Information recommendation / determination method and device and information recommendation system
By analyzing users' click history and querying the relevance information between items and polymer products, the system matches and recommends combinations of polymer products and their target items, thus solving the problem of inaccurate user interest recommendations in existing systems and improving user experience.
Patent Information
- Application Number
- CN202310595489.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-24
- Publication Date
- 2024-11-26
AI Technical Summary
Existing information recommendation systems struggle to accurately recommend polymer products that users are interested in and their target items, resulting in a poor user experience.
By utilizing the target user's item click history sequence, querying the primary relevance information between different items and the secondary relevance information between items and polymer items, matching polymer items in the polymer item set with items in the item set, identifying the target item, and recommending combinations of polymer items and the target item to the user.
It enables accurate recommendations of polymer products and target items that users are interested in, thus improving the user experience.
Smart Images

Figure CN119027205A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of computers, and in particular to an information recommendation / determination method and device and an information recommendation system. Background Art
[0002] With the development of information technology, more and more users choose to browse and purchase items in e-commerce systems. Information recommendation systems in e-commerce systems can mine items of interest to users based on their browsing and purchasing behaviors, and recommend them to users to improve user experience.
[0003] In the product display form of the e-commerce system, in addition to single items, there are also aggregated products that include multiple items. The display form of aggregated products is usually: the image of the target item is used as the main image for display, and the description information related to the aggregated product is displayed below the main image. The target item is one of the items in the aggregated product and is the item that drives traffic to the aggregated product page.
[0004] Research has found that users' interest in aggregated products is mainly reflected in two aspects: on the one hand, users' interest in aggregated products, and on the other hand, users' interest in target items displayed in aggregated products. How to recommend aggregated products of interest to users and how to select the items that users are most interested in from aggregated products as target items for diversion are challenges facing current information recommendation services. Summary of the invention
[0005] The disclosed embodiment utilizes the target user's item click history behavior sequence to query the first correlation information between different items and the second correlation information between items and aggregated items, respectively, to obtain the item set and aggregated item set that the target user may be interested in related to the target user's historical clicked items, and then by matching the aggregated items in the aggregated item set with the items in the item set, determine the target items of the target user for the aggregated items in the aggregated item set, and recommend the aggregated items in the aggregated item set and the combination of the target items to the target user. Thus, the aggregated items and target items that each user is interested in are accurately mined and located, and the aggregated items and target items that each user is interested in are personalized for combination recommendation, and different combinations of aggregated items and target items are recommended to users with different interests, so as to achieve personalized information recommendation effect and improve user experience.
[0006] Some embodiments of the present disclosure provide an information recommendation method, including:
[0007] Using the target user's item click history behavior sequence, query the first correlation information between different items to obtain an item set related to the target user's historically clicked items;
[0008] Using the target user's item click history behavior sequence, query the second correlation information between the item and the aggregated item, and obtain the aggregated item set related to the target user's historically clicked items;
[0009] Determining target items of the polymerized items in the polymerized items set for the target user by matching the polymerized items in the polymerized items set with the items in the item set;
[0010] The combination of the polymer items in the polymer item set and the target items is recommended to the target user.
[0011] In some embodiments, determining the target items of the aggregated items in the aggregated items set for the target user comprises:
[0012] Matching the polymer items in the polymer item set with items in the item set according to the attribute information satisfied by the target item of the polymer items set in the matching rule;
[0013] The items in the item set that match the polymer item in the polymer item set are determined as target items of the polymer item for the target user.
[0014] In some embodiments, recommending a combination of a polymer item in the polymer item set and a target item thereof to a target user comprises:
[0015] Calculating the weight of the combination of the aggregated items in the aggregated items set and the target items thereof, so as to represent the degree of interest of the target user in the combination;
[0016] Filter combinations based on the weights of each combination;
[0017] Recommend the selected combinations to target users.
[0018] In some embodiments, calculating the weight of a combination of a polymer item in the polymer item set and its target item includes:
[0019] For any combination of the polymer product in the polymer product set and its target product:
[0020] Acquire a weight of the target item in the combination from the first correlation information, which is used to represent the degree of interest of the target user in the target item in the combination;
[0021] Obtaining the weight of the aggregated items in the combination from the second correlation information, which is used to represent the degree of interest of the target user in the aggregated items in the combination;
[0022] The weight of the combination is calculated according to the weight of the target item in the combination and the weight of the aggregated items in the combination.
[0023] In some embodiments, it also includes:
[0024] According to the item and aggregate item click history behavior sequence of each user, determine the item feature matrix and the aggregate item feature matrix;
[0025] Determining first correlation information between different items according to the item feature matrix;
[0026] According to the item feature matrix and the polymer item feature matrix, second correlation information between the items and the polymer items is determined.
[0027] In some embodiments, determining the item feature matrix and the aggregate item feature matrix includes: using each user's item and aggregate item click history behavior sequence to train the language representation model to obtain the item feature matrix and the aggregate item feature matrix, wherein: the item feature matrix includes a feature vector for each item, and the aggregate item feature matrix includes a feature vector for each aggregate item.
[0028] In some embodiments, training the language representation model includes:
[0029] Divide the item and aggregate item click history behavior sequence of each user into multiple sessions, where the same session includes the object sequence clicked by the same user within a unit time, where the objects include items and aggregate items, and each session is used as a training sample;
[0030] The language representation model is trained using each training sample, wherein the training parameters include an item feature matrix and a polymer item feature matrix of an encoding layer of the language representation model;
[0031] Output the item feature matrix and aggregate item feature matrix of the encoding layer of the trained language representation model.
[0032] In some embodiments, training the language representation model using each training sample includes:
[0033] Input any first training sample in the same training batch into the language representation model, and the output layer of the language representation model outputs the object probability distribution of the target position in the first training sample based on the first vector,
[0034] Wherein: the first vector is obtained by concatenating the feature vectors of the items or aggregated items corresponding to the positive labels in each second training sample, and the second training samples are other training samples in the same training batch except the first training samples.
[0035] In some embodiments, the language representation model includes: a BERT model, an ELMo model, a GTP model, an XLNet model, or a T5 model.
[0036] In some embodiments, the first correlation information between different items includes: a weight of the second item relative to the first item, which is used to represent the degree of interest of the user in the second item when the user has clicked on the first item.
[0037] In some embodiments, determining first correlation information between different items includes:
[0038] determining a weight of the second item relative to the first item according to an inner product between a feature vector of the first item and a feature vector of the second item in the item feature matrix;
[0039] or,
[0040] The weight of the second item relative to the first item is determined according to the inner product between the feature vector of the first item and the feature vector of the second item in the item feature matrix and the posterior click rate of the second item.
[0041] In some embodiments, the second correlation information between the item and the aggregate item includes: a weight of the first aggregate item relative to the first item, which is used to represent the degree of interest of the user in the first aggregate item if the user has clicked on the first item.
[0042] In some embodiments, determining second correlation information between the item and the polymer item includes:
[0043] Determining a weight of the first polymerized item relative to the first item based on an inner product between a feature vector of the first item in the item feature matrix and a feature vector of the first polymerized item in the polymerized item feature matrix;
[0044] or,
[0045] The weight of the first aggregated item relative to the first item is determined according to the inner product between the feature vector of the first item in the item feature matrix and the feature vector of the first aggregated item in the aggregated item feature matrix and the posterior click rate of the first aggregated item.
[0046] In some embodiments, it also includes: before recommending to the target user, screening or sorting the combination of the polymer items in the polymer item set and the target items thereof according to one or more of the preset refinement rules or re-ranking rules.
[0047] In some embodiments, aggregate items include activities, stores, and aggregate pages.
[0048] Some embodiments of the present disclosure provide an information determination method, including:
[0049] According to the item and aggregate item click history behavior sequence of each user, determine the item feature matrix and the aggregate item feature matrix;
[0050] Determining first correlation information between different items according to the item feature matrix;
[0051] According to the item feature matrix and the polymer item feature matrix, second correlation information between the items and the polymer items is determined.
