CN110245299B - Sequence recommendation method and system based on dynamic interaction attention mechanism - Google Patents

Sequence recommendation method and system based on dynamic interaction attention mechanism Download PDF

Info

Publication number
CN110245299B
CN110245299B CN201910533753.1A CN201910533753A CN110245299B CN 110245299 B CN110245299 B CN 110245299B CN 201910533753 A CN201910533753 A CN 201910533753A CN 110245299 B CN110245299 B CN 110245299B
Authority
CN
China
Prior art keywords
term
user
short
long
preference
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201910533753.1A
Other languages
Chinese (zh)
Other versions
CN110245299A (en
Inventor
蔡飞
陈洪辉
刘俊先
罗爱民
舒振
陈涛
罗雪山
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
National University of Defense Technology
Original Assignee
National University of Defense Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by National University of Defense Technology filed Critical National University of Defense Technology
Priority to CN201910533753.1A priority Critical patent/CN110245299B/en
Publication of CN110245299A publication Critical patent/CN110245299A/en
Application granted granted Critical
Publication of CN110245299B publication Critical patent/CN110245299B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
    • G06F16/2474—Sequence data queries, e.g. querying versioned data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/06—Buying, selling or leasing transactions
    • G06Q30/0601—Electronic shopping [e-shopping]
    • G06Q30/0631—Item recommendations

Abstract

The invention provides a sequence recommendation method based on a dynamic interaction attention mechanism, which comprises the steps of obtaining initial short-term preference and initial long-term preference of a user; obtaining a long-term preference and a short-term preference according to the interactive attention network in combination with the initial short-term preference and the initial long-term preference; and according to the sequence recommendation model, the long-term preference and the short-term preference are combined to score the corresponding articles, and the articles are recommended to the user according to the scoring result. The dynamic interaction attention mechanism network model for sequential recommendation (DCN-SR) constructed by the invention can learn the common dependency representation of long-term and short-term interaction of the user through the model, and combines long-term preference and short-term preference to enable the recommendation result to be more accurate.

