CN111310050B - Recommendation method based on multilayer attention - Google Patents

Abstract

The invention discloses a recommendation method based on multilayer attention, which comprises the steps of obtaining historical behaviors of a user to be recommended and generating a user behavior sequence according to the historical behaviors; determining a recommendation score corresponding to each article in a preset article set based on the user behavior sequence and the trained recommendation network model; and determining a recommended article corresponding to the user to be recommended according to the recommendation score, and pushing the recommended article to the user to be recommended. According to the method and the system, the historical behaviors of the user are used as input items, and the context characteristics of the historical behaviors of the user are learned through the recommendation network model to determine the recommended articles, so that the accuracy of recommending the articles can be improved.

Description

A recommendation method based on multi-layer attention

Technical Field

The present invention relates to the field of artificial intelligence technology, and in particular to a recommendation method based on multi-layer attention.

Background Art

With the continuous development of mobile communication technology and the deepening of big data services, various Internet applications have emerged, resulting in a surge in information. Information overload has become an urgent problem to be solved. For example, there are hundreds of millions of video information in short video platforms such as Douyin, Kuaishou, and Tencent Weishi, and dazzling product data in e-commerce platforms such as Taobao, JD.com, and Amazon. However, the information available for Internet applications has complex characteristics such as multi-source heterogeneity, uneven distribution, and large scale. For recommendation strategies, although the data information seems to be rich, it actually has strong limitations: (1) Classic collaborative filtering methods cannot utilize the deep features of users and items; (2) Content-based recommendation methods require effective feature extraction. Traditional shallow models rely on artificially designed features, and their effectiveness and scalability are limited, which restricts the performance of recommendation algorithms; (3) The explicit feedback provided by users is far less than the implicit feedback. The application scenarios of using users' explicit feedback for recommendation are limited.

Summary of the invention

The technical problem to be solved by the present invention is to provide a recommendation method based on multi-layer attention in view of the deficiencies of the prior art.

In order to solve the above technical problems, the technical solution adopted by the present invention is as follows:

A recommendation method based on multi-layer attention, the method comprising:

Obtaining historical behaviors of the user to be recommended, and generating a user behavior sequence based on the historical behaviors;

Determine a recommendation score corresponding to each item in a preset item set based on the user behavior sequence and the trained recommendation network model;

Determine the recommended item corresponding to the to-be-recommended user according to the recommendation score, and push the recommended item to the to-be-recommended user.

The recommendation method based on multi-layer attention, wherein the step of obtaining the historical behavior of the user to be recommended and generating the user behavior sequence according to the historical behavior specifically includes:

Obtaining historical behaviors of the user to be recommended, wherein each historical behavior includes an item identifier and a behavior time;

The historical behaviors are sorted according to the behavior time to obtain a user behavior sequence.

The recommendation method based on multi-layer attention, wherein the determining of the recommendation score corresponding to each item in the preset item set based on the user behavior sequence and the trained recommendation network model specifically includes:

For each item in the preset item set, obtain the item vector corresponding to the item;

An item sequence is generated based on the user behavior sequence and the item vector, and the item sequence is input into a trained recommendation network model so as to output a recommendation score corresponding to the item through the recommendation network model.

The recommendation method based on multi-layer attention, wherein the training process of the recommendation network model includes:

Dividing the training user behavior sequence into a first behavior sequence and a second behavior sequence, wherein the second behavior sequence includes the last behavior record in the training user behavior sequence;

Performing masking processing on the first behavior sequence according to a preset masking strategy to obtain a masked first behavior sequence;

A generated behavior sequence corresponding to the first behavior sequence after masking output by the mask network model to be trained;

The recommendation network model to be trained is trained based on the generated behavior sequence and the second behavior sequence to obtain a trained recommendation network model.

The recommendation method based on multi-layer attention, wherein the recommendation network model includes a multi-layer attention structure; the training of the recommendation network model to be trained based on the generated behavior sequence and the second behavior sequence to obtain a trained recommendation network model specifically includes:

Divide the generated behavior sequence and the second behavior sequence into a plurality of sub-vectors respectively;

Inputting the equally divided generated sequence and the second behavior sequence into a multi-layer attention structure, and outputting the attention score of the generated behavior sequence relative to the second behavior sequence through the multi-layer attention structure;

The network parameters of the recommendation network model to be trained are modified based on the attention score to obtain a trained recommendation network model.