[0052] Some embodiments of the present disclosure provide an information recommendation device, including:
[0053] Memory; and
[0054] A processor is coupled to the memory, and the processor is configured to execute an information recommendation method based on instructions stored in the memory.
[0055] Some embodiments of the present disclosure provide an information determination device, including:
[0056] Memory; and
[0057] A processor is coupled to the memory, and the processor is configured to execute an information determination method based on instructions stored in the memory.
[0058] Some embodiments of the present disclosure provide a non-transitory computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps of the method of each embodiment are implemented.
[0059] Some embodiments of the present disclosure provide an information recommendation device, including:
[0060] An item set acquisition unit is configured to query first correlation information between different items by using the item click history behavior sequence of the target user, and obtain an item set related to the items clicked by the target user in history;
[0061] The aggregated item set acquisition unit is configured to query the second correlation information between the item and the aggregated item by using the item click history behavior sequence of the target user, and obtain the aggregated item set related to the item clicked by the target user in history;
[0062] a target item determination unit configured to determine a target item of the polymer items in the polymer item set for a target user by matching the polymer items in the polymer item set with the items in the item set;
[0063] The information recommendation unit is configured to recommend a combination of the polymer items in the polymer item set and the target items to the target user.
[0064] Some embodiments of the present disclosure provide an information determination device, including:
[0065] A feature determination unit is configured to determine an item feature matrix and an aggregate item feature matrix according to the item and aggregate item click history behavior sequence of each user;
[0066] a first correlation information determining unit, configured to determine first correlation information between different items according to the item feature matrix;
[0067] The second correlation information determination unit is configured to determine second correlation information between the items and the polymer items based on the item feature matrix and the polymer item feature matrix. BRIEF DESCRIPTION OF THE DRAWINGS
[0068] The following is a brief introduction to the drawings required for use in the embodiments or related technical descriptions. The present disclosure can be more clearly understood according to the following detailed description with reference to the drawings.
[0069] Obviously, the drawings described below are only some embodiments of the present disclosure, and a person skilled in the art can obtain other drawings based on these drawings without creative work.
[0070] Figure 1 Schematic diagram showing a polymer article and its target article according to some embodiments of the present disclosure.
[0071] Figure 2 A schematic diagram showing an information recommendation system according to some embodiments of the present disclosure.
[0072] Figure 3 A schematic diagram of the BERT language representation model according to some embodiments of the present disclosure is shown.
[0073] Figure 4 A schematic diagram showing an information determination method according to some embodiments of the present disclosure.
[0074] Figure 5 A schematic diagram showing an information recommendation method according to some embodiments of the present disclosure.
[0075] Figure 6 A schematic diagram showing an information determination device according to some embodiments of the present disclosure.
[0076] Figure 7 A schematic diagram showing an information recommendation device according to some embodiments of the present disclosure. DETAILED DESCRIPTION
[0077] The technical solutions in the embodiments of the present disclosure will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present disclosure.
[0078] Unless otherwise specified, descriptions such as “first” and “second” in the present disclosure are used to distinguish different objects and are not used to indicate meanings such as size or time sequence.
[0079] Figure 1 Schematic diagram of polymer products and target products of some embodiments of the present disclosure. The target product is one of the products in the polymer products and is the product that attracts traffic to the polymer product page. Therefore, it is also called a traffic-attracting product. Figure 1 As shown, the display format of aggregated items is usually: the image of the target item is used as the main image for display, and the description information related to the aggregated item is displayed below the main image. After the user clicks the image of the target item, the aggregated item page is entered, and the user can continue to browse the individual items on the aggregated item page. For example, aggregated items include but are not limited to collections of multiple items such as activities, stores, and aggregated pages.
[0080] Figure 2 A schematic diagram of an information recommendation system according to some embodiments of the present disclosure is shown. Figure 2 As shown, the information recommendation system 200 includes, for example, an identification acquisition module 210, a data acquisition module 220, an offline mining module 230, and an online recall module 240. In other words, the information recommendation system 200 includes an information determination device 600 and an information recommendation device 700. The information determination device 600 includes, for example, an offline mining module 230, and may also include an identification acquisition module 210 and a data acquisition module 220. The offline mining module 230 includes, for example, a feature determination unit 231, a first correlation information determination unit 232, and a second correlation information determination unit 233. The information recommendation device 700 includes, for example, an online recall module 240, and may also include an identification acquisition module 210 and a data acquisition module 220. The online recall module 240 includes, for example, an item set acquisition unit 241, an aggregate item set acquisition unit 242, a target item determination unit 243, and an information recommendation unit 244.
[0081] Below Figure 2 Describe the various modules and units in it.
[0082] The identification acquisition module 210 is configured to acquire the identification of the user browsing information in the e-commerce system (used to uniquely identify the user) and pass the identification to the data acquisition module 220. When the identification acquisition module 210 is applied to the online recall module 240, in the information flow scenario, the acquired user identification is passed to the data acquisition module 220 to trigger the online recall module 240 to recommend information. When the identification acquisition module 210 is applied to the offline mining module 230, in the information flow scenario or the search scenario, the acquired user identification can be passed to the data acquisition module 220, so that the offline mining module 230 determines the first correlation information between different items and the second correlation information between items and aggregated items based on the user's historical behavior.
[0083] The data collection module 220 is configured to collect the user's click history behavior data on the e-commerce system according to the user's identification and update it regularly. The click history behavior data includes item click history behavior sequence and aggregate item click history behavior sequence.
[0084] The offline mining module 230 is configured to mine first correlation information between different items and second correlation information between items and aggregated items based on the historical click behavior sequences of items and aggregated items of each user collected by the data collection module 220, and can also transmit the mined information to the network disk for use by the online recall module 240.
[0085] The offline mining module 230 includes, for example, a feature determination unit 231, a first correlation information determination unit 232, and a second correlation information determination unit 233. The first correlation information determination unit 232 and the second correlation information determination unit 233 may also be combined into one unit, referred to as a correlation information determination unit. The feature determination unit 231 is configured to determine an item feature matrix and an aggregate item feature matrix according to the item and aggregate item click history behavior sequence of each user. The first correlation information determination unit 232 is configured to determine the first correlation information between different items according to the item feature matrix. For example, the first correlation information between different items includes: the weight of the second item relative to the first item, which is used to characterize the degree of interest of the user in the second item when the user has clicked the first item. The first correlation information may be in the form of a vocabulary. The form of each record in the first correlation information vocabulary is, for example, <first item, second item, weight of the second item relative to the first item>. The second correlation information determination unit 233 is configured to determine the second correlation information between items and aggregate items according to the item feature matrix and the aggregate item feature matrix. The second correlation information between the item and the aggregated item includes: the weight of the first aggregated item relative to the first item, which is used to characterize the degree of interest of the user in the first aggregated item when the user has clicked on the first item. The second correlation information may be in the form of a word list. The form of each record in the second correlation information word list is, for example, <first item, first aggregated item, weight of the first aggregated item relative to the first item>.
[0086] The various units in the offline mining module 230 are described in detail below.
[0087] In some embodiments, the feature determination unit 231 is configured to train the language representation model using the click history behavior sequences of items and aggregated items of each user to obtain an item feature matrix and an aggregated item feature matrix, wherein: the item feature matrix includes a feature vector of each item, and the aggregated item feature matrix includes a feature vector of each aggregated item. In this case, the feature determination unit 231 is also called a pre-training unit.
[0088] The feature determination unit 231 is described in detail below from three aspects: data preprocessing, model training, and feature generation.
[0089] The data preprocessing part is described in detail below.
[0090] The historical click behavior sequence of each user's items and aggregated items is divided into multiple sessions. The same session includes a sequence of objects clicked by the same user within a unit time. The objects include items and aggregated items. Each session is used as a training sample.