Description

Sequence recommendation method and system based on dynamic interaction attention mechanism
Technical Field
The invention belongs to the field of sequence recommendation, and particularly relates to a sequence recommendation method and a sequence recommendation system based on a dynamic interaction attention mechanism.
Background
Recommendation systems are an effective solution to help people deal with increasingly complex information environments. Conventional recommendation systems typically ignore sequential information and focus on mining static correlations between users and items from interactions. For example, a typical conventional recommendation system based on matrix factorization may learn from the entire interaction history to effectively model the general preferences of a user, but it does not model the user's short-term, sequential interaction behavior. Unlike the conventional recommendation system, the sequential recommendation system predicts the next object that may be interested by the user according to the interaction history of the user over a period of time.
The modeling method of the existing sequential recommendation system mainly comprises a Markov chain and a Recurrent Neural Network (RNN). For example, decomposing a personalized Markov chain (FPMC) model combines Markov chains with matrix decomposition to achieve good recommendation performance. Also included in the prior art is a Hierarchical Representation Model (HRM) model that extends the idea of FPMC by building representations of users and items using a two-layer structure. The markov chain based approach can only model the local sequence pattern between every two adjacent interactions. RNN-based models can effectively model multi-step sequential behavior. The Hierarchical Recurrent Neural Network (HRNN) model and the dynamic bAsket model (DREAM) model combine the long and short term tastes of users. Significant improvement was achieved compared to both HRM and FPMC. However, these methods do not capture the change in relative importance of the user's long-term preferences versus short-term preferences.
Disclosure of Invention
The invention aims to provide a sequence recommendation method and a system thereof based on a dynamic interaction attention mechanism, so as to solve the technical problems in the prior art.
The invention discloses a sequence recommendation method based on a dynamic interactive attention mechanism, which comprises the following steps:
acquiring initial short-term preference and initial long-term preference of a user;
obtaining a long-term preference and a short-term preference according to the interactive attention network in combination with the initial short-term preference and the initial long-term preference;
and according to the sequence recommendation model, the long-term preference and the short-term preference are combined to score the corresponding articles, and the articles are recommended to the user according to the scoring result.
Preferably, after the corresponding article is scored, the scoring result needs to be corrected through a loss function, where the loss function is:
Figure GDA0003308475580000011
Figure GDA0003308475580000012
indicating the degree of preference, y, of the predicted user u for item iuiRepresenting the real result and V representing the number of all items.
Preferably, the way to obtain the initial short-term preference of the user is through a context-based GRU model: the CGRU performs initial short-term preference queries, the CGRU being:
Figure GDA0003308475580000021
zt＝σ(Wzxt+Vzat+Uzht-1)
Figure GDA0003308475580000022
rt＝σ(Wrxt+Vrat+Urht-1)。
Wz，Vz，Uz，Uz，Wr，Vr，Urw, V, U are model parameters, which can be obtained by training, xtVector representation representing items input at time t, atThe final calculated h is the vector representation of the behavior actiontσ represents a sigmoid function and tan h represents a tangent function, which are hidden layer state representations at time.
Preferably, the method for obtaining the initial long-term preference of the user is to use a multi-layer perceptron to model the long-term preference of the user, and the modeling process of the multi-layer perceptron is as follows:
Figure GDA0003308475580000023
Wmand bmParameters representing the m-th layer perceptron, phi the activation function, XiDenotes xiThe final output state vector.
Preferably, the interactive attention network is:
Figure GDA0003308475580000024
Hl＝tan h[WlUl+(WsUs+Wths,T)CT]
Figure GDA0003308475580000025
Hs＝tan h[WsUs+Wths,T+(WlUl)C]
Figure GDA0003308475580000026
Uland UsIs the user's initial long and short term state vector representation, Wc,Wl,Ws,Wt,Whl,WhsAs model parameters, can be obtained by training, hs,TRepresenting the last hidden layer state vector of the CGRU output, a of the final outputlAnd aSA weight of attention that is a long-short term preference of the user.
Preferably, the short-term preference and the long-term preference of the user are respectively:
Figure GDA0003308475580000027
Figure GDA0003308475580000028
Uco-sand Uco-lIndicating passing of interactive attentionAnd (4) making the final long-term and short-term preference representation of the user obtained by calculation.
Preferably, the manner of scoring the pairs of articles is:
Figure GDA0003308475580000031
Figure GDA0003308475580000032
Figure GDA0003308475580000033
Figure GDA0003308475580000034
Bl，Bs，BTall model parameters can be obtained through training and finally output
Figure GDA0003308475580000035
Indicating the degree of preference of user u for all items.