The recommendation method based on multi-layer attention, wherein the mask network model includes a multi-layer attention structure; the generated behavior sequence corresponding to the first behavior sequence after the mask network model to be trained outputs the mask specifically includes:

The masked first behavior sequence is input into the multi-layer attention structure, and the behavior sequence is generated through the output of the multi-layer attention structure.

The recommendation method based on multi-layer attention, wherein the mask network model further includes a recognition structure, and after the mask network model to be trained outputs a generated behavior sequence corresponding to the first behavior sequence after masking, the method further includes:

Inputting the generated behavior sequence into the recognition structure, and generating an item identification through the recognition structure output;

Based on the generated object identification, the network parameters of the mask network model to be trained are modified to train the mask network model to be trained.

The recommendation method based on multi-layer attention, wherein the step of correcting the network parameters of the recommendation network model to be trained based on the attention score to obtain the trained recommendation network model specifically includes:

Based on the attention score, the network parameters of the recommended network model to be trained are modified, and when the network parameters of the recommended network model meet the preset conditions, the first behavior sequence and the second behavior sequence are input into the network parameters of the recommended network model whose network parameters meet the preset conditions;

Outputting the attention score corresponding to the second behavior sequence through the recommendation network model whose network parameters meet the preset conditions;

Based on the attention score, the network parameters of the recommendation network model whose network parameters meet the preset conditions are corrected to obtain a trained recommendation network model.

A computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps in any of the multi-layer attention-based recommendation methods described above.

A terminal device comprises: a processor, a memory and a communication bus; the memory stores a computer-readable program executable by the processor;

The communication bus realizes the connection and communication between the processor and the memory;

When the processor executes the computer-readable program, the processor implements the steps in any of the above-described multi-layer attention-based recommendation methods.

Beneficial effects: Compared with the prior art, the present invention provides a recommendation method based on multi-layer attention, which obtains the historical behavior of the user to be recommended and generates a user behavior sequence based on the historical behavior; determines the recommendation score corresponding to each item in the preset item set based on the user behavior sequence and the trained recommendation network model; determines the recommended item corresponding to the user to be recommended based on the recommendation score, and pushes the recommended item to the user to be recommended. The present invention uses the user's historical behavior as an input item, and learns the contextual features of the user's historical behavior through the recommendation network model to determine the recommended item, which can improve the accuracy of the recommended item.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a multi-layer attention-based recommendation method provided by the present invention.

FIG2 is a flow chart of the multi-layer attention-based recommendation method provided by the present invention.

FIG3 is a flow chart of the training process of the recommendation network model to be trained in the multi-layer attention-based recommendation method provided by the present invention.

FIG. 4 is a schematic diagram of the structure of the terminal device provided by the present invention.

DETAILED DESCRIPTION

The present invention provides a recommendation method based on multi-layer attention. In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

It will be understood by those skilled in the art that, unless expressly stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and/or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and/or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The term "and/or" used herein includes all or any unit and all combinations of one or more associated listed items.

It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless specifically defined as herein.

This embodiment provides a recommendation method based on multi-layer attention, which can be applied to electronic devices, and the electronic devices can be implemented in various forms. For example, mobile phones, tablet computers, PDAs, personal digital assistants (PDAs), etc. In addition, the functions implemented by the method can be implemented by calling program codes by a processor in the electronic device, and of course the program codes can be stored in a computer storage medium. It can be seen that the electronic device at least includes a processor and a storage medium.

As shown in FIG. 1 and FIG. 2 , this implementation provides a recommendation method based on multi-layer attention, and the method may include the following steps:

S10: Obtain historical behaviors of the user to be recommended, and generate a user behavior sequence according to the historical behaviors.

Specifically, the historical behavior of the recommended user can be recorded through an Internet application (for example, Taobao application, Tencent video application, etc.), and all the behavior records of the users of the Internet application on the Internet application. In addition, since the market directions of Internet applications may be different, the user behaviors recorded by each Internet application may also be different. For example, a shopping website records the purchase records of each user, and a video website records the viewing records of each user. However, for each Internet application, each user (user_id) in the Internet application can have multiple behavior records (event), wherein each behavior record includes an item identifier (item_id) and a behavior time (timestamp). Of course, it is worth noting that for each historical behavior of the user to be recommended, the behavior type corresponding to the historical behavior is the same, for example, all historical behaviors are shopping behaviors, all historical behaviors are video watching behaviors, etc.