[0091] For example, the user's historical click behavior sequence of items and aggregated items in the past 10 days is selected as training data, and the historical click behavior sequence of items and aggregated items in 10 days is divided into multiple sessions, for example, with 10 minutes as the boundary. The same session includes the items or aggregated items clicked by the same user within 10 minutes, and the session length (the number of items or aggregated items clicked by the user in one session) is approximately between 4 and 16. If the click time of two clicked objects is greater than 10 minutes apart, the two clicked objects are divided into two sessions. Furthermore, the same session can include the items or aggregated items clicked by the same user on the same access device within a unit time, so as to distinguish the click behavior of the same user ID on different access devices.
[0092] The model training part is explained in detail below.
[0093] The language representation model is trained using the training samples obtained above. The language representation model includes a coding layer, a hidden layer, and an output layer. The training parameters of the language representation model include: an item feature matrix and a polymer item feature matrix of the coding layer, a position vector matrix of the coding layer, linear transformation parameters such as weights and biases of the hidden layer, and linear transformation parameters such as weights and biases of the output layer.
[0094] Language representation models include, but are not limited to, BERT (Bidirectional Encoder Representation from Transformer), ELMo (Embeddings from Language Models), GTP (Generative Pre-Training), XLNet (a generalized autoregressive language model for learning unsupervised representation of text sequences), T5 (Transfer Text-to-Text Transformer), and other models.
[0095] The following uses the BERT model as an example to illustrate the language representation model and its training process.
[0096] like Figure 3 The BERT model shown, the Bert4Rec model, mainly includes three sub-networks: encoding layer, self-attention layer, and output layer.
[0097] The training process of the BERT model includes: for each training text, first replace some words in the text in the encoding layer, then encode the processed text into a vector, and then input it into the self-attention layer. Through the self-attention mechanism, the potential semantic information in the text is learned, and the replaced word labels are predicted in the output layer. Using each training text, the BERT model is continuously trained until the BERT model can accurately predict the replaced word labels.
[0098] Specifically to the present embodiment, as mentioned above, each training sample is a session, and the same session includes a sequence of objects clicked by the same user within a unit time, and the objects include articles and aggregated articles. Therefore, the training process of the BERT model in the present embodiment includes: for each session, in the encoding layer, an object (article or aggregated article) in the session is randomly replaced, and the replaced object is replaced by a mask, and the processed session is position-encoded and feature-encoded, and then input into the self-attention layer, and the potential semantic information in the session is learned through the self-attention mechanism, and the mask position is predicted in the output layer. Using each session for training, the BERT model is continuously trained until the BERT model can accurately predict the mask position of the replaced object. The accuracy of the BERT model prediction (projection) satisfies the preset conditions, and the preset conditions can be, for example, that the prediction accuracy reaches a certain threshold. Thereby, the trained BERT model can predict the object that the user may click at the current position based on the object clicked by the user before and / or the object clicked later.
[0099] Combine the following Figure 3 The BERT language representation model (BERT model for short) shown in the figure details the training process of the BERT model in this embodiment.
[0100] Encoding layer
[0101] The encoding layer generates a semantic vector for any clicked object in the click object sequence of the session.
[0102] First, in each session’s click object sequence, one of the objects (items or aggregated items) is randomly replaced, and the replaced object v t Use mask instead.
[0103] Then, the encoding layer generates a specific position vector for each clicked object (item or aggregated object) in the click object sequence of the session to represent the click time interval between the clicked objects in the click object sequence of the session. i , by querying the position vector matrix P, we get v i The corresponding position vector P i For example, for v t , by querying the position vector matrix P, we get v t The corresponding position vector P t Among them, the position vector matrix P is one of the parameters that need to be trained, and its value can be randomly initialized before training. The position vector matrix P can be set using the tf.Variable function of the Tensorflow framework.
[0104] Then, the encoding layer generates a feature vector for each clicked object (item or aggregated object) in the clicked object sequence of the session. In some embodiments, for any clicked object v in the session, i , if v i is an item, by querying the item feature matrix V S , we get v i The corresponding eigenvector V i ; if v i It is a polymer product. By querying the polymer product feature matrix V A , we get v i The corresponding eigenvector V i Among them, the item feature matrix V S and polymer product characteristic matrix V A It is a parameter that needs to be trained, and its value can be randomly initialized before training. S and polymer product characteristic matrix V A It can be set using the tf.Variable function of the Tensorflow framework.
[0105] Finally, for any clicked object v in the session i , v i The corresponding position vector P i and v i The corresponding eigenvector V i Add them together to get v i The semantic vector E at the encoding layer i .
[0106] Self-Attention Layer
[0107] The self-attention layer captures the correlation between different objects in the click object sequence of the same session through the self-attention mechanism, including the correlation between different items (first correlation information) and the correlation between items and aggregated items (second correlation information).
[0108] In order to fully extract the semantic information between different objects, the self-attention layer usually includes multiple layers, and each layer structure can be defined as Trm l Each self-attention layer contains two sub-modules: a multi-head self-attention layer and a feed-forward neural network layer. After each sub-module, a residual connection and layer normalization module can be connected.
[0109] (a) Multi-head self-attention layer
[0110] The input vector of the current self-attention layer is It is the output vector of the previous self-attention layer. It maps the input vector to multiple independent feature subspaces through a linear transformation mechanism, and then captures the semantic information between different dimensions through multiple self-attention heads.
[0111] First, for the i-th self-attention head, the input vector is mapped to three independent subspaces through three different linear transformation matrices, that is, the input vector is multiplied by three different linear transformation matrices to obtain three linearly transformed output vectors.
[0112] That is, assuming: are three linear transformation matrices, take Represents the output vector after linear transformation.
[0113] Then, dot product attention calculation is performed on the output feature vectors Q, K, and V. The attention calculation method includes, for example: performing a dot product operation on the feature vector Q and the transpose of the feature vector K, calculating the quotient of the dimension of the input vector and the number of self-attention heads, calculating the square root of the quotient, performing a softmax operation on the quotient of the dot product operation result and the square root result, obtaining the attention weights between different objects in the click object sequence of the session, performing a dot product operation on the attention weights and the feature vector V, obtaining the attention between different objects in the click object sequence of the session, and using the output vector of the i-th self-attention head as the output vector.
[0114] The calculation formula for the output vector of the i-th self-attention head is as follows:
[0115]
[0116] Among them: softmax(*) represents the attention weights between different objects in the click object sequence of the session, and d represents the input vector H l The dimension of h represents the number of self-attention heads.
[0117] Multiple self-attention heads in the self-attention layer are calculated in parallel, and each self-attention head is based on the input vector H l The corresponding output vector head is obtained i .
[0118] Finally, the output vectors of multiple self-attention heads of the same layer are concatenated and linearly transformed using the linear transformation matrix to obtain the output of the multi-head self-attention layer of this layer. The formula is expressed as:
[0119] MultiHead l (H l ,H l ,H l )=Concat(head1,head2,…,head h )W o
[0120] Among them, head i represents the output vector of the i-th attention head, where i ranges from 1,…,h, h represents the number of attention heads, and Concat() represents concatenating the output vectors of each self-attention head by dimension. O represents the linear transformation matrix, H l Represents the input vector of the lth self-attention layer, MultiHead l Represents the output of the l-th multi-head self-attention layer.
[0121] (b) Feedforward neural network layer
[0122] In order to ensure the nonlinearity of the network structure and the correlation between different dimensions of the feature vector, a feed-forward neural network is added after the multi-head attention layer to perform a linear transformation on the output of the multi-head attention layer, wherein: linear transformation parameters such as weights and biases are used for linear transformation; the linear transformation can be performed multiple times, for example, twice.