The invention also provides a sequence recommendation system based on the dynamic interaction attention mechanism, which is based on the method, and comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor implements the steps of any one of the methods when executing the computer program.
The beneficial effects of the invention include:
1. the invention constructs a dynamic interaction attention mechanism network model for sequential recommendation (DCN-SR), and the model can learn the common dependency representation of long-term and short-term interaction of a user and combine the relationship between the long term and the short term to ensure that the recommendation result is more accurate.
2. The context-gated cyclic unit CGRU is constructed, contains different types of short-term user action behaviors, and can better estimate the next interest preference of the user.
Drawings
FIG. 1 is a frame diagram of a sequence recommendation method based on a dynamic interaction attention mechanism according to the present invention;
FIG. 2 is a CGRU framework diagram of the present invention;
FIG. 3 is a schematic diagram of an interactive attention network of the present invention;
FIG. 4 is a comparison graph of the recommendation effect of different models at different session lengths according to the preferred embodiment of the present invention;
FIG. 5 is a diagram illustrating a comparison of recommendation effects of different models for different lengths of history information according to an embodiment of the present invention;
fig. 6 is a schematic diagram illustrating the influence of the session importance on the recommendation effect in the preferred embodiment of the present invention.
Detailed Description
Example 1:
the sequence recommendation is to predict the next object possibly interested by the user according to the interaction history of the user in a period of time. A reasonably efficient sequence recommendation model should be able to efficiently combine the long-term and short-term preferences of the user. The existing method always regards the long-term interest preference of the user as a static fixed vector, and the long-term preference of the user is not changed when different dynamic preferences of the user are processed. But this is not reasonable because in the face of different dynamic preferences of the user, their long-term preferences should have different importance. The relative importance of events in a user's long-term interaction history depends on the events in their short-term interaction history, and vice versa. If a user has searched for a camera in the current session, the long-term interaction with the electronic product should be more important than her clothing-related interaction in the user's long-term search history when deciding what to recommend next. Conversely, if the user's past interactions indicate a greater interest in the Sony brand, then during the current session, interactions related to that brand may be more important than other interactions in predicting the next recommended item. At the same time, different user actions (e.g., clicking, adding to a shopping cart, or purchasing) provide different types of information about the user's interests. Clicking on a camera may indicate that the current recommendation is unsatisfactory and thus may proceed to recommend another camera, whereas if a user purchases a camera, we should later recommend some camera related products instead of the camera, e.g. memory cards, etc. Based on this consideration, we propose a recommendation model DCN-SR based on a dynamic interaction attention mechanism, which employs an interactive attention network to model the interaction between the user's long-term and short-term preferences, and includes the following three parts:
(1) short-term preferences of the user are modeled. The present invention uses a context-gated cycle unit (CGRU) network to incorporate information contained in user operations. The short-term preference is expressed as a combination of the interaction states hidden in the current session.
(2) A multi-layered perceptron (MLP) is used to process the user's historical interactions and infer its general preferences.
(3) Using the outputs of the first two parts, we apply an interaction attention mechanism network to capture the interaction between the user's behaviors in the long-term and short-term interaction histories and generate a co-dependent representation of their long-term and short-term preferences. We compute a recommendation score for each candidate based on these co-dependent representations.
Referring to fig. 1, fig. 1 is a block diagram of the method of the present invention, comprising the three components described above: short term query preferences, long term query preferences, interactive attention networks.
Referring to fig. 2, the present invention considers different behaviors of users and the time-sequence characteristics of the behaviors, and therefore proposes a context-based GRU model: CGRU. Namely, the behavior information is added to the input gate, the forgetting gate and the updating gate of the GRU, and the CGRU is expressed as follows:
Figure GDA0003308475580000041
zt＝σ(Wzxt+Vzat+Uzht-1)
Figure GDA0003308475580000042
rt＝σ(Wrxt+Vrat+Urht-1)
Wz，Vz，Uz，Uz，Wr，Vr，Urw, V, U are model parameters, which can be obtained by training, xtVector representation representing items input at time t, atThe final calculated h is the vector representation of the behavior actiontIs a hidden layer state representation of the time of day. σ denotes a sigmoid function, and tan h denotes a tangent function.
This may result in a hidden layer representation for each interaction representing the user's initial short-term preference with their set.