Furthermore, since each acquired historical behavior includes the behavior time, after all the historical behaviors are acquired, all the acquired historical behaviors can be sorted according to the behavior time to obtain a user behavior sequence, wherein the behavior time corresponding to each historical behavior in the user behavior sequence is earlier than the moment when the historical behavior of the user to be recommended is acquired, and the moment when the historical behavior of the user to be recommended is acquired can be the moment when the user logs into the Internet application. Accordingly, in an implementation of this embodiment, the acquisition of the historical behavior of the user to be recommended and the generation of the user behavior sequence according to the historical behavior specifically include:

S11, obtaining historical behaviors of the user to be recommended, wherein each historical behavior includes an item identifier and a behavior time;

S12. Sort the historical behaviors according to the behavior time to obtain a user behavior sequence.

Specifically, the historical behavior of the user to be recommended can be one or more, so after obtaining the historical behavior corresponding to the user to be recommended, all the historical behaviors obtained can be sorted into: [user_id: event ¹ → event ² → event ³ ...... → event ⁿ ]], where user_id represents the user ID of the user to be recommended, event represents the behavior record, and n represents the number of historical behaviors. In addition, since each behavior record includes an item ID and a behavior time, all the historical behaviors obtained can be arranged according to the behavior time to obtain the user behavior sequence corresponding to the user to be recommended, which can be expressed as [user_id: item_id ¹ → item_id ² → item_id ³ ...... → item_id ⁿ ], where item_id represents the item ID, and n represents the number of historical behaviors.

S20: Determine a recommendation score corresponding to each item in a preset item set based on the user behavior sequence and the trained recommendation network model.

Specifically, the recommendation network model is pre-trained and is used to determine the recommendation score of the user for the recommended item based on the user behavior sequence formed by the historical behavior of the user to be recommended. It can be understood that the input items of the recommendation network model are the user behavior sequence and the item to be recommended, and the output is the score corresponding to the item to be recommended. The preset item set includes all items contained in the Internet application to which the recommendation method is applied. For example, if the Internet application is Taobao, then the preset item set is all the goods sold in Taobao. Therefore, in one implementation of this embodiment, the recommendation score corresponding to each item in the preset item set is determined based on the user behavior sequence and the trained recommendation network model as follows: for each item in the preset item set, obtain the item vector corresponding to the item; generate an item sequence based on the user behavior sequence and the item vector, and input the item sequence into the trained recommendation network model to output the recommendation score corresponding to the item through the recommendation network model.

In an implementation of this embodiment, as shown in FIG3 , the training process of the recommendation network model includes:

H10, dividing the training user behavior sequence into a first behavior sequence and a second behavior sequence, wherein the second behavior sequence includes the last behavior record in the training user behavior sequence;

H20, performing mask processing on the first behavior sequence according to a preset masking strategy to obtain a masked first behavior sequence;

H30, based on the mask network model to be trained outputting the generated behavior sequence corresponding to the first behavior sequence after masking;

H40. Train the recommendation network model to be trained based on the generated behavior sequence and the second behavior sequence to obtain a trained recommendation network model.

Specifically, in the step H10, the user behavior sequence to be trained is included in a pre-training sample set, and the training sample set includes several user behavior sequences to be trained, and for each user behavior sequence to be trained in the training sample set, the behavior categories corresponding to the user behavior sequence to be trained are the same, for example, all are shopping categories, all are video watching categories, etc. In addition, each user behavior sequence to be trained includes item identifiers corresponding to multiple behavior records, and the item identifiers corresponding to the multiple behavior records are arranged in chronological order of behavior time. For details, please refer to the process of obtaining the user behavior sequence corresponding to the user to be recommended, which will not be repeated here. Of course, it is worth noting that the first behavior sequence includes all the last behavior records in the user behavior sequence to be trained except the last one, and the second behavior sequence includes the last behavior record in the user behavior sequence to be trained.

Further, in an implementation of this embodiment, before training the recommendation network model Nxet_item to be trained based on the preset training sample set, the training sample set can be expanded to improve the diversity of training data in the training sample set, thereby improving the accuracy of the recommendation network model obtained by training based on the training sample set. The expansion process of the training sample set may include: for each training user behavior sequence in the training sample set, obtaining the last behavior record in the training user behavior sequence, randomly selecting an item identifier to replace the item identifier in the behavior record, so as to obtain a negative sample corresponding to the training user sequence, and the randomly selected item identifier is different from the item identifier in the behavior record. Among them, the randomly selected item identifier is randomly selected from the preset item library, and the behavior type of the user behavior based on the random item identifier and the behavior record form is the same as the behavior type of the user behavior in the last behavior record. For example, if the last behavior record is a purchase behavior, then the random item identifier and the behavior record are based on the purchase behavior.