[0123] For example, the feedforward neural network uses the linear transformation parameters in the first layer of the feedforward neural network to perform the first linear transformation on the output of the connected multi-head attention layer. The first linear transformation result is used by the nonlinear activation function, and the linear transformation parameters in the second layer of the feedforward neural network are used to perform the second linear transformation on the above result to produce the output of the feedforward neural network. The calculation formula is as follows:
[0124] FFN l (x) = GELU(xW (1) +b (1) )W (2) +b (2)
[0125] Among them, x represents the output of the multi-head self-attention layer, FFN l (x) represents the output of the feedforward neural network, W (1) ,b (1) Represents the linear transformation parameters in the first layer of the feedforward neural network, which are weights and biases, W (2) ,b (2) represents the linear transformation parameters in the second layer of feedforward neural network, which are weight and bias respectively, and GELU represents the nonlinear activation function in the first layer of feedforward neural network.
[0126] (c) Residual connection and layer normalization module
[0127] In order to accelerate model convergence and ensure the stability of data distribution during model training, residual connection and layer normalization modules are added after the multi-head self-attention layer and the feedforward neural network layer, respectively. This module performs residual connection and normalizes the residual connection result, and finally obtains the output H of the l-th self-attention layer l+1 , the calculation process is as follows:
[0128] f out (x)=LayerNorm(x+SubLayer(x))
[0129] Where x represents the input of the multi-head self-attention layer or the feedforward neural network layer, SubLayer(x) represents the output of the multi-head self-attention layer or the feedforward neural network layer, x+SubLayer(x) represents the residual connection, LayerNorm(*) represents layer normalization, that is, normalizing all neurons in an intermediate layer, and f out (x) represents the output of the residual connection and layer normalization module, where: When x represents the input of the feedforward neural network layer, f out (x) is the output H of the lth self-attention layer l+1 .
[0130] Output Layer
[0131] The output layer is used to predict the mask position. The output layer is normalized by the softmax function through a nonlinear feedforward neural network to obtain the object probability distribution of the target position t in the click object sequence of the session. For example, the output of the last self-attention layer is linearly transformed using the linear transformation parameters of the output layer, and then the linear transformation result is processed using the nonlinear activation function in the output layer. The processed result is multiplied with the learnable item feature matrix and the aggregate item feature matrix in the encoding layer, and the sum of the multiplication result and the bias term is normalized using the softmax function to obtain the object probability distribution of the target position t in the click object sequence of the session. The calculation formula is:
[0132]
[0133] in, represents the output of the last (Lth) self-attention layer, W P ,b P represents the linear transformation parameter of the output layer, GELU represents the nonlinear activation function in the output layer, and V S 、V A They represent the learnable item feature matrix and aggregate item feature matrix in the encoding layer, respectively, and b O represents the bias term, and P(v) represents the probability distribution.
[0134] If the number of object types involved in the session is relatively large, for example, an e-commerce system includes hundreds of millions of items and aggregated items, considering that directly performing multiple classification on hundreds of millions of objects will bring huge computing overhead, it will seriously reduce the accuracy of model estimation. In order to improve the efficiency of model training, ensure that the model can run stably on large-scale training sets (such as hundreds of millions), and prevent the model from exceeding the memory limit during training, this embodiment inputs any first training sample in the same training batch into the language representation model, and the output layer of the language representation model is based on the first vector Output the object probability distribution of the target position in the first training sample, where: the first vector The second training samples are obtained by concatenating the feature vectors of the items or aggregated items corresponding to the positive labels in each second training sample, and the second training samples are other training samples in the same training batch except the first training samples.
[0135] That is, the output of the last self-attention layer is linearly transformed using the linear transformation parameters of the output layer, and then the linear transformation result is processed using the nonlinear activation function in the output layer, and the processed result is combined with the first vector Multiply, and use the softmax function to normalize the sum of the multiplication result and the bias term to obtain the object probability distribution of the target position t in the click object sequence of the session. The calculation formula is improved to:
[0136]
[0137] According to the above method, the BERT model is continuously trained using each session as a training sample until the BERT model can accurately predict the mask position of the replaced object. Then, the feature generation part is performed.
[0138] The feature generation part is described in detail below.
[0139] After the language representation model training is completed, the item feature matrix and aggregate item feature matrix of the encoding layer of the trained language representation model are output and can be saved in an offline file. Among them, the item feature matrix records the mapping relationship between each item and its feature vector, and the aggregate item feature matrix records the mapping relationship between each aggregate item and its feature vector.
[0140] The correlation information determination units (the first correlation information determination unit 232, the second correlation information determination unit 233) are described in detail below.
[0141] The first correlation information determination unit 232 is configured to determine the first correlation information between different items according to the item feature matrix. The first correlation information between different items includes: the weight of the second item relative to the first item, which is used to characterize the degree of interest of the user in the second item when the user has clicked on the first item. Two methods for determining the first correlation information are given below. The first method: determine the weight of the second item relative to the first item according to the inner product between the feature vector of the first item and the feature vector of the second item in the item feature matrix. For example, the inner product is used as the weight of the second item relative to the first item. The second method: determine the weight of the second item relative to the first item according to the inner product between the feature vector of the first item and the feature vector of the second item in the item feature matrix and the posterior click rate of the second item. For example, the product of the inner product and the posterior click rate of the second item is used as the weight of the second item relative to the first item. In addition, as mentioned above, one of the records of the first correlation information in the word list form is, for example, <first item, second item, weight of the second item relative to the first item>.
[0142] The second correlation information determination unit 233 is configured to determine the second correlation information between the item and the aggregated item according to the item feature matrix and the aggregated item feature matrix. The second correlation information between the item and the aggregated item includes: the weight of the first aggregated item relative to the first item, which is used to characterize the degree of interest of the user in the first aggregated item when the first item has been clicked. Two methods for determining the second correlation information are given below. The first method: determine the weight of the first aggregated item relative to the first item according to the inner product between the feature vector of the first item in the item feature matrix and the feature vector of the first aggregated item in the aggregated item feature matrix. For example, the inner product is used as the weight of the first aggregated item relative to the first item. The second method: determine the weight of the first aggregated item relative to the first item according to the inner product between the feature vector of the first item in the item feature matrix and the feature vector of the first aggregated item in the aggregated item feature matrix and the posterior click rate of the first aggregated item. For example, the product of the inner product and the posterior click rate of the first aggregated item is used as the weight of the first aggregated item relative to the first item. In addition, as mentioned above, one of the records of the second correlation information in the word list form is, for example, <first item, first aggregated item, weight of the first aggregated item relative to the first item>.
[0143] The inner product between the two vectors mentioned above is calculated by calculating the product of the first vector and the second vector, taking the modulus, and then dividing it by the product of the modulus of the first vector and the modulus of the second vector, and taking the final quotient as the inner product between the first vector and the second vector. In addition, the inner product between two vectors can be calculated using the Faiss tool.
[0144] In some embodiments, the first correlation information and the second correlation information may also be filtered. For example, according to a set weight threshold, the first correlation information and the second correlation information with weights higher than the threshold may be filtered out, thereby filtering out items or aggregate items of relatively high user interest.
[0145] The online recall module 240 and its involved item set acquisition unit 241 , aggregate item set acquisition unit 242 , target item determination unit 243 , and information recommendation unit 244 are described in detail below.
[0146] Based on the first correlation information between different items and the second correlation information between items and aggregate items generated by the offline mining module 230, the online recall module 240 accurately locates the aggregate items and combinations of target items that each target user is interested in, and makes personalized combination recommendations for each target user.
[0147] The item set acquisition unit 241 is configured to use the target user's item click history behavior sequence to query the first correlation information between different items, obtain an item set related to the target user's historically clicked items, and also obtain the weights of the items in the item set relative to the target user's historically clicked items. Thus, the item set that the target user may be interested in is recalled, and the user's interest in the recalled items is obtained.
