Multi-round dialogue recommendation method based on user positive and negative interest modeling

Abstract

The invention provides a multi-round dialogue recommendation method based on user positive and negative interest modeling, and belongs to the technical field of multi-round dialogue recommendation CRS in a recommendation system. The method comprises the following steps: constructing a historical interest graph, a positive interest graph and a negative interest graph; extracting multiple interests of the user through a graph neural network and comparative learning based on the historical interest graph, the positive interest graph and the negative interest graph; determining the next action through reinforcement learning based on multiple interests of the user; feeding back an evaluation action according to the user; the comparative learning comprises the steps of forming a positive sample pair by the historical interests and the positive interests, forming a negative sample pair by the positive interests and the negative interests, and enabling the interests of the user to be close to the positive sample and far away from the negative sample through comparative learning. A self-supervised training model using comparative learning is designed from the angle of negative interest of the user so as to enhance the expression of the interest of the user; therefore, the model can learn more accurate user interests.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-round dialogue recommendation (CRS) in a recommendation system, and in particular to a multi-round dialogue recommendation method based on user positive and negative interest modeling. Background Art

[0002] In today's era of information explosion, recommendation systems have become an indispensable tool to help users discover and obtain content of interest. Recommendation systems provide personalized recommendations by analyzing users' historical behaviors and interests. From e-commerce to social media to entertainment, they have been widely used and have profoundly changed the user experience. Traditional recommendation systems mainly rely on static user behavior data, and often have difficulty capturing the dynamic characteristics of user preferences, lacking real-time and personalized adaptability. To this end, conversational recommendation systems have emerged, which enable users to dynamically obtain preference information through interactive conversations, thereby better understanding users' needs and intentions and providing more accurate recommendations. Multi-round CRSs usually consist of two parts: a conversation part that interacts with users and a recommendation part that estimates user preferences. In a multi-round conversation, the agent can ask the user whether he likes a given attribute or a list of recommended items. If the upper limit of the number of conversation rounds is not reached, the user rejects the recommended item and the conversation continues. However, in actual applications, users may not have obvious preferences, so a simple yes or no answer cannot meet customer needs.

[0003] The existing technology "Multiple Choice Questions based Multi-Interest Policy Learning for Conversational Recommendation" proposes to extract multiple interests of users through historical interests and current conversation status, and select items or query attributes recommended to users based on multiple interests through a decision network. However, this method only considers the data set used for current conversation recommendations and does not have positive and negative preferences at the user attribute level; and the current preference represents the user's current interest in a certain item, which will conflict with the historical preference to a certain extent and cannot balance the long-term and short-term positive and negative preferences.

[0004] Therefore, a multi-round dialogue recommendation method is needed that combines users' negative interests and captures the correlation between positive and negative interests. Summary of the invention

[0005] In view of this, the present invention provides a multi-round dialogue recommendation method based on user positive and negative interest modeling, which solves the problem in the prior art that the negative interests of users affect the accuracy of multi-round dialogue recommendations.

[0006] To this end, the present invention provides the following technical solutions:

[0007] A multi-round dialogue recommendation method based on user positive and negative interest modeling, comprising:

[0008] Construct historical interest graphs, positive interest graphs, and negative interest graphs;

[0009] Based on historical interest graphs, positive interest graphs, and negative interest graphs, we extract users’ multiple interests through graph neural networks and contrastive learning.

[0010] Determine the next action based on the user's multiple interests through reinforcement learning;

[0011] Evaluate actions based on user feedback about the actions;

[0012] The contrastive learning includes: forming a positive sample pair from historical interests and positive interests, forming a negative sample pair from positive interests and negative interests, and making the user's interests approach the positive sample and move away from the negative sample through contrastive learning.