For example, if the training user behavior sequence is: item_id ¹ →item_id ² →item_id ³ →item_id ⁴ →item_id ⁵ , and the training user behavior sequence is recorded as a positive sample, then the negative sample generated based on the positive sample can be: item_id ¹ →item_id ² →item_id ³ →item_id ⁴ →rand_item, where rand_item is the randomly selected item identifier.

Further, in the step H20, the preset mask strategy is pre-set, and the training samples are masked according to the preset mask strategy to determine the item identifiers that need to be masked in the training user behavior sequence. In this embodiment, the mask strategy is to randomly select a number of item identifiers in the training user behavior sequence according to the first preset probability, and then mask the selected item identifiers according to the mask rule. Among them, the first preset probability can be 15%, etc., the mask rule can be replaced by the "[Mask]" mark for the second preset probability, the probability of the third preset probability is randomly replaced by other item identifiers, and the fourth preset probability is unchanged, wherein the sum of the second preset probability, the third preset probability and the fourth preset probability is 1. For example, the second preset probability is 80%, the third preset probability is 10%, and the fourth preset probability is 10%.

For example, the training user behavior sequence is: item_id ¹ →item_id ² →item_id ³ →item_id ⁴ →item_id ⁵ , where item_id ² is selected to replace the hidden mark according to the mask strategy. Then, after replacing item_id ² with [mask], we can get item_id ¹ →[mask]→item_id ³ →item_id ⁴ →item_id ⁵ . In addition, after obtaining item_id ¹ →[mask]→item_id ³ →item_id ⁴ →item_id ⁵ , item_id ^{1 →[mask]→item_id 3} →item_id ⁴ →item_id ⁵ is updated according to the ^mask rule. The update result is: 80% probability of using the "[Mask]" tag to replace the selected item, and the training user behavior sequence after mask processing is item_id ¹ →[mask]→item_id ³ →item_id ⁴ →item_id ⁵ ; 10% probability of keeping item_id ² unchanged, and the training user behavior sequence after mask processing is item_id ¹ →item_id ² →item_id ³ →item_id ⁴ →item_id ⁵ ; 10% probability of replacing item_id ² with a randomly selected item identifier, and the training user behavior sequence after mask processing is item_id ¹ →rand_item→item_id ³ →item_id ⁴ →item_id ^5. It is understandable that, for each training user behavior sequence, the probability of random replacement of item identifiers in the training user behavior sequence is only 1.5% (ie, 10% of 10%), which does not affect the distribution of user interests contained in the user behavior sequence.

Furthermore, since each training user behavior sequence in the training sample set corresponds to a negative sample, after taking negative samples and masking the training user behavior sequence, the training user behavior sequence can correspond to two user behavior sequences, which are respectively recorded as item_id ¹ →[mask]→item_id ³ →...→item_id ^n-1 →item_id ⁿ obtained according to the positive sample mask and item_id ¹ →item_id ² →[mask]→...→item_id ^n-1 →rand_item obtained according to the negative sample mask. It can be understood that the user behavior sequence corresponding to each user user_id in the training sample set can correspond to two user behavior sequences after taking negative samples and masking, which can be expressed as:

Further, in the step H30, the masked network model Masked_lm to be trained is used to reconstruct the masked user behavior sequence through the self-attention score, so that the generated behavior sequence after reconstruction contains semantic features. It can be understood that the generated behavior sequence corresponding to the masked first behavior sequence can be generated by the masked network model to be trained. In addition, in order to input the masked first behavior sequence into the masked network model to be trained, each item identification needs to be identified with a vector to convert the first behavior sequence and the second behavior sequence into behavior vectors, wherein the first behavior sequence can be converted into a first behavior vector, and the second behavior sequence can be converted into a second behavior sequence.