The long-term preferences of the user are modeled using a multi-tier perceptron, where we use only the items that the user has collected and purchased, as these behaviors are more representative of the user's preferences. The modeling process of the multilayer perceptron is as follows:
Figure GDA0003308475580000043
Figure GDA0003308475580000051
Wmand bmRepresenting the parameters of the m-th layer perceptron and phi the activation function. XiDenotes xiThe final output state vector.
The multilayer perceptron modeling is adopted mainly because the multilayer perceptron modeling has good nonlinear modeling capability and is widely applied to a collaborative filtering method.
It is beneficial to combine the short-term and long-term preferences of the user when making recommendations. However, conventional approaches treat these two types of preferences as independent, ignoring the (potential) interdependencies between them. Furthermore, conventional attention mechanisms assign weights to events in the user's history and recent interactions, respectively. Historical interactions and recent interactions should provide context for each other in calculating the importance of each event. Therefore, the present invention designs an interactive attention network to model the long and short term interest preferences of users. As shown in fig. 3. The interactive attention network is:
Figure GDA0003308475580000052
Hl＝tan h[WlUl+(WsUs+Wths,T)CT]
Figure GDA0003308475580000053
Hs＝tan h[WsUs+Wths,T+(WlUl)C]
Figure GDA0003308475580000054
Uland UsIs the user's initial long and short term state vector representation, Wc,Wl,Ws,Wt,Whl,WhsAs model parameters, can be obtained by training, hs,TRepresenting the last hidden layer state vector output by the CGRU. A of final outputlAnd asA weight of attention that is a long-short term preference of the user.
And finally, respectively obtaining the short-term preference and the long-term preference of the user as follows:
Figure GDA0003308475580000055
Figure GDA0003308475580000056
Uco-sand Uco-lThe final user long-short term preference representation obtained through the calculation of the interactive attention mechanism is represented.
In the prediction stage, the long-term and short-term interest preferences of the user are combined to score the corresponding articles:
Figure GDA0003308475580000057
Figure GDA0003308475580000058
Figure GDA0003308475580000059
Figure GDA00033084755800000510
Bl，Bs，BTall model parameters can be obtained through training and finally output
Figure GDA00033084755800000511
Indicating the degree of preference of user u for all items.
Wherein the loss function at training is as follows:
Figure GDA00033084755800000512
Figure GDA00033084755800000513
indicating the degree of preference, y, of the predicted user u for item iuiRepresenting the real result and V representing the number of all items.
Comparing the DCN-SR model provided by the invention with GRURec and NARM models, the DCN-SR model can be well popularized to the GRURec and NARM models.
GRU4Rec represents the user's interest preference with the last hidden layer state of the GRU, namely:
hT＝GRUsess(vT,hT-1)。
GRUsessdenotes a GRU calculation unit, vTRepresenting the input state vector at time t.
When the DCN-SR model does not consider the user's long-term retrieval history and behavior information, the DDCN-SR model can simplify:
Figure GDA0003308475580000061
hs,Trepresenting the last hidden lamina vector of the CGRU. v. ofiRepresenting the ith item State vector representation
For NARM, it uses the attention mechanism and combines the attention mechanism-based vector and the last GRU hidden layer vector to form the user's interest preference representation:
Figure GDA0003308475580000062
aiindicating the attention weight of the ith item.
When the DCN-SR model does not consider the long-term history of the user and simultaneously sets reasonable parameters and activation functions, the following results can be obtained:
Figure GDA0003308475580000063
therefore, the DCN-SR model is a relatively general model.
In conclusion, the invention designs a dynamic interaction attention mechanism network model for sequential recommendation (DCN-SR), which can learn the co-dependent representation of the long-term and short-term interactions of a user and utilize the relationship between her long-term and short-term; the present invention designs a context-gated loop unit CGRU to incorporate different types of short-term user action behavior in order to better estimate the user's next interest preferences.
Example 2:
in this embodiment, a heaven and cat electronic commerce data set and a heaven and battery electronic commerce data set are selected, and specific information of the data set is shown in table 1 below:
TABLE 1
Figure GDA0003308475580000064
Figure GDA0003308475580000071
The following models were selected as comparison targets in this experiment: item-pop is a method of sorting items according to their number of interactions and then recommending according to the sort, which is a non-personalized method. FPMC is a recommendation method based on markov chains and collaborative filtering. GRU4Rec, a recommendation model based on RNN and query sessions, is a non-personalized approach. The HRNN is used for a personalized recommendation method based on query sessions, adopts a layered RNN-structure, namely a session-level RNN and a user-level RNN, and is used for simulating short-term and long-term preferences of a user. NARM is based on a model of RNN, applying an attention mechanism to capture user preference information, but it is also a non-personalized approach. STAMP is a query recommendation method that combines a memory network and an attention mechanism.