[0013] Furthermore, a historical interest graph is constructed through the historical interactions between users and items and the correlations between items and attributes:

[0014] G g =(N,ε)

[0015] Among them, N=u∪v∪p, ε=ε u，v ∪ε p，v ;

[0016] Extract the relevance of users’ long-term interests and global interest items and attribute instances through LightGCN:

[0017]

[0018] Aggregate the representations of all layers to obtain the history graph representation e h :

[0019]

[0020] Furthermore, constructing the negative interest graph includes:

[0021]

[0022] Where N represents nodes, ε represents edges; for the weight of each edge The edge weight between the user and each rejected attribute is 1;

[0023] The weight of the edge between the rejected attribute and each item including the rejected attribute is:

[0024]

[0025] where σ(·) is the activation function, u, v, and p represent users, items, and attributes, respectively;

[0026] LightGCN captures the connectivity between graphs and nodes, and obtains higher quality node representations in the current state:

[0027]

[0028] Aggregate the representations of all layers to obtain the negative interest graph representation e n :

[0029]

[0030] Furthermore, constructing the positive interest graph includes:

[0031]

[0032]

[0033] Through L-layer LightGCN, the connection relationship between graph nodes is captured and higher-quality node representation is obtained in the current state:

[0034]

[0035] Aggregate the representations of all layers to obtain the positive interest graph representation e p :

[0036]

[0037] Furthermore, the step of determining the next action based on the user's multiple interests through reinforcement learning includes:

[0038] When the agent decides to ask, it chooses the attribute with the highest entropy value to ask;

[0039] When the agent decides to recommend, it selects the top K items for recommendation.

[0040] Further, evaluating the action according to the user's feedback on the action includes:

[0041] When the agent chooses to ask for an attribute, if the user accepts the recommended attribute, the attribute is defined as: Otherwise define the reject attribute as The candidate attributes are updated as:

[0042]

[0043]

[0044]

[0045] in, Represents candidate attributes, Indicates the set of attributes that the user has accepted. Indicates the set of attributes that the user has rejected;

[0046] When the agent selects a recommended item, if the user rejects the recommended item, the rejected item is defined as The candidate projects are updated to:

[0047]

[0048]

[0049] If the user accepts the recommendation, the recommendation dialog ends successfully.

[0050] Furthermore, the contrastive learning loss function is:

[0051]

[0052] Where s(·) is the cosine similarity function, τ is the temperature coefficient, and e h 、e p and e n They are historical interest representation, positive interest representation and negative interest representation respectively.

[0053] Advantages and positive effects of the present invention:

[0054] The present invention constructs a negative sample pair of negative interests and positive interests for comparative learning, and enhances the user's positive interest by capturing the user's negative interest, thereby improving the accuracy of the model in extracting user interests, so that the model can learn more accurate user interests. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0056] Figure 1 This is a structural diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0058] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0059] The present invention provides a multi-round dialogue recommendation method based on user positive and negative interest modeling, combining Figure 1 As shown in the structural diagram, the present invention constructs the user's negative interests separately; and compares and studies the negative interests, positive interests, and historical interests to obtain more accurate user interests, thereby optimizing the recommendation experience for the user. In the multi-round dialogue recommendation process, the intelligent agent recommends attributes or items to the user, and the user not only likes the content but also dislikes the content. Therefore, each round distinguishes the user's positive and negative preferences, and uses the negative preferences to enhance the positive preferences so that more accurate choices can be made in the next round, thereby capturing the user's accurate interests in a shorter dialogue. On the other hand, comparative learning is an unsupervised learning that constructs the user's positive interests and historical interests into a positive sample pair, and constructs the positive interests and negative interests into a negative sample pair. It can effectively learn the user's current real interests.

[0060] The process of the present invention is based on a multi-interest multi-round conversational recommendation scenario (MIMCR), where the agent has multiple attribute instances or items for the user to choose in each round t. Define attribute set instances P, V and attribute types C, and each item v∈V is associated with a set of attribute instances Pv. Each attribute instance p∈P has a corresponding attribute type cp∈C.

[0061] S1. Build user preference model:

[0062] For state s tEncodes the information of all the following conversation rounds. The state consists of seven parts:

[0063]

[0064] Among them, u represents the user, Represents the attributes that the user has accepted. Indicates the attribute rejected by the user, Represents an item that the user has rejected, represents the candidate attribute, Indicates candidate items, Indicates the item to which the attribute is associated.