In one implementation of the present embodiment, for each item identifier, the item identifier can be converted into a continuous dense vector of a fixed length, and the length of the continuous dense vector corresponding to each item identifier is equal. Among them, the process of converting the item identifier into a continuous dense vector can be: for each item identifier in the training user sequence behavior, the one-hot encoding vector of the item identifier is vector-multiplied with a preset dense initialization matrix, the product result is the value of a row in the initialization matrix, and the product result is used as the continuous dense vector corresponding to the item identifier. In this way, each item identifier will be mapped into a dense vector of a fixed length. In addition, the one-hot encoding vector of the item identifier is obtained based on the position sequence encoding of the item identifier in the corresponding training user behavior sequence, and the position of the item identifier is represented by 1, and the remaining positions are represented by.

For example: Assume that the one-hot encoding vector corresponding to the item representation is [0 0 0 1 0], and the preset dense initialization matrix is:

Then, the continuous dense vector corresponding to the item identifier is:

Of course, it is worth mentioning that during the training process of the untrained recommendation network model, as the number of iterations of the gradient descent method on the network parameters of the untrained recommendation network model increases, the preset dense initialization matrix will be continuously updated, so that the essential meaning of the item will be given to the continuous dense vector mapped by the item.

Further, in an implementation of this embodiment, the mask network model to be trained includes a multi-layer attention structure, through which the self-attention score is reconstructed on the masked first behavior sequence to obtain a generated behavior sequence corresponding to the first behavior sequence. Accordingly, the generated behavior sequence corresponding to the masked first behavior sequence output based on the mask network model to be trained is specifically:

H30a. Input the masked first behavior sequence into the multi-layer attention structure, and generate a behavior sequence through the output of the multi-layer attention structure.

Specifically, the multi-layer attention mechanism adopts a multi-layer attention mechanism, and the multi-head attention mechanism calculates the attention mechanism in multiple semantic spaces, so that the multi-semantic space helps to improve the model performance of the mask network model. The first behavior sequence is pre-converted into a first behavior vector based on the above conversion process, and the specific conversion process can be referred to. In addition, the multi-layer attention structure can be expressed as:

Among them, _h0 is the first behavior vector corresponding to the first behavior sequence after masking, and _h3 is the generated behavior vector output by the multi-layer attention mechanism, wherein the generated behavior vector is the generated behavior sequence output by the multi-layer attention structure.

Further, in an implementation of this embodiment, inputting the masked first behavior sequence into the multi-layer attention structure, and generating the behavior sequence through the multi-layer attention structure output may specifically include the following three steps:

The first step: divide the first behavior vector into h equal parts, that is, for each item vector code w _i in the first behavior sequence, divide it into w _i = {w _i1 ,w _i2 ,...,w _ih }. For the behavior sequence h ₀ ∈R ^N*K (N represents the number of items, K represents the length of the dense vector) containing N items, the vector h ₀ ∈R ^hN*K/h (N represents the number of items, K represents the length of the dense vector, represents the equal division fraction) after equal division.

Step 2: The divided vectors are mapped in different semantic spaces through multi-layer attention mechanisms to learn the actual meanings of different items. For example, a wool sweater and a wool blanket may have a high similarity in the first semantic space, but a low similarity in the second semantic space. The actual meaning of the items represented by the first semantic space is the material of the items, and the actual meaning represented by the second semantic space is the purpose of the items. In addition, in this embodiment, the attention scores in the semantic space are calculated in parallel, so that compared with not dividing the semantic space, no additional calculation amount is added. Among them, the mapping process of the multi-layer attention mechanism can be shown in formula (1), and the calculation process of the self-attention score can be shown in formula (2), wherein formula (1) and formula (2) are as follows:

Q＝QW _i ^Q ,K＝KW _i ^K ,V＝VW _i ^V (1)

表示查询向量的映射矩阵，

表示键向量的映射矩阵,

表示值向量的映射矩阵，K表示第一行为向量的大小，h表示语义空间的数量，

表示将第一行为向量均分成h份，也就是将第一行为向量分在h个语义空间里。Q，K,V均表示第一行为向量，d_k表示语义空间中向量的维度。in,

represents the mapping matrix of the query vector,

represents the mapping matrix of the key vector,

represents the mapping matrix of the value vector, K represents the size of the vector in the first row, h represents the number of semantic spaces,

It means that the vector of the first line is divided into h parts, that is, the vector of the first line is divided into h semantic spaces. Q, K, V all represent the vector of the first line, and d _k represents the dimension of the vector in the semantic space.