[0148] That is, the items historically clicked by the target user are taken as the first items in the first correlation information, and the first correlation information is queried. The second items in the queried first correlation information and the weight of the second items relative to the first item are taken as the set of items related to the items historically clicked by the target user, and the weight of each item in the set of items.
[0149] For example, the set of items that the target user may be interested in is {sku1,sku2,…,sku m}, the target users’ interest in each recalled item is: {sku1_weight,sku2_weight,…,sku m _weight}.
[0150] The aggregated item set acquisition unit 242 is configured to use the target user's item click history behavior sequence to query the second correlation information between the item and the aggregated item, obtain the aggregated item set related to the target user's historically clicked items, and also obtain the weight of the aggregated items in the aggregated item set relative to the target user's historically clicked items. Thus, the aggregated item set that the target user may be interested in is recalled, and the user's interest in the recalled aggregated items is obtained.
[0151] That is, the items historically clicked by the target user are taken as the first items in the second correlation information, and the second correlation information is queried. The first aggregated items in the queried second correlation information and the weight of the first aggregated items relative to the first item are taken as the aggregated item set related to the items historically clicked by the target user, and the weight of each aggregated item in the aggregated item set.
[0152] For example, the set of aggregate items (such as activities) that the target user may be interested in is {activity1,activity2,…,activity n}, the target users’ interest in each of the recalled polymer products is:
[0153] {activity1_weight,activity2_weight,…,activity m _weight}.
[0154] The target item determination unit 243 is configured to determine the target item of the aggregated items in the aggregated item set for the target user by matching the aggregated items in the aggregated item set with the items in the item set.
[0155] In some embodiments, the target item determination unit 243 is configured to match the polymer items in the polymer item set with the items in the item set according to the attribute information satisfied by the target items of the polymer items set in the matching rule; and determine the items in the item set that match the polymer items in the polymer item set as the target items of the polymer items for the target user.
[0156] Among them, the attribute information satisfied by the target item of the polymer items set by the matching rule includes, but is not limited to, one or more of the store information, category information, brand information, etc. to which the target item should belong.
[0157] In some embodiments, before recommending the combination of the polymer items in the polymer item set and the target items to the target user, the weight of the combination of the polymer items in the polymer item set and the target items is calculated to characterize the degree of interest of the target user in the combination; the combination is screened according to the weight of each combination; and the screened combination is then recommended to the target user. This process can be performed by the target item determination unit 243, or the information recommendation unit 244, or another separate unit. Thus, the combination of the polymer items and the target items that the target user is more interested in is recommended to the target user.
[0158] In some embodiments, calculating the weight of a combination of the polymer items in the set of polymer items and their target items includes: for any combination of the polymer items in the set of polymer items and their target items: obtaining the weight of the target item in the combination from the first correlation information, which is used to characterize the degree of interest of the target user in the target item in the combination; obtaining the weight of the polymer items in the combination from the second correlation information, which is used to characterize the degree of interest of the target user in the polymer items in the combination; and calculating the weight of the combination based on the weight of the target item in the combination and the weight of the polymer items in the combination.
[0159] Wherein, obtaining the weight of the target item in the combination from the first correlation information includes: querying the record that the second item is the target item in the combination from the first correlation information, and the weight of the target item relative to the first item in the record is the weight of the target item.
[0160] Among them, obtaining the weight of the polymer item in the combination from the second correlation information includes: querying the record that the first polymer item is a polymer item in the combination from the second correlation information, and the weight of the polymer item relative to the first item in the record is the weight of the polymer item.
[0161] According to the weight of the target item in the combination and the weight of the aggregated items in the combination, calculating the weight of the combination includes: multiplying the weight of the target item in the combination by the weight of the aggregated items in the combination to determine the weight of the combination. Thus, the user's interest in the aggregated items and the user's interest in the target item are considered at the same time, and the user's interest in the aggregated items as a whole is determined more accurately.
[0162] The information recommendation unit 244 is configured to recommend the aggregated items and the combination of the target items in the aggregated item set to the target user. Thus, the aggregated items and the combination of the target items that the target user is interested in are recommended to the target user, that is, different aggregated items and the combination of the target items are recommended to users with different interests, so as to achieve the personalized information recommendation effect and improve the user experience.
[0163] In some embodiments, the information recommendation unit 244 is configured to filter or sort the combination of the polymer items in the polymer item set and the target items according to one or more of the preset fine sorting rules or rearrangement rules before recommending to the target user, and then recommend the filtered or sorted results to the target user.
[0164] For example, the refined ranking rule is: score the combination of the aggregate item and its target item according to the predicted click-through rate (pCTR). The higher the predicted click-through rate, the higher the score, the higher the ranking, and the combination can be selected for recommendation and display first.
[0165] For another example, the precise ranking rule is to score the combination of polymer items and their target items according to the predicted revenue. The predicted revenue is determined based on the predicted click-through rate and the cost per click. For example, the product of the predicted click-through rate and the cost per click is used as the predicted revenue. The higher the predicted revenue, the higher the score, the higher the ranking, and the combination can be selected for recommendation and display first.
[0166] For another example, the rearrangement rule is to score the combination of the aggregated item and its target item according to the rearrangement score. The higher the score, the higher the ranking, and the combination can be selected for recommendation and display. The rearrangement score is determined, for example, based on at least one of the novelty score and the window rule score. This takes into account both accuracy and richness. k Probability distribution P(w k |S) and the i-th item s in the set S i Product word w k Probability distribution P(w k |s i ), and calculate the novelty score based on the Kullback-Leibler distance. Among them, the diversity constraint can be used as a window rule to limit the frequency of occurrence of objects of the same category / same brand / same store / same material in the same window. The window rule can include one or more diversity constraints, and the weights of each diversity constraint can be different. The weights of the diversity constraints can be set according to importance. The sum of the weights of each diversity constraint satisfied by the object can be used as the window rule score of the object.
[0167] Figure 4 A schematic diagram of an information determination method according to some embodiments of the present disclosure is shown, wherein the determined information includes first correlation information between different items and second correlation information between an item and a polymer item.
[0168] like Figure 4 As shown, the information determination method includes the following steps.
[0169] In step 410, an item feature matrix and an aggregate item feature matrix are determined based on the item and aggregate item click history behavior sequences of each user.
[0170] In some embodiments, the language representation model is trained using the historical click behavior sequences of each user's items and aggregated items to obtain an item feature matrix and an aggregated item feature matrix, wherein the item feature matrix includes a feature vector for each item, and the aggregated item feature matrix includes a feature vector for each aggregated item.
[0171] In some embodiments, using the item and aggregate item click history behavior sequences of each user to train the language representation model includes:
[0172] Divide the item and aggregate item click history behavior sequence of each user into multiple sessions, where the same session includes the object sequence clicked by the same user within a unit time, where the objects include items and aggregate items, and each session is used as a training sample;
[0173] The language representation model is trained using each training sample, wherein the training parameters include an item feature matrix and a polymer item feature matrix of an encoding layer of the language representation model;
[0174] Output the item feature matrix and aggregate item feature matrix of the encoding layer of the trained language representation model.
[0175] In some embodiments, training a language representation model using various training samples includes: inputting any first training sample in the same training batch into the language representation model, and the output layer of the language representation model outputs the object probability distribution of the target position in the first training sample based on the first vector, wherein: the first vector is obtained by concatenating the feature vectors of the items or aggregate items corresponding to the positive labels in each second training sample, and the second training samples are other training samples in the same training batch except the first training samples.
[0176] This improves the efficiency of model training, ensures that the model can run stably on large-scale training sets (such as billions), and prevents the model from exceeding memory limits during training.
[0177] In step 420, first correlation information between different items is determined based on the item feature matrix.
[0178] The first correlation information between different items includes: the weight of the second item relative to the first item, which is used to represent the degree of interest of the user in the second item when the user has clicked on the first item.
[0179] In some embodiments, the weight of the second item relative to the first item is determined based on the inner product between the feature vector of the first item and the feature vector of the second item in the item feature matrix. For example, the inner product is used as the weight of the second item relative to the first item.