[0065] S2, after the system completes the user's initial preference modeling; according to the state s t , the agent selects an action a from the action space a t The purpose is to decide whether to ask for attributes or recommend items in this round. The user learning strategy is adopted based on the reinforcement learning method of multi-interest extraction, and the Q-value of each action is calculated using the Q-network.

[0066] When the agent decides to ask, it chooses the attribute with the highest entropy value to ask;

[0067] When the agent decides to recommend, it selects the top K items for recommendation.

[0068] When the agent chooses an action, the state s t Will transition to the next state s t+1 .

[0069] When the agent chooses to ask for an attribute, if the user accepts the recommended attribute, the attribute is defined as: Otherwise define the reject attribute as And update the candidate attributes:

[0070]

[0071]

[0072]

[0073] in, Represents candidate attributes, Indicates the set of attributes that the user has accepted. Represents a collection of attributes that the user has rejected.

[0074] When the agent selects a recommended item, if the user rejects the recommended item, the rejected item is defined as And update the candidates:

[0075]

[0076]

[0077] If the user accepts the recommendation, the recommendation dialog ends successfully.

[0078] S3. Model building rewards. In this embodiment, the rewards include:

[0079] r ask_suc , when the user accepts the attribute being asked, a slight reward is given;

[0080] r ask_fail , a slight penalty is imposed when the user rejects the requested attribute;

[0081] r rec_succ , giving a strong reward when the user accepts the recommended item;

[0082] r rec_fail , giving a slight penalty when a user rejects a recommended item;

[0083] r cand_none , when there are no projects in the candidate pool, a slight penalty is given;

[0084] r quit When a session reaches the maximum number of rounds, a strong penalty is imposed.

[0085] The final reward for multi-round dialogue recommendation is:

[0086]

[0087] S4, constructing history graph representation, positive interest graph representation and negative interest graph representation;

[0088] Existing MCR methods focus more on attributes accepted by users and ignore the negative preferences represented by attributes rejected by users.

[0089] S41. Use historical interactions between users and items and correlations between items and attribute instances to build a heterogeneous global graph:

[0090] G g =(N,ε)

[0091] Among them, N=u∪v∪p, ε=ε u，v ∪ε p，v ;

[0092] The Lightweight Graph Neural Network (LightGCN) is used to extract the relevance of users’ long-term interests and global interest items and attribute instances:

[0093]

[0094] S42. Construct a negative interest graph to represent the user's negative preference in the tth session:

[0095]

[0096]

[0097] Where N represents nodes, ε represents edges; for the weight of each edge The edge weight between the user and each rejected attribute is 1;

[0098] The weight of the edge between the rejected attribute and each item including the rejected attribute is:

[0099]

[0100] where σ(·) is the activation function, and u, v, and p represent users, items, and attributes, respectively.

[0101] LightGCN captures the connectivity between graphs and nodes, and obtains higher quality node representations in the current state:

[0102]

[0103] S43, constructing positive interest graph:

[0104] A positive interest graph is constructed based on the user's acceptance attributes at the node in the tth round Its towel Similarly, the weight of each edge is calculated in the same way as the negative interest graph. At the same time, the connection relationship between the graph nodes is captured through the L-layer LightGCN to obtain a higher quality node representation in the current state:

[0105]

[0106] S44, respectively, by aggregating the representations of all layers, we get the historical graph representation e h , positive interest graph representation e p and negative interest graph representation e n :

[0107]

[0108] S5. Contrastive learning: In order to fully integrate the user's negative interests, positive interests, and historical interests, the current embedding can express the user's interests as accurately as possible. Therefore, the historical interests and positive interests are combined to form a positive sample pair, and the positive interests and negative interests are combined to form a negative sample pair. Contrastive learning is a self-supervised learning method that can effectively integrate the different interests of users. Make the current positive interests more similar to the historical long-term interests, and expand the difference between negative interests and positive interests. In order to better capture the relevance of user interests and fully integrate the different interests of users, the loss function of contrastive learning is designed based on InfoNCE:

[0109]

[0110] Where s(·) is the cosine similarity function, τ is the temperature coefficient, and e h 、e p and e n They are historical interest representation, positive interest representation of the current conversation round, and negative interest representation.