Step 3: After calculating the attention scores in the semantic space, we reconnect the outputs of the semantic space as follows:

MultiHead＝Concat(head ₁ ,...,head _h )W ⁰ wherehead _i ＝Attention(Q,K,V)

Among them, W ⁰ represents the mapping parameter, head ₁ ,...,head _h all represent semantic spaces, and h is the number of semantic spaces.

For example: the first line sequence is [a,b,c,d], the positive sample is [e], the first line sequence after masking is [a,mask,c,d], the first line vector is obtained after vector conversion, the first line vector is divided into h parts, and h semantic spaces are obtained. Assuming that the vector in one of the semantic spaces is [E _a ,E _mask ,E _c ,E _d ], the vector in the semantic space is mapped into a query vector, a key vector and a value vector, where the query vector is [Q _a ,Q _mask ,Q _c ,Q _d ], the key vector is [K _a ,K _mask ,K _c ,K _d ], and the value vector is [V _a ,V _mask ,V _c ,V _d ].

Using the query vector and key vector, we get the attention score matrix:

的结果。Where, Q _m represents Q _mask and K _m represents K _mask . The numbers in the figure represent

result.

The attention score matrix is then multiplied by the value vector to obtain the value vector weighted by the attention score. For example, the value of the vector E _a in the semantic space after the multi-head attention mechanism is [0.4*V _a +0.2*V _mask +0.2*V _c +0.2*V _d ], the value of E _mask after the multi-head attention mechanism is [0.3*V _a +0.3*V _mask +0.2*V _c +0.2*V _d ], the value of E _c after the multi-head attention mechanism is [0.1*V _a +0.3*V _mask +0.3*V _c +0.3*V _d ], and the value of E _d after the multi-head attention mechanism is [0.3*V _a +0.2*V _mask +0.3*V _c +0.2*V _d ].

Furthermore, in an implementation of this embodiment, in order to sequence the mask network model to be trained, after the generated behavior vector is output through the multi-layer attention mechanism, the generated code corresponding to the item marked or replaced by "[Mask]" in the generated behavior vector is searched, and the network parameters of the mask network model to be trained are modified based on the generated code. Accordingly, the mask network model also includes a recognition structure, and after the generated behavior sequence corresponding to the first behavior sequence after the mask network model to be trained outputs the mask, the method also includes:

Inputting the generated behavior sequence into the recognition structure, and generating an item identification through the recognition structure output;

Based on the generated object identification, the network parameters of the mask network model to be trained are modified to train the mask network model to be trained.

Specifically, the recognition structure can be a softmax layer, which determines the probability that the generated code is the real code corresponding to the generated code in the training user behavior sequence through the softmax layer, and modifies the network parameters of the mask network model to be trained based on the probability to train the mask network model to be trained. The softmax function corresponding to the softmax layer can be expressed as:

为训练样本集对应的所有物品对应的编码矩阵。Where M is the generated code corresponding to the “[Mask]” mark or the replaced item in the generated behavior sequence.

It is the encoding matrix corresponding to all items in the training sample set.

Further, in the step H40, the recommendation network model includes a multi-layer attention structure; the training of the recommendation network model to be trained based on the generated behavior sequence and the second behavior sequence to obtain the trained recommendation network model specifically includes:

Divide the generated behavior sequence and the second behavior sequence into a plurality of sub-vectors respectively;

Inputting the equally divided generated sequence and the second behavior sequence into a multi-layer attention structure, and outputting the attention score of the generated behavior sequence relative to the second behavior sequence through the multi-layer attention structure;

The network parameters of the recommendation network model to be trained are modified based on the attention score to obtain a trained recommendation network model.

Specifically, before the second behavior sequence is equally divided into several sub-vectors, it needs to be converted into a second behavior vector first, wherein the process of converting the second behavior sequence into the second behavior vector is the same as the process of converting the first behavior sequence into the first behavior vector. For details, please refer to the process of converting the first behavior sequence into the first behavior vector. The generated behavior sequence is generated by the mask network model to be trained based on the first behavior vector. The multi-layer attention structure is a single-item attention model from left to right, which predicts whether a user with historical behavior will have behavior on a new item. Taking an e-commerce website as an example, if a user buys a wool coat, jeans, snow boots, and plush toys, when this user visits the e-commerce website, we need to predict the next item she will buy, such as a new wool coat. In addition, it can be obtained by expanding the preset training sample set that each training user behavior sequence in the preset sample set corresponds to a negative sample. Therefore, the second behavior sequence can be the last item identifier in the training user behavior sequence, or the last item identifier in the negative sample corresponding to the training user behavior sequence. In this way, the positive and negative samples can be used as the second behavior sequence, which can improve the diversity of the training samples of the recommendation network model to be trained, thereby improving the accuracy of the recommendation network model obtained by training.