The overall effect of these several reference models and the DDCN-SR model proposed by the present invention is shown in table 2 below:
table 2:
Figure GDA0003308475580000072
as can be seen from Table 2, the RNN-based methods, i.e., Item-pop and FPMC, are superior to the conventional methods. At the same time, HRNN's results are only higher than RNN-based methods, such as GRU4Rec, which means that combining the user's history with recent interactions can help improve recommendation performance. In addition NARM and STAMP also showed improvement on HRNN. STAMP achieves better performance than others. We therefore used STAMP as the best benchmark for comparison in later experiments. Next, comparison was made with the DCN-SR model. In Recall @10, both NARM and STAMP are weak DCN-SR. This indicates that applying the dynamic interactive attention network helps to improve the recommendation effect. This is because the dynamic interaction attention network can capture the relationship between the user's history and recent interactions. Improvement of DCN-SR over the best reference model: in Recall @10, the data set for the skatecat was 2.58% and the data set for the skull was 3.08%. The improvement in the day cat dataset at MRR @10 was 3.78% and the improvement in the day pool dataset at MRR @10 was 4.05%.
To demonstrate the utility of the CGRU network, the different behaviors of the user were treated as search Contextual meetings in a short time, and the recommended performance of DCN-SR, DCN-SRGRU (with simple GRU network) working and DCN-SR CGRU (with context GRU network), was examined under different settings. Table 3 compares its performance to the best baseline model (STAMP):
TABLE 3
Figure GDA0003308475580000081
Experimental results show that the model of the invention has better recommendation effect all the time. For different numbers of recommended items, it can be seen that the overall performance of Recall and MRR improves, with the size of the recommendation list ranging from 5 to 15.
The short-term conversation of the user is divided into short, medium and long according to different lengths, and then the recommendation effects of all models are compared. The comparison results are shown in FIGS. 4(a) to 4 (d). From fig. 4 we can see that the performance of all models improves on the tianmao dataset as the session length increases. The DCN-SR model always achieves the highest score for different session lengths on the two data sets. Specifically, for Recall @10, STAMP is superior to NARM, except for short-term sessions; this may be because short-term sessions contain less information than long sessions, and therefore location and order information may be provided based on the RNN model, i.e., NARM. STAMP lacks such position and sequence information.
The user history information is divided into eight types according to different lengths, and then recommendation effects of all models are compared. The comparison results are shown in FIGS. 5(a) to 5 (d). From fig. 5 the best performance can be achieved with DCN-SR, all eight groups in both datasets. For the day cat dataset, as the user's historical number of interactions increased, the performance of all models began to fluctuate first, but showed an overall upward trend.
For the sky-pool dataset, the performance of all models will degrade, in terms of two metrics. However, DCN-SR and HRNN descent speeds were slower than STAMP and NARM models, consistent with our findings shown in FIG. 5(a) and FIG. 5 (b). These results demonstrate that the DCN-SR model can improve the recommendation effect in combination with the advantages of long-term and short-term interest preferences.
To illustrate the role of the common focus mechanism, the present embodiment presents an example of two users. We randomly chosen two sessions from the test set of the day pool dataset because the day pool dataset contains category information of the items, which helps us to evaluate the recommendation effect to some extent. In fig. 6, the depth of the color indicates the importance of the event, and the darker the color, the more important the event. The number above the column is the category of the corresponding item. As shown in fig. 6, first, although two sessions of a single user share the same historical interactions, the historical interactions are weighted differently. In addition, some items have greater attention in the same category as the target item. Secondly, the interactions in the session also have different weights for predicting the user's preferences, which indicates that the DCN-SR can select important events and ignore unexpected interactions. Third, there are some important interactions in the session that are not near the user's last click. For example, in Session2, the sixth event is more important than the last event. This may be due to a shift in the interest of the user. However, DCN-SR may also give them higher weight.
In summary, the DCN-SR applies a common interest network to capture the relationship between the long-term and short-term of the user and generate a common dependency representation of the user that represents the long-term and short-term preferences. The model is tested on two electric quotient data sets, and the result shows that the method has better effect than the existing method, on the Tmax data set, the Recall @10 is improved by 2.58%, the MRR @10 is improved by 3.78%, on the Tianchi data set, the Recall @10 is improved by 3.08%, and the MRR @10 is improved by 4.05%. The improvement of the DCN-SR model on short-session and active users is more obvious in the sensitivity and stability of the model.
The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (3)