[0111] In order to evaluate the proposed method, two standard datasets, Yelp and LastFM, which are commonly used in existing multi-round conversations, were used for experiments to evaluate recommendation tasks. According to previous conversation recommendation research, SR@T is used to evaluate the rate of successful recommendations in T-round conversations, and the average number of rounds (AT) is used to evaluate the number of conversation rounds required to successfully complete the recommendation. In addition, hDCG@(T, K), a recommendation system evaluation indicator with hierarchical data, is used to evaluate the model. The improvements of the SR@T, AT and hDCG indicators in the above-mentioned Yelp dataset are 1.82%, 2.78% and 7.5% respectively.

[0112] The present invention designs a training model using contrastive learning self-supervision from the perspective of user negative interests to enhance the representation of user interests; it not only pays attention to the user's dynamically changing positive interests during the conversation, but also pays attention to the user's dynamically changing negative interests.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-round dialogue recommendation method based on user positive and negative interest modeling, characterized in that: include: Construct historical interest graphs, positive interest graphs, and negative interest graphs; Based on historical interest graphs, positive interest graphs, and negative interest graphs, we extract users’ multiple interests through graph neural networks and contrastive learning. Determine the next action based on the user's multiple interests through reinforcement learning; Evaluate actions based on user feedback about the actions; The contrastive learning includes: forming a positive sample pair with historical interests and positive interests, forming a negative sample pair with positive interests and negative interests, and making the user's interests approach the positive sample and move away from the negative sample through contrastive learning.

2. According to claim 1, a multi-round dialogue recommendation method based on user positive and negative interest modeling is characterized in that: Construct a historical interest graph through historical interactions between users and items and correlations between items and attributes: G g =(N,e) Where, N=u∪v∪p, ε=ε u,v ∪e p,v ; Extract the relevance of users’ long-term interests and global interest items and attribute instances through LightGCN: Aggregate the representations of all layers to obtain the history graph representation e h :

3. According to claim 1, a multi-round dialogue recommendation method based on user positive and negative interest modeling is characterized in that: The constructing of the negative interest graph comprises: Where N represents nodes, ε represents edges; for the weight of each edge The edge weight between the user and each rejected attribute is 1; The weight of the edge between the rejected attribute and each item including the rejected attribute is: where σ(·) is the activation function, u, v, and p represent users, items, and attributes, respectively; LightGCN captures the connectivity between graphs and nodes, and obtains higher quality node representations in the current state: Aggregate the representations of all layers to obtain the negative interest graph representation e n :

4. According to claim 1, a multi-round dialogue recommendation method based on user positive and negative interest modeling is characterized in that: The constructing of the positive interest graph comprises: Through L-layer LightGCN, the connection relationship between graph nodes is captured and higher-quality node representation is obtained in the current state: Aggregate the representations of all layers to obtain the positive interest graph representation e p :

5. According to claim 1, a multi-round dialogue recommendation method based on user positive and negative interest modeling is characterized in that: The method of determining the next action based on the user's multiple interests through reinforcement learning includes: When the agent decides to ask, it chooses the attribute with the highest entropy value to ask; When the agent decides to recommend, it selects the top K items for recommendation.

6. According to claim 1, a multi-round dialogue recommendation method based on user positive and negative interest modeling is characterized in that: The evaluating the action according to the user's feedback on the action includes: When the agent chooses to ask for an attribute, if the user accepts the recommended attribute, the attribute is defined as: Otherwise define the reject attribute as The candidate attributes are updated as: in, Represents candidate attributes, Indicates the set of attributes that the user has accepted. Indicates the set of attributes that the user has rejected; When the agent selects a recommended item, if the user rejects the recommended item, the rejected item is defined as The candidate projects are updated to: If the user accepts the recommendation, the recommendation dialog ends successfully.

7. According to claim 1, a multi-round dialogue recommendation method based on user positive and negative interest modeling is characterized in that: The contrastive learning loss function: Where s(·) is the cosine similarity function, τ is the temperature coefficient, e h 、e p and e n They are historical interest representation, positive interest representation and negative interest representation respectively.

Images