Furthermore, the recommendation network model to be trained is used to calculate the attention score between the target item vector and the historical behavior vector, and learn the user's interest preferences from the user's historical behavior sequence. Among them, the multi-layer attention structure is the same as the multi-layer attention structure in the mask network model to be trained, but the output items of the multi-layer attention structure and the mask network model to be trained are different. The output items of the multi-layer attention structure are the second behavior training and the generated behavior sequence output by the mask network model to be trained, and the input items of the mask network model to be trained are the first behavior sequence after mask processing. In addition, the processing process of the multi-layer attention structure also includes three steps, among which the first step and the third step are the same as the first step and the third step in the processing process of the multi-layer attention structure in the mask network model to be trained, and will not be repeated here, and only the second step will be explained.

为第二行为序列和生成行为序列的自注意力分数。The specific processing process of the second step can be: calculating the attention score between the second behavior sequence and the generated behavior sequence, and its input is Q=I _target represents the vector of the second behavior sequence after k equal divisions, K=V=I _hist represents the vector of the generated behavior sequence after k equal divisions,

is the self-attention score of the second behavior sequence and the generated behavior sequence.

Furthermore, the recommendation network model to be trained is jointly trained with the mask network model to be trained during the training process, and the recommendation network model to be trained is used alone after the training is completed, and the mask model to be trained uses training samples with mask tags as input items, while in actual applications, the user behavior sequence obtained does not carry mask tags. Therefore, in order to make the trained recommendation network model adaptable to the training of user behaviors without mask tags, after the network model of the recommendation network to be trained meets the preset conditions, the network parameters of the recommendation network model whose network parameters meet the preset conditions can be corrected using user behaviors without mask tags.

Correspondingly, in an implementation of the present embodiment, the correcting the network parameters of the recommended network model to be trained based on the attention score to obtain a trained recommended network model specifically includes: correcting the network parameters of the recommended network model to be trained based on the attention score, and when the network parameters of the recommended network model meet preset conditions, inputting the first behavior sequence and the second behavior sequence into the model parameters of the recommended network model whose network parameters meet the preset conditions; outputting the attention score corresponding to the second behavior sequence through the recommended network model whose network parameters meet the preset conditions; and correcting the model parameters of the recommended network model whose network parameters meet the preset conditions based on the attention score to obtain a trained recommended network model.

S30: Determine the recommended item corresponding to the user to be recommended according to the recommendation score, and push the recommended item to the user to be recommended.

Specifically, the recommendation score is the self-attention score of the item to be recommended and the user's historical behavior sequence. For each item to be recommended in the preset item set, the item to be recommended corresponds to a self-attention score, and the self-attention score is used as the recommendation score corresponding to the item to be recommended. In addition, after obtaining the recommendation score corresponding to each item to be recommended, the recommendation scores corresponding to each item to be recommended can be compared to determine the recommended item corresponding to the user to be recommended. Among them, the recommended item corresponding to the user to be recommended can be the item to be recommended with the highest recommendation score among all recommended items.

Based on the above-mentioned recommendation method based on multi-layer attention, this embodiment provides a computer-readable storage medium, which stores one or more programs. The one or more programs can be executed by one or more processors to implement the steps in the recommendation method based on multi-layer attention as described in the above embodiment.

Based on the above-mentioned recommendation method based on multi-layer attention, the present invention also provides a terminal device, as shown in FIG4, which includes at least one processor (processor) 20; display screen 21; and memory (memory) 22, and may also include a communication interface (Communications Interface) 23 and a bus 24. Among them, the processor 20, the display screen 21, the memory 22 and the communication interface 23 can communicate with each other through the bus 24. The display screen 21 is configured to display a preset user guide interface in the initial setting mode. The communication interface 23 can transmit information. The processor 20 can call the logic instructions in the memory 22 to execute the method in the above-mentioned embodiment.

In addition, the logic instructions in the memory 22 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product.