1. A sequence recommendation method based on a dynamic interaction attention mechanism is characterized by comprising the following steps:
acquiring initial short-term preference and initial long-term preference of a user;
combining the initial short-term preference and the initial long-term preference to derive a long-term preference and a short-term preference according to an interactive attention network;
according to a sequence recommendation model, combining the long-term preference and the short-term preference to score corresponding articles, and recommending the articles for the user according to a scoring result;
the way to obtain the initial short-term preference of the user is through a context-based GRU model: performing the initial short-term preference query by a CGRU, the CGRU being:
Figure FDA0003308475570000011
zt＝σ(Wzxt+Vzat+Uzht-1)
Figure FDA0003308475570000012
rt＝σ(Wrxt+Vrat+Urht-1)
wherein, Wz，Vz，Uz，Wr，Vr，UrW, V and U are model parameters and are obtained by training, xtVector representation representing items input at time t, atThe final calculated h is the vector representation of the behavior actiontRepresenting hidden layer state at moment, wherein sigma represents a sigmoid function, and tan h represents a tangent function;
the method for obtaining the initial long-term preference of the user is to adopt a multi-layer perceptron to model the long-term preference of the user, and the modeling process of the multi-layer perceptron is as follows:
Figure FDA0003308475570000013
wherein, WmAnd bmParameters representing the m-th layer perceptron, phi the activation function, tan h the tangent function, xiIndicating the ith item, X, in the user's long-term interactioniDenotes xiThe final output state vector of (2);
the interactive attention network is:
Figure FDA0003308475570000014
Hl＝tan h[WlUl+(WsUs+Wths，T)CT]
Figure FDA0003308475570000015
Hs＝tan h[WsUs+Wths，T+(WlUl)C]
Figure FDA0003308475570000016
wherein, UlAnd UsIs the user's initial long and short term state vector representation,
Figure FDA0003308475570000021
transposed matrix, W, representing a representation of a user's long-term interactive itemsc，Wl，Ws，Wt，Whl，WhsObtained by training as a model parameter, hs，TThe last hidden layer state vector, representing the CGRU output, CT represents the transpose of the correlation matrix between long and short term interactions of the user,
Figure FDA0003308475570000022
represents a weight parameter for learning the importance of the user's long-term interactive items,
Figure FDA0003308475570000023
a representing a weight parameter for learning the importance of user short-term interactive articles, and a finally outputlAnd aSAttention weight that is a long-short term preference of the user;
the short-term preference and long-term preference of the user are respectively:
Figure FDA0003308475570000024
Figure FDA0003308475570000025
wherein, Uco-sAnd Uco-lRepresenting calculation through an interactive attention mechanismThe resulting final user long-short term preference representation, T represents the number of items in the user's short term interaction,
Figure FDA0003308475570000026
representing attention weight, h, of the t-th item in the user's short-term interactions，tA representation representing the t-th item in the user's short-term interaction, N representing the number of items in the user's long-term interaction,
Figure FDA0003308475570000027
representing the attention weight, X, of the nth item in the user's long-term interactionnA representation representing the nth item in the user's long-term interaction output by the multi-tier perceptron;
the mode of scoring the article pairs is as follows:
Figure FDA0003308475570000028
Figure FDA0003308475570000029
Figure FDA00033084755700000210
Figure FDA00033084755700000211
wherein, Bl，Bs，BTAll the parameters are obtained by training the model parameters,
Figure FDA00033084755700000212
representing a score on item i predicted from the user's long-term preference,
Figure FDA00033084755700000213
representing a predicted score on item i based on the user's short-term preferences,
Figure FDA00033084755700000214
represents the score on item i of the final prediction, sigma represents the sigmoid function,
Figure FDA00033084755700000215
middle viRepresents the ith item state vector, T represents the number of items in the user's short-term interaction, UlongIndicating the user's long-term preference, UshortIndicating short-term preferences of the user, finally output
Figure FDA00033084755700000216
Indicates the preference degree of the user U for all the items, UlongBy Uco-lIs calculated to obtain UshortBy Uco-sConcatenated with the representation of the last item in the short-term interaction.
2. The sequence recommendation method based on the dynamic interaction attention mechanism as claimed in claim 1, wherein the scoring result is further modified by a loss function after the corresponding item is scored, and the loss function is:
Figure FDA00033084755700000217
Figure FDA00033084755700000218
indicating the degree of preference, y, of the predicted user u for item iuiRepresenting the real result and V representing the number of all items.
3. A sequence recommendation system based on a dynamic interactive attention mechanism, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the method according to any one of the preceding claims 1 or 2 when executing the computer program.
CN201910533753.1A 2019-06-19 2019-06-19 Sequence recommendation method and system based on dynamic interaction attention mechanism Active CN110245299B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910533753.1A CN110245299B (en) 2019-06-19 2019-06-19 Sequence recommendation method and system based on dynamic interaction attention mechanism