[0180] In some embodiments, the weight of the second item relative to the first item is determined based on the inner product between the feature vector of the first item and the feature vector of the second item in the item feature matrix and the posterior click rate of the second item. For example, the product of the inner product and the posterior click rate of the second item is used as the weight of the second item relative to the first item.
[0181] In step 430, second correlation information between the items and the polymer items is determined based on the item feature matrix and the polymer item feature matrix.
[0182] The second correlation information between the item and the aggregated item includes: a weight of the first aggregated item relative to the first item, which is used to represent the degree of interest of the user in the first aggregated item when the user has clicked on the first item.
[0183] In some embodiments, the weight of the first polymer item relative to the first item is determined based on the inner product between the feature vector of the first item in the item feature matrix and the feature vector of the first polymer item in the polymer item feature matrix. For example, the inner product is used as the weight of the first polymer item relative to the first item.
[0184] In some embodiments, the weight of the first aggregated item relative to the first item is determined based on the inner product between the feature vector of the first item in the item feature matrix and the feature vector of the first aggregated item in the aggregated item feature matrix and the posterior click rate of the first aggregated item. For example, the product of the inner product and the posterior click rate of the first aggregated item is used as the weight of the first aggregated item relative to the first item.
[0185] Therefore, based on the click history of each user's items and aggregated items, the first correlation information between different items and the second correlation information between items and aggregated items are mined to provide support for personalized information recommendation. In addition, the training process is optimized to improve the efficiency of model training, ensure that the model can run stably on large-scale training sets (such as hundreds of millions), and prevent the model from exceeding the memory limit during training.
[0186] Figure 5 A schematic diagram showing an information recommendation method according to some embodiments of the present disclosure, wherein the recommended information includes a combination of a polymer item and a target item.
[0187] like Figure 5 As shown, the information recommendation method includes the following steps.
[0188] In step 510, the first correlation information between different items is queried using the target user's item click history behavior sequence, and an item set related to the target user's historically clicked items is obtained, and the weights of the items in the item set relative to the target user's historically clicked items can also be obtained. Thus, the item set that the target user may be interested in is recalled, and the user's interest in the recalled items is obtained.
[0189] That is, the items historically clicked by the target user are taken as the first items in the first correlation information, and the first correlation information is queried. The second items in the queried first correlation information and the weight of the second items relative to the first item are taken as the set of items related to the items historically clicked by the target user, and the weight of each item in the set of items.
[0190] In step 520, the second correlation information between the items and the aggregated items is queried using the target user's item click history behavior sequence, and an aggregated item set related to the target user's historically clicked items is obtained, and the weight of the aggregated items in the aggregated item set relative to the target user's historically clicked items can also be obtained. Thus, the aggregated item set that the target user may be interested in is recalled, and the user's interest in the recalled aggregated items is obtained.
[0191] That is, the items historically clicked by the target user are taken as the first items in the second correlation information, and the second correlation information is queried. The first aggregated items in the queried second correlation information and the weight of the first aggregated items relative to the first item are taken as the aggregated item set related to the items historically clicked by the target user, and the weight of each aggregated item in the aggregated item set.
[0192] In step 530 , target items of the aggregated items in the aggregated item set for the target user are determined by matching the aggregated items in the aggregated item set with the items in the item set.
[0193] In some embodiments, the polymer items in the polymer item set are matched with items in the item set according to the attribute information satisfied by the target items of the polymer items set according to the matching rules; and the items in the item set that match the polymer items in the polymer item set are determined as the target items of the polymer items for the target user.
[0194] Among them, the attribute information satisfied by the target item of the polymer items set by the matching rule includes, but is not limited to, one or more of the store information, category information, brand information, etc. to which the target item should belong.
[0195] In step 540 , a combination of the aggregated items in the aggregated item set and the target items is recommended to the target user.
[0196] In some embodiments, the weights of the combinations of the polymer items and their target items in the polymer item set are calculated to represent the degree of interest of the target user in the combinations; combinations are screened according to the weights of the combinations; and the screened combinations are recommended to the target user. Thus, the combinations of the polymer items and their target items that the target user is more interested in are recommended to the target user.
[0197] Calculating the weight of the combination of the polymer item in the polymer item set and the target item includes:
[0198] For any combination of the polymer product in the polymer product set and its target product:
[0199] Acquire a weight of the target item in the combination from the first correlation information, which is used to represent the degree of interest of the target user in the target item in the combination;
[0200] Obtaining the weight of the aggregated items in the combination from the second correlation information, which is used to represent the degree of interest of the target user in the aggregated items in the combination;
[0201] The weight of the combination is calculated according to the weight of the target item in the combination and the weight of the aggregated items in the combination. For example, the weight of the combination is determined as the product of the weight of the target item in the combination and the weight of the aggregated items in the combination. Thus, the user's interest in the aggregated items and the user's interest in the target item are considered at the same time, and the user's interest in the aggregated items as a whole is determined more accurately.
[0202] In some embodiments, before recommending to the target user, the combination of the polymer items in the polymer item set and the target items is screened or sorted according to one or more of the preset refinement rules or rearrangement rules. The refinement rules and rearrangement rules are specifically referred to above and will not be described in detail here.
[0203] Thus, the aggregate items and target items that each user is interested in are accurately mined and located, and personalized combinations of aggregate items and target items that each user is interested in are recommended, and different combinations of aggregate items and target items are recommended to users with different interests, so as to achieve personalized information recommendation effects and improve user experience. In addition, the user's interest in the aggregate items and the user's interest in the target items are considered at the same time, so as to more accurately determine the user's overall interest in the aggregate items.
[0204] Figure 6 A schematic diagram showing an information determination device according to some embodiments of the present disclosure.
[0205] like Figure 6As shown, the information determination device 600 of this embodiment includes: a memory 610 and a processor 620 coupled to the memory 610, and the processor 620 is configured to execute the information determination method in any of the aforementioned embodiments based on the instructions stored in the memory 610.
[0206] The device 600 may further include an input / output interface 630 , a network interface 640 , a storage interface 650 , etc. These interfaces 630 , 640 , 650 , the memory 610 , and the processor 620 may be connected via a bus 660 , for example.
[0207] Figure 7 A schematic diagram showing an information recommendation device according to some embodiments of the present disclosure.
[0208] like Figure 7 As shown, the information recommendation device 700 of this embodiment includes: a memory 710 and a processor 720 coupled to the memory 710 , and the processor 720 is configured to execute the information recommendation method in any of the aforementioned embodiments based on instructions stored in the memory 710 .
[0209] The device 700 may further include an input / output interface 730 , a network interface 740 , a storage interface 750 , etc. These interfaces 730 , 740 , 750 , the memory 710 , and the processor 720 may be connected via a bus 760 , for example.
[0210] The memory 610, 710 may include, for example, a system memory, a fixed non-volatile storage medium, etc. The system memory may store, for example, an operating system, an application program, a boot loader, and other programs.
[0211] Among them, the processors 620, 720 can be implemented by discrete hardware components such as a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistors.
[0212] Among them, the input and output interfaces 630, 730 provide connection interfaces for input and output devices such as displays, mice, keyboards, touch screens, etc. The network interfaces 640, 740 provide connection interfaces for various networked devices. The storage interfaces 650, 750 provide connection interfaces for external storage devices such as SD cards and USB flash drives. The buses 660, 760 can use any of a variety of bus structures. For example, the bus structure includes but is not limited to the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MCA) bus, and the Peripheral Component Interconnect (PCI) bus.
[0213] The embodiment of the present disclosure provides a non-transitory computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the information recommendation method and / or the information determination method are implemented.