The memory 22, as a computer-readable storage medium, can be configured to store software programs, computer executable programs, such as program instructions or modules corresponding to the methods in the embodiments of the present disclosure. The processor 20 executes functional applications and data processing by running the software programs, instructions or modules stored in the memory 22, that is, implementing the methods in the above embodiments.

The memory 22 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and at least one application required for a function; the data storage area may store data created according to the use of the terminal device, etc. In addition, the memory 22 may include a high-speed random access memory and may also include a non-volatile memory. For example, a variety of media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, may also be a transient storage medium.

In addition, the specific process of loading and executing the multiple instruction processors in the above-mentioned storage medium and the terminal device has been described in detail in the above-mentioned method, and will not be described one by one here.

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 make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (7)

1. A multi-tier attention-based recommendation method, the method comprising:

acquiring historical behaviors of a user to be recommended, and generating a user behavior sequence according to the historical behaviors;

determining a recommendation score corresponding to each article in a preset article set based on the user behavior sequence and the trained recommendation network model;

determining a recommended article corresponding to the user to be recommended according to the recommendation score, and pushing the recommended article to the user to be recommended;

wherein the recommended network model comprises a multi-layer attention structure, and the training process of the recommended network model comprises:

dividing the training user behavior sequence into a first behavior sequence and a second behavior sequence, wherein the second behavior sequence comprises the last behavior record in the training user behavior sequence;

performing mask processing on the first behavior sequence according to a preset mask strategy to obtain a masked first behavior sequence;

outputting a generated behavior sequence corresponding to the first behavior sequence after the mask is output based on a mask network model to be trained;

equally dividing the generated behavior sequence and the second behavior sequence into a plurality of sub-vectors respectively;

inputting the generated sequence and the second behavior sequence after being divided into the multilayer attention structure, and outputting an attention score of the generated behavior sequence relative to the second behavior sequence through the multilayer attention structure;

correcting network parameters of a recommended network model to be trained based on the attention score, and inputting the first behavior sequence and the second behavior sequence into the network parameters of the recommended network model of which the network parameters meet preset conditions when the network parameters of the recommended network model meet the preset conditions;

outputting the attention score corresponding to the second behavior sequence through a recommended network model with the network parameter meeting a preset condition;

and correcting the network parameters of the recommended network model with the network parameters meeting the preset conditions based on the attention scores to obtain the trained recommended network model.

2. The multi-layer attention-based recommendation method according to claim 1, wherein the obtaining of the historical behaviors of the user to be recommended and the generating of the user behavior sequence according to the historical behaviors specifically comprise:

acquiring historical behaviors of a user to be recommended, wherein each historical behavior comprises an article identifier and behavior time;

and sequencing the historical behaviors according to the behavior time to obtain a user behavior sequence.

3. The multi-tier attention-based recommendation method according to claim 1, wherein the determining recommendation scores corresponding to respective items in a preset item set based on the user behavior sequence and a trained recommendation network model specifically comprises:

for each article in a preset article set, acquiring an article vector corresponding to the article;

and generating an article sequence based on the user behavior sequence and the article vector, and inputting the article sequence into a trained recommendation network model so as to output a recommendation score corresponding to the article through the recommendation network model.

4. The multi-tier attention-based recommendation method of claim 1, wherein said mask network model comprises a multi-tier attention structure; the generating behavior sequence corresponding to the first behavior sequence after the mask is output based on the mask network model to be trained specifically includes:

inputting the masked first behavior sequence into the multi-layer attention structure, and generating a behavior sequence through the multi-layer attention structure output.

5. The multi-layer attention-based recommendation method according to claim 4, wherein the mask network model further comprises an identification structure, and after outputting a generation behavior sequence corresponding to the masked first behavior sequence based on the mask network model to be trained, the method further comprises;

inputting the generated behavior sequence into the recognition structure, and outputting a generated article identifier through the recognition structure;

and correcting the network parameters of the mask network model to be trained based on the generated article identifier so as to train the mask network model to be trained.

6. A computer-readable storage medium storing one or more programs, the one or more programs being executable by one or more processors to perform the steps of the multi-tiered attention recommendation method as recited in any one of claims 1-5.

7. A terminal device, comprising: a processor, a memory, and a communication bus; the memory has stored thereon a computer readable program executable by the processor;

the communication bus realizes connection communication between the processor and the memory;

the processor, when executing the computer readable program, implements the steps in the multi-tiered attention recommendation method as recited in any of claims 1-5.

Images