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910533753.1A CN110245299B (en) 2019-06-19 2019-06-19 Sequence recommendation method and system based on dynamic interaction attention mechanism

Publications (2)

Publication Number Publication Date
CN110245299A CN110245299A (en) 2019-09-17
CN110245299B true CN110245299B (en) 2022-02-08

Family

ID=67888192

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910533753.1A Active CN110245299B (en) 2019-06-19 2019-06-19 Sequence recommendation method and system based on dynamic interaction attention mechanism

Country Status (1)

Country Link
CN (1) CN110245299B (en)

Families Citing this family (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110955826B (en) * 2019-11-08 2023-06-20 上海交通大学 Recommendation system based on improved cyclic neural network unit
CN110930219B (en) * 2019-11-14 2022-10-18 电子科技大学 Personalized merchant recommendation method based on multi-feature fusion
CN110992137B (en) * 2019-11-27 2023-09-01 上海风秩科技有限公司 Product recommendation method and device, electronic equipment and storage medium
CN110941764A (en) * 2019-12-03 2020-03-31 腾讯科技（深圳）有限公司 Object recommendation method and device, computer equipment and storage medium
CN111242729A (en) * 2020-01-07 2020-06-05 西北工业大学 Serialization recommendation method based on long-term and short-term interests
CN111506814B (en) * 2020-04-09 2023-11-28 苏州大学 Sequence recommendation method based on variational self-attention network
CN111563205A (en) * 2020-04-26 2020-08-21 山东师范大学 Cross-domain information recommendation method and system based on self-attention mechanism in shared account
CN111581520B (en) * 2020-05-25 2022-04-19 中国人民解放军国防科技大学 Item recommendation method and system based on item importance in session
CN112000873B (en) * 2020-06-19 2022-10-18 南京理工大学 Session-based recommendation system, method, device and storage medium
CN111753209B (en) * 2020-07-02 2023-07-18 南京工业大学 Sequence recommendation list generation method based on improved time sequence convolution network
WO2022011681A1 (en) * 2020-07-17 2022-01-20 国防科技大学 Method for fusing knowledge graph based on iterative completion
CN112364976B (en) * 2020-10-14 2023-04-07 南开大学 User preference prediction method based on session recommendation system
CN112269927A (en) * 2020-10-22 2021-01-26 辽宁工程技术大学 Recommendation method based on session sequence dynamic behavior preference coupling relation analysis
CN112256971B (en) * 2020-10-29 2023-06-20 清华大学深圳国际研究生院 Sequence recommendation method and computer readable storage medium
CN112861012B (en) * 2021-03-09 2022-11-08 河南工业大学 Recommendation method and device based on context and user long-term and short-term preference adaptive learning
CN113420058B (en) * 2021-07-01 2022-07-01 宁波大学 Conversational academic conference recommendation method based on combination of user historical behaviors
CN113468229B (en) * 2021-07-16 2023-04-25 南京信息工程大学 Recommendation system weighted similarity measurement method based on continuous scoring
CN115858926B (en) * 2022-11-29 2023-09-01 杭州电子科技大学 Sequence recommendation method based on complex multi-mode interest extraction and modeling of user
CN116304279B (en) * 2023-03-22 2024-01-26 烟台大学 Active perception method and system for evolution of user preference based on graph neural network
CN116127199B (en) * 2023-04-17 2023-06-16 昆明理工大学 User preference modeling method for clothing sequence recommendation

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101339563A (en) * 2008-08-15 2009-01-07 北京航空航天大学 Interest model update method facing to odd discovery recommendation
CN102056335A (en) * 2009-11-06 2011-05-11 华为技术有限公司 Mobile search method, device and system
CN102402766A (en) * 2011-12-27 2012-04-04 纽海信息技术（上海）有限公司 User interest modeling method based on web page browsing
CN103412757A (en) * 2013-08-19 2013-11-27 南京大学 Realizing method of personalized integrated framework of mobile applications
CN105488216A (en) * 2015-12-17 2016-04-13 上海中彦信息科技有限公司 Recommendation system and method based on implicit feedback collaborative filtering algorithm
CN108648049A (en) * 2018-05-03 2018-10-12 中国科学技术大学 A kind of sequence of recommendation method based on user behavior difference modeling
CN109213861A (en) * 2018-08-01 2019-01-15 上海电力学院 In conjunction with the tourism evaluation sensibility classification method of At_GRU neural network and sentiment dictionary
CN109359140A (en) * 2018-11-30 2019-02-19 苏州大学 A kind of sequence of recommendation method and device based on adaptive attention
CN109492229A (en) * 2018-11-23 2019-03-19 中国科学技术大学 A kind of cross-cutting sensibility classification method and relevant apparatus
CN110060097A (en) * 2019-04-01 2019-07-26 苏州市职业大学 User behavior sequence of recommendation method based on attention mechanism and convolutional neural networks