[0214] (1) An information recommendation method, comprising:
[0215] Using the target user's item click history behavior sequence, query the first correlation information between different items to obtain an item set related to the target user's historically clicked items;
[0216] Using the target user's item click history behavior sequence, query the second correlation information between the item and the aggregated item, and obtain the aggregated item set related to the target user's historically clicked items;
[0217] Determining target items of the polymerized items in the polymerized items set for the target user by matching the polymerized items in the polymerized items set with the items in the item set;
[0218] The combination of the polymer items in the polymer item set and the target items is recommended to the target user.
[0219] (2) According to (1), determining the target items of the aggregated items in the aggregated items set for the target user includes:
[0220] Matching the polymer items in the polymer item set with items in the item set according to the attribute information satisfied by the target item of the polymer items set in the matching rule;
[0221] An item in the item set that matches the polymer item in the polymer item set is determined as a target item of the polymer item for a target user.
[0222] (3) According to (1) or (2), recommending to a target user a combination of a polymer item in the polymer item set and a target item thereof comprises:
[0223] Calculating the weight of the combination of the aggregated items in the aggregated items set and the target items thereof, so as to represent the degree of interest of the target user in the combination;
[0224] Filter combinations based on the weights of each combination;
[0225] Recommend the selected combinations to target users.
[0226] (4) According to (3), calculating the weight of the combination of the polymer product in the polymer product set and the target product includes:
[0227] For any combination of the polymer product in the polymer product set and its target product:
[0228] Acquire a weight of the target item in the combination from the first correlation information, which is used to represent the degree of interest of the target user in the target item in the combination;
[0229] Obtaining the weight of the aggregated items in the combination from the second correlation information, which is used to represent the degree of interest of the target user in the aggregated items in the combination;
[0230] The weight of the combination is calculated according to the weight of the target item in the combination and the weight of the aggregated items in the combination.
[0231] (5) According to (4), calculating the weight of the combination includes: multiplying the weight of the target item in the combination by the weight of the aggregate item in the combination to determine the weight of the combination.
[0232] (6) According to any one of (1) to (5), further comprising:
[0233] According to the item and aggregate item click history behavior sequence of each user, determine the item feature matrix and the aggregate item feature matrix;
[0234] Determining first correlation information between different items according to the item feature matrix;
[0235] According to the item feature matrix and the polymer item feature matrix, second correlation information between the items and the polymer items is determined.
[0236] (7) According to (6), determining the item feature matrix and the aggregate item feature matrix includes: using the item and aggregate item click history behavior sequences of each user to train the language representation model to obtain the item feature matrix and the aggregate item feature matrix, wherein: the item feature matrix includes the feature vector of each item, and the aggregate item feature matrix includes the feature vector of each aggregate item.
[0237] (8) According to (7), training the language representation model includes:
[0238] Divide the item and aggregate item click history behavior sequence of each user into multiple sessions, where the same session includes the object sequence clicked by the same user within a unit time, where the objects include items and aggregate items, and each session is used as a training sample;
[0239] The language representation model is trained using each training sample, wherein the training parameters include an item feature matrix and a polymer item feature matrix of an encoding layer of the language representation model;
[0240] Output the item feature matrix and aggregate item feature matrix of the encoding layer of the trained language representation model.
[0241] (9) According to (8), using each training sample to train the language representation model includes: inputting any first training sample in the same training batch into the language representation model, and the output layer of the language representation model outputs the object probability distribution of the target position in the first training sample based on the first vector, wherein: the first vector is obtained by concatenating the feature vectors of the objects or aggregated objects corresponding to the positive labels in each second training sample, and the second training samples are other training samples in the same training batch except the first training samples.
[0242] (10) According to any one of (7) to (9), the language representation model includes: a BERT model, an ELMo model, a GTP model, an XLNet model, or a T5 model.
[0243] (11) Any of (6) to (10):
[0244] The first correlation information between different items includes: a weight of the second item relative to the first item, which is used to represent the degree of interest of the user in the second item when the user has clicked on the first item;
[0245] Alternatively, determining the first correlation information between different items includes:
[0246] determining a weight of the second item relative to the first item according to an inner product between a feature vector of the first item and a feature vector of the second item in the item feature matrix;
[0247] or,
[0248] The weight of the second item relative to the first item is determined according to the inner product between the feature vector of the first item and the feature vector of the second item in the item feature matrix and the posterior click rate of the second item.
[0249] (12) Any of (6) to (11) above:
[0250] The second correlation information between the item and the aggregated item includes: a weight of the first aggregated item relative to the first item, which is used to represent the degree of interest of the user in the first aggregated item if the user has clicked on the first item;
[0251] Alternatively, determining the second correlation information between the item and the polymer item includes:
[0252] determining a weight of the first polymerized item relative to the first item based on an inner product between a feature vector of the first item in the item feature matrix and a feature vector of the first polymerized item in the polymerized item feature matrix;
[0253] or,
[0254] The weight of the first aggregated item relative to the first item is determined according to the inner product between the feature vector of the first item in the item feature matrix and the feature vector of the first aggregated item in the aggregated item feature matrix and the posterior click rate of the first aggregated item.
[0255] (13) According to any one of items (1) to (12), it also includes: before recommending to the target user, screening or sorting the combination of the polymer items in the polymer item set and the target items thereof according to one or more of the preset refinement rules or re-ranking rules.
[0256] (14) According to any one of (1) to (13), the aggregated items include, but are not limited to, activities, stores, and aggregated pages.
[0257] (15) An information determination method comprising:
[0258] According to the item and aggregate item click history behavior sequence of each user, determine the item feature matrix and the aggregate item feature matrix;
[0259] Determining first correlation information between different items according to the item feature matrix;
[0260] According to the item feature matrix and the polymer item feature matrix, second correlation information between the items and the polymer items is determined.
[0261] (16) According to (15), determining the item feature matrix and the aggregate item feature matrix includes: using the item and aggregate item click history behavior sequences of each user to train the language representation model to obtain the item feature matrix and the aggregate item feature matrix, wherein: the item feature matrix includes the feature vector of each item, and the aggregate item feature matrix includes the feature vector of each aggregate item.
[0262] (17) According to (16), training the language representation model includes:
[0263] Divide the item and aggregate item click history behavior sequence of each user into multiple sessions, where the same session includes the object sequence clicked by the same user within a unit time, where the objects include items and aggregate items, and each session is used as a training sample;
[0264] The language representation model is trained using each training sample, wherein the training parameters include an item feature matrix and a polymer item feature matrix of an encoding layer of the language representation model;
[0265] Output the item feature matrix and aggregate item feature matrix of the encoding layer of the trained language representation model.
[0266] (18) According to (17), using each training sample to train the language representation model includes:
[0267] Input any first training sample in the same training batch into the language representation model, and the output layer of the language representation model outputs the object probability distribution of the target position in the first training sample based on the first vector,
[0268] Wherein: the first vector is obtained by concatenating the feature vectors of the items or aggregated items corresponding to the positive labels in each second training sample, and the second training samples are other training samples in the same training batch except the first training samples.
[0269] (19) According to any one of (16) to (18), the language representation model includes: a BERT model, an ELMo model, a GTP model, an XLNet model, or a T5 model.
[0270] (20) According to any of (15) to (19):
[0271] The first correlation information between different items includes: a weight of the second item relative to the first item, which is used to represent the degree of interest of the user in the second item when the user has clicked on the first item;
[0272] Alternatively, determining the first correlation information between different items includes:
[0273] determining a weight of the second item relative to the first item according to an inner product between a feature vector of the first item and a feature vector of the second item in the item feature matrix;
[0274] or,
[0275] The weight of the second item relative to the first item is determined according to the inner product between the feature vector of the first item and the feature vector of the second item in the item feature matrix and the posterior click rate of the second item.
[0276] (21) According to any of (15) to (20):
[0277] The second correlation information between the item and the aggregated item includes: a weight of the first aggregated item relative to the first item, which is used to represent the degree of interest of the user in the first aggregated item if the user has clicked on the first item;
[0278] Alternatively, determining the second correlation information between the item and the polymer item includes:
[0279] determining a weight of the first polymerized item relative to the first item based on an inner product between a feature vector of the first item in the item feature matrix and a feature vector of the first polymerized item in the polymerized item feature matrix;
[0280] or,
[0281] The weight of the first aggregated item relative to the first item is determined according to the inner product between the feature vector of the first item in the item feature matrix and the feature vector of the first aggregated item in the aggregated item feature matrix and the posterior click rate of the first aggregated item.
[0282] (22) According to any one of (15)-(21): Aggregate items include, but are not limited to, activities, stores, and aggregate pages.
[0283] It should be noted that the collection, collection, updating, analysis, processing, use, transmission, storage and other aspects of user personal information involved in the technical solution of this disclosure are in compliance with the provisions of relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. Necessary measures are taken for user personal information to prevent illegal access to user personal information data and maintain the security of user personal information, network security and national security.
[0284] Those skilled in the art will appreciate that the embodiments of the present disclosure may be provided as methods, systems, or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present disclosure may take the form of a computer program product implemented on one or more non-transient computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer program code.
[0285] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0286] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0287] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0288] The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present disclosure should be included in the protection scope of the present disclosure.
Claims
1. An information recommendation method, characterized in that: include: Using the target user's item click history behavior sequence, query the first correlation information between different items to obtain an item set related to the target user's historically clicked items; Using the target user's item click history behavior sequence, query the second correlation information between the item and the aggregated item, and obtain the aggregated item set related to the target user's historically clicked items; Determining target items of the polymerized items in the polymerized items set for the target user by matching the polymerized items in the polymerized items set with the items in the item set; The combination of the polymer items in the polymer item set and the target items is recommended to the target user.
2. The method according to claim 1, characterized in that Determining the target items of the polymerized items in the polymerized items set for the target user includes: Matching the polymer items in the polymer item set with items in the item set according to the attribute information satisfied by the target item of the polymer items set in the matching rule; The items in the item set that match the polymer item in the polymer item set are determined as target items of the polymer item for the target user.
3. The method according to claim 1, characterized in that: Recommending a combination of the polymer item in the polymer item set and the target item to the target user includes: Calculating the weight of the combination of the aggregated items in the aggregated items set and the target items thereof, so as to represent the degree of interest of the target user in the combination; Filter combinations based on the weights of each combination; Recommend the selected combinations to target users.
4. The method according to claim 3, characterized in that Calculating the weight of the combination of the polymer item in the polymer item set and the target item includes: For any combination of the polymer product in the polymer product set and its target product: Acquire a weight of the target item in the combination from the first correlation information, which is used to represent the degree of interest of the target user in the target item in the combination; Obtaining the weight of the aggregated items in the combination from the second correlation information, which is used to represent the degree of interest of the target user in the aggregated items in the combination; The weight of the combination is calculated according to the weight of the target item in the combination and the weight of the aggregated items in the combination.
5. The method according to any one of claims 1 to 4, characterized in that: Also includes: According to the item and aggregate item click history behavior sequence of each user, determine the item feature matrix and the aggregate item feature matrix; Determining first correlation information between different items according to the item feature matrix; According to the item feature matrix and the polymer item feature matrix, second correlation information between the items and the polymer items is determined.
6. The method according to claim 5, characterized in that Determine the item feature matrix and the aggregate item feature matrix including: The language representation model is trained using the historical click behavior sequences of each user's items and aggregated items to obtain an item feature matrix and an aggregated item feature matrix, where the item feature matrix includes the feature vector of each item, and the aggregated item feature matrix includes the feature vector of each aggregated item.
7. The method according to claim 6, characterized in that Training a language representation model includes: Divide the item and aggregate item click history behavior sequence of each user into multiple sessions, where the same session includes the object sequence clicked by the same user within a unit time, where the objects include items and aggregate items, and each session is used as a training sample; The language representation model is trained using each training sample, wherein the training parameters include an item feature matrix and a polymer item feature matrix of an encoding layer of the language representation model; Output the item feature matrix and aggregate item feature matrix of the encoding layer of the trained language representation model.
8. The method according to claim 7, characterized in that Using various training samples to train the language representation model includes: Input any first training sample in the same training batch into the language representation model, and the output layer of the language representation model outputs the object probability distribution of the target position in the first training sample based on the first vector, Wherein: the first vector is obtained by concatenating the feature vectors of the items or aggregated items corresponding to the positive labels in each second training sample, and the second training samples are other training samples in the same training batch except the first training samples.
9. The method according to any one of claims 6 to 8, characterized in that: The language representation model includes: a BERT model, an ELMo model, a GTP model, an XLNet model, or a T5 model.
10. The method according to claim 5, characterized in that The first correlation information between different items includes: a weight of the second item relative to the first item, which is used to represent the degree of interest of the user in the second item when the user has clicked on the first item; Alternatively, determining the first correlation information between different items includes: determining a weight of the second item relative to the first item according to an inner product between a feature vector of the first item and a feature vector of the second item in the item feature matrix; or, The weight of the second item relative to the first item is determined according to the inner product between the feature vector of the first item and the feature vector of the second item in the item feature matrix and the posterior click rate of the second item.
11. The method according to claim 5, characterized in that The second correlation information between the item and the aggregated item includes: a weight of the first aggregated item relative to the first item, which is used to represent the degree of interest of the user in the first aggregated item if the user has clicked on the first item; Alternatively, determining the second correlation information between the item and the polymer item includes: determining a weight of the first polymerized item relative to the first item based on an inner product between a feature vector of the first item in the item feature matrix and a feature vector of the first polymerized item in the polymerized item feature matrix; or, The weight of the first aggregated item relative to the first item is determined according to the inner product between the feature vector of the first item in the item feature matrix and the feature vector of the first aggregated item in the aggregated item feature matrix and the posterior click rate of the first aggregated item.
12. The method according to any one of claims 1 to 4, characterized in that: Also includes: Before recommending to the target user, the combination of the polymer items in the polymer item set and the target items is screened or sorted according to one or more of the preset refined sorting rules or rearrangement rules.
13. A method for determining information, characterized in that: include: According to the item and aggregate item click history behavior sequence of each user, determine the item feature matrix and the aggregate item feature matrix; Determining first correlation information between different items according to the item feature matrix; According to the item feature matrix and the polymer item feature matrix, second correlation information between the items and the polymer items is determined.
14. An information recommendation device, comprising: Memory; as well as A processor coupled to the memory, wherein the processor is configured to execute the information recommendation method according to any one of claims 1 to 12 based on instructions stored in the memory.
15. An information determination device, comprising: Memory; as well as A processor coupled to the memory, the processor being configured to execute the information determination method of claim 13 based on instructions stored in the memory.
16. A non-transitory computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the steps of the method according to any one of claims 1 to 13 are implemented.
17. An information recommendation device, characterized in that: include: An item set acquisition unit is configured to query first correlation information between different items by using the item click history behavior sequence of the target user, and obtain an item set related to the items clicked by the target user in history; The aggregated item set acquisition unit is configured to query the second correlation information between the item and the aggregated item by using the item click history behavior sequence of the target user, and obtain the aggregated item set related to the item clicked by the target user in history; a target item determination unit configured to determine a target item for a target user from the aggregated items in the aggregated item set by matching the aggregated items in the aggregated item set with the items in the item set; The information recommendation unit is configured to recommend a combination of the polymer items in the polymer item set and the target items to the target user.
18. An information determination device, characterized in that: include: A feature determination unit is configured to determine an item feature matrix and an aggregate item feature matrix according to the item and aggregate item click history behavior sequence of each user; a first correlation information determining unit, configured to determine first correlation information between different items according to the item feature matrix; The second correlation information determination unit is configured to determine second correlation information between the items and the polymer items based on the item feature matrix and the polymer item feature matrix.
