ t 1 ∈R |V| corresponds to a prediction score of each item v i ∈V, z t currentT is a transpose matrix of Z t crrent ;
obtaining a corresponding neighbor conversation session representation by identifying a neighbor conversation session of the current conversation session from a plurality of conversation sessions occurring before the current conversation session according to the user intention:
obtaining and recommending the target item to the user according to the current conversation session representation and the neighbor conversation session representation, wherein obtaining and recommending the target item to the user according to the current conversation session representation and the neighbor conversation session representation comprises:
obtaining a preference u t of the user by adaptively integrating the current conversation session representation and the neighbor conversation session representation, wherein the preference u t is expressed as: u t =f t z t current +(1−f t )z t neighbor , f t =σ(W 1 z t current +W 2 z t neighbor ), wherein σ in the formula is a sigmoid activation function, and W 1 ,W 2 ∈R d×d is a trainable parameter, z t current is the current conversation session representation and z t neighbor is the neighbor conversation session representation; and
recommending an optimization target to the user by obtaining the target item according to the preference u t , thus achieving the item recommendation method by modeling time sequence signals and latest interests of the current conversation session at a same time, thereby improving an accuracy of the item recommendation, wherein the step of the recommending the optimization target is to learn the trainable parameter of the trained recommendation model by using a cross entropy function as the optimization target.

2. The item recommendation method according to claim 1 , wherein the hidden state vector is obtained by allowing the gate recurrent unit (GRU) to act on the item embedding vector.

3. The item recommendation method according to claim 1 , wherein obtaining the user intention in the current conversation session according to the current conversation session representation comprises:
generating a prediction score of each item in the item set as the target item according to the current conversation session representation;
sorting items in the item set in a descending order according to the corresponding prediction scores, and selecting a vector formed by a plurality of items ranked top in the item set as the user intention in the current conversation session.

4. The item recommendation method according to claim 1 , wherein obtaining the corresponding neighbor conversation session representation by identifying the neighbor conversation session of the current conversation session from a plurality of conversation sessions occurring before the current conversation session according to the user intention comprises:
forming a conversation session memory by collecting conversation session representations of a plurality of conversation sessions occurring before the current conversation session and corresponding target item;
forming a to-be-identified neighbor conversation session representation by searching out each conversation session representation corresponding to the target item identical to one of a plurality of items in the user intention in the conversation session memory according to the user intention;
sorting various conversation session representations in the to-be-identified neighbor conversation session representation according to an order of the plurality of items in the user intention;
forming a candidate neighbor conversation session of the current conversation session by selecting a plurality of conversation session representations ranked top in the sorted to-be-identified neighbor conversation session representation;
calculating a similarity between each conversation session in the candidate neighbor conversation session and the current conversation session, and selecting a plurality of conversation sessions ranked top in similarity in the candidate neighbor conversation session as each neighbor conversation session of the current conversation session;
calculating a weight sum of various neighbor conversation sessions as the neighbor conversation session representation.

5. The item recommendation method according to claim 4 , wherein the conversation session memory is updated according to a first-in-first-out mechanism.

Assignees

• National Univ Of Defense Technology

Inventors

• Cai Fei
• Chen Wanyu
• Pan Zhiqiang
• Wang Yitong
• Song Chengyu
• Ling Yanxiang
• Zhang Xin
• Chen Honghui

Classifications

• G06N3/044

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

• G06N3/09

  Supervised learning · CPC title

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/08

  Learning methods · CPC title

• G06F16/9535 Primary

  Search customisation based on user profiles and personalisation · CPC title