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6733295B2 (en) * 1996-09-25 2004-05-11 Sylvan Learning Systems, Inc. Learning system for enabling separate teacher-student interaction over selected interactive channels
US7987491B2 (en) * 2002-05-10 2011-07-26 Richard Reisman Method and apparatus for browsing using alternative linkbases
US20050131716A1 (en) * 2003-12-15 2005-06-16 Hanan Martin D. Method for determining compatibility
CN109522474B (en) * 2018-10-19 2021-05-18 上海交通大学 Recommendation method for mining deep user similarity based on interactive sequence data

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101339563A (en) * 2008-08-15 2009-01-07 北京航空航天大学 Interest model update method facing to odd discovery recommendation
CN102056335A (en) * 2009-11-06 2011-05-11 华为技术有限公司 Mobile search method, device and system
CN102402766A (en) * 2011-12-27 2012-04-04 纽海信息技术（上海）有限公司 User interest modeling method based on web page browsing
CN103412757A (en) * 2013-08-19 2013-11-27 南京大学 Realizing method of personalized integrated framework of mobile applications
CN105488216A (en) * 2015-12-17 2016-04-13 上海中彦信息科技有限公司 Recommendation system and method based on implicit feedback collaborative filtering algorithm
CN108648049A (en) * 2018-05-03 2018-10-12 中国科学技术大学 A kind of sequence of recommendation method based on user behavior difference modeling
CN109213861A (en) * 2018-08-01 2019-01-15 上海电力学院 In conjunction with the tourism evaluation sensibility classification method of At_GRU neural network and sentiment dictionary
CN109492229A (en) * 2018-11-23 2019-03-19 中国科学技术大学 A kind of cross-cutting sensibility classification method and relevant apparatus
CN109359140A (en) * 2018-11-30 2019-02-19 苏州大学 A kind of sequence of recommendation method and device based on adaptive attention
CN110060097A (en) * 2019-04-01 2019-07-26 苏州市职业大学 User behavior sequence of recommendation method based on attention mechanism and convolutional neural networks

Also Published As

Publication number Publication date
CN110245299A (en) 2019-09-17

Similar Documents

Publication Publication Date Title
CN110245299B (en) Sequence recommendation method and system based on dynamic interaction attention mechanism
US10896459B1 (en) Recommendation system using improved neural network
US20220198289A1 (en) Recommendation model training method, selection probability prediction method, and apparatus
US20210248651A1 (en) Recommendation model training method, recommendation method, apparatus, and computer-readable medium
Wu et al. Improving performance of tensor-based context-aware recommenders using bias tensor factorization with context feature auto-encoding
US10824940B1 (en) Temporal ensemble of machine learning models trained during different time intervals
CN111080400B (en) Commodity recommendation method and system based on gate control graph convolution network and storage medium
CN108431833A (en) End-to-end depth collaborative filtering
US20230153857A1 (en) Recommendation model training method, recommendation method, apparatus, and computer-readable medium
CN108230058A (en) Products Show method and system
CN110955826B (en) Recommendation system based on improved cyclic neural network unit
CN110807150A (en) Information processing method and device, electronic equipment and computer readable storage medium
CN111859149A (en) Information recommendation method and device, electronic equipment and storage medium
CN111737578A (en) Recommendation method and system
Wu et al. Neural modeling of buying behaviour for e-commerce from clicking patterns
Santana et al. MARS-Gym: A Gym framework to model, train, and evaluate Recommender Systems for Marketplaces
Salehi An effective recommendation based on user behaviour: a hybrid of sequential pattern of user and attributes of product
CN113688306A (en) Recommendation strategy generation method and device based on reinforcement learning
CN112819575B (en) Session recommendation method considering repeated purchasing behavior
Zhou et al. Towards real time team optimization
Diamantaras et al. Predicting Shopping Intent of e-Commerce Users using LSTM Recurrent Neural Networks.
CN116757747A (en) Click rate prediction method based on behavior sequence and feature importance
Kabra et al. Potent real-time recommendations using multimodel contextual reinforcement learning
CN116484092A (en) Hierarchical attention network sequence recommendation method based on long-short-term preference of user
Yao et al. Hierarchical attention based recurrent neural network framework for mobile MOBA Game recommender systems

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant