Data recommendation method, device, equipment and medium

Abstract

The present application provides a data recommendation method, apparatus, device, and medium. The method involves a deep neural network model in artificial intelligence technology and can be applied in item recommendation scenarios. The method includes: obtaining item information fusion features of a historical set of items held by a target object at time t, and obtaining object information fusion features of a historical set of objects holding the target object at time t; determining the social information features of the target object based on the social relationship sequence associated with the target object at time t; combining the item information fusion features, object information fusion features, social information features, target object features, and target item features into a cascade feature to be predicted corresponding to time t; determining a temporal migration feature based on a cascade feature sequence containing the cascade features to be predicted at N times, and determining the recommendation status of the target item based on the temporal migration feature. The present application improves the accuracy of item recommendations.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and in particular to a data recommendation method, apparatus, device, and medium. Background Art

[0002] In a large-scale data environment, most of the data is redundant for objects, and objects may only be interested in certain information. Therefore, providing personalized information for each object has become a daily requirement.

[0003] In existing item recommendation scenarios, items of interest can be recommended to different users based on their interests, preferences, and behavioral data. However, since users' preferences can change over time—for example, a user might have liked type A items a month ago but type B a month later—if recommended items are based on their prior interests and preferences and their own behavioral data, it's difficult to guarantee that the recommended items will be of interest to them, resulting in low accuracy in recommended items. Summary of the Invention

[0004] The embodiments of the present application provide a data recommendation method, apparatus, device, and medium, which can improve the accuracy of item recommendations.

[0005] An embodiment of the present application provides a data recommendation method, including:

[0006] Obtain the item information fusion features corresponding to the historical item set held by the object to be predicted at time t, and obtain the object information fusion features corresponding to the historical object set holding the item to be predicted at time t;

[0007] According to the social associated objects associated with the object to be predicted at time t, a social relationship sequence corresponding to the object to be predicted is constructed, and the social information features corresponding to the object to be predicted are determined according to the social relationship sequence;

[0008] Obtain the object features to be predicted corresponding to the object to be predicted and the item features to be predicted corresponding to the item to be predicted, and combine the item information fusion features, the object information fusion features, the social information features, the object features to be predicted, and the item features to be predicted into the cascade features to be predicted corresponding to time t;

[0009] The temporal migration characteristics between the object to be predicted and the item to be predicted are determined based on the cascade feature sequence, and the recommendation status of the item to be predicted for the object to be predicted is determined based on the temporal migration characteristics; the cascade feature sequence includes the cascade features to be predicted corresponding to N moments, where N moments include moment t, and N is a positive integer greater than 1.

[0010] An embodiment of the present application provides a data recommendation device, including:

[0011] A fusion feature acquisition module is used to obtain the item information fusion features corresponding to the historical item set held by the object to be predicted at time t, and obtain the object information fusion features corresponding to the historical object set holding the item to be predicted at time t;

[0012] A social information acquisition module is used to construct a social relationship sequence corresponding to the object to be predicted based on the social associated objects associated with the object to be predicted at time t, and to determine the social information features corresponding to the object to be predicted based on the social relationship sequence;

[0013] A feature cascade module is used to obtain the object features to be predicted corresponding to the object to be predicted and the item features to be predicted corresponding to the item to be predicted, and combine the item information fusion features, object information fusion features, social information features, object features to be predicted, and item features to be predicted into the cascade features to be predicted corresponding to time t;

[0014] An item recommendation module is configured to determine the temporal migration characteristics between the object to be predicted and the item to be predicted based on a cascade feature sequence, and to determine the recommendation status of the item to be predicted for the object to be predicted based on the temporal migration characteristics; the cascade feature sequence includes cascade features to be predicted corresponding to N moments, where N moments include moment t, and N is a positive integer greater than 1.

[0015] Optionally, the fusion feature acquisition module includes:

[0016] A first acquisition unit is configured to acquire a set of historical items held by the subject to be predicted at time t, and to acquire a first evaluation triplet between the subject to be predicted and the items contained in the set of historical items; the first evaluation triplet includes the subject to be predicted, item k in the set of historical items, and the subject to be predicted's preference value for item k, where k is a positive integer;

[0017] a second acquisition unit, configured to acquire historical item features corresponding to item k, and convert the preference value contained in the first evaluation triplet into a first preference vector associated with item k;

[0018] A first output unit is configured to input a first cascade feature between the historical item feature and the first preference vector into a first neural network layer in the recommendation model, and output an item output feature corresponding to the first cascade feature through the first neural network layer;

[0019] The item information fusion unit is used to obtain the item weight corresponding to each item in the historical item set, and determine the item information fusion feature corresponding to the historical item set based on the item weight and the item output feature associated with each item in the historical item set.

[0020] Optionally, the first acquiring unit includes:

[0021] The identification acquisition subunit is used to obtain the object identification corresponding to the object to be predicted and the item identification corresponding to item k in the historical item set;

[0022] The indicator acquisition subunit is used to obtain the usage time interval and usage frequency of the object to be predicted for the item k, and multiply the usage time interval and the usage frequency to determine the item usage attenuation factor corresponding to the item k;

[0023] The preference value determination subunit is used to determine the balance factor corresponding to item k based on the unit usage count of item k by the subject to be predicted, and to determine the preference value of the subject to be predicted for item k based on the usage time interval, usage frequency, item usage attenuation factor, and balance factor;

[0024] The triplet determination subunit is configured to determine the object identifier, the item identifier, and the preference value of the object to be predicted for the item k as a first evaluation triplet between the object to be predicted and the item k.

[0025] Optionally, the second acquiring unit includes:

[0026] A feature conversion subunit is used to obtain attribute description information associated with item k and convert the attribute description information into an initial attribute feature vector corresponding to item k;

[0027] a feature preprocessing subunit, configured to obtain a minimum eigenvalue and a maximum eigenvalue in an initial attribute feature vector, obtain a first difference between the maximum eigenvalue and the minimum eigenvalue, and obtain a second difference between a target eigenvalue and the minimum eigenvalue; the target eigenvalue being any eigenvalue in the initial attribute feature vector;

[0028] The first item feature determination subunit is configured to determine the ratio between the second difference and the first difference as a normalized target feature value, and determine the historical item feature corresponding to the item k according to the normalized target feature value.

[0029] Optionally, the item information fusion unit includes:

[0030] The first object feature determination subunit is configured to obtain an item operation feature corresponding to the object to be predicted from the object log, and determine the object feature corresponding to the object to be predicted based on the item operation feature and the basic object profile of the object to be predicted;

[0031] An item similarity acquisition subunit is used to acquire the item similarity between the item to be predicted and item k based on the item features corresponding to the item to be predicted and the historical item features corresponding to item k;

[0032] A first prior information input subunit is used to input the characteristics of the object to be predicted, the characteristics of the item to be predicted, and the item similarity into the first attention network layer in the recommendation model;

[0033] The item weight acquisition subunit is used to perform a product operation on the feature matrix determined by the features of the object to be predicted, the features of the item to be predicted, and the item similarity according to the network parameter matrix of the first attention network layer, to obtain the item weight corresponding to each item in the historical item set.

[0034] Optionally, the fusion feature acquisition module includes:

[0035] a third acquisition unit, configured to acquire a set of historical objects holding the item to be predicted at time t, and acquire a second evaluation triplet between each object in the set of historical objects and the item to be predicted; the second evaluation triplet includes object i in the set of historical objects, the item to be predicted, and a preference value of object i for the item to be predicted, where i is a positive integer;

[0036] a fourth acquiring unit, configured to acquire historical object features corresponding to the object i, and convert the preference value included in the second evaluation triplet into a second preference vector corresponding to the object i;

[0037] A second output unit is configured to input a second cascade feature between the historical object feature and the second preference vector into a second neural network layer in the recommendation model, and output an object output feature corresponding to the second cascade feature through the second neural network layer;

[0038] The object information fusion unit is used to obtain the object weight corresponding to each object in the historical object set, and determine the object information fusion feature corresponding to the historical object set according to the object weight and the object output feature associated with each object in the historical object set.

[0039] Optionally, the object information fusion unit includes:

[0040] The object similarity acquisition subunit is used to obtain the object similarity between the object to be predicted and object i based on the object features corresponding to the object to be predicted and the historical object features corresponding to object i;

[0041] A second prior information input subunit is used to input the features of the object to be predicted, the features of the item to be predicted corresponding to the item to be predicted, and the object similarity into the second attention network layer in the recommendation model;

[0042] The object weight acquisition subunit is used to perform a product operation on the feature matrix determined by the features of the object to be predicted, the features of the item to be predicted, and the item similarity according to the network parameter matrix of the second attention network layer, so as to obtain the object weight corresponding to each object in the historical object set.

[0043] Optionally, the social information acquisition module includes:

[0044] A random walk unit is used to obtain the socially associated objects associated with the object to be predicted at time t, generate an object relationship chain including the object to be predicted and the socially associated objects, and perform a random walk in the object relationship chain starting from the object to be predicted to obtain a sequence of associated objects;

[0045] Similar item determination unit, used to obtain the associated object u in the associated object sequence [m] , in the associated object u [m] Among the items held, the item with the greatest item similarity to the item to be predicted is determined as the similar item corresponding to the item to be predicted;

[0046] The social relationship sequence determination unit is used to associate the object u [m] , similar items and related objects u [m] The preference values ​​for similar items are determined as social sequence elements, and the social sequence elements associated with each associated object in the associated object sequence are determined as the social relationship sequence corresponding to the object to be predicted.

[0047] Optionally, the social information acquisition module includes:

[0048] Similar item feature acquisition unit, used to obtain the associated object u [m] The corresponding associated object features, and the similar item features corresponding to the similar items, convert the preference values ​​in the social relationship sequence into a third preference vector;

[0049] a third output unit, configured to input a third cascade feature among the associated object feature, the similar item feature, and the third preference vector into a third neural network layer in the recommendation model, and output a sequence output feature corresponding to the third cascade feature through the third neural network layer;

[0050] A bidirectional cyclic encoding unit is used to input the sequence output features into the first recurrent neural network in the recommendation model, and perform bidirectional cyclic encoding on the sequence output features according to the first recurrent neural network to obtain forward encoding features and reverse encoding features;

[0051] The social information fusion unit is used to splice the forward coding features and the reverse coding features into latent state features, obtain the social weight corresponding to each associated object in the social relationship sequence, and determine the social information features corresponding to the object to be predicted based on the social weight and the latent state features.

[0052] Optionally, the social information fusion unit includes:

[0053] A social similarity acquisition subunit, configured to acquire the social similarity between the object to be predicted and each associated object according to the associated object features corresponding to each associated object in the social relationship sequence and the predicted object features corresponding to the object to be predicted;

[0054] An intimacy determination subunit is used to determine the social intimacy between the object to be predicted and each associated object based on the social parameters between the object to be predicted and each associated object; the social parameters include social duration, number of common social objects, interaction duration, and interaction frequency;

[0055] A third prior information input subunit is used to input the characteristics of the object to be predicted, the social similarity, the social intimacy, and the characteristics of the item to be predicted corresponding to the item to be predicted into the third attention network layer in the recommendation model;

[0056] The social weight acquisition subunit is used to perform a product operation on the feature matrix determined by the features of the object to be predicted, social similarity, social intimacy and the features of the item to be predicted according to the network parameter matrix of the third attention network layer, so as to obtain the social weight corresponding to each associated object.

[0057] Optionally, the number of social relationship sequences is P, and the P social relationship sequences include social relationship sequence S ρ , P is a positive integer, ρ is a positive integer less than P;

[0058] The social information fusion unit includes:

[0059] The social output feature determination subunit is used to determine the social relationship sequence S ρ The social weight corresponding to the associated object in the social relationship sequence S ρ The implicit state features of the association determine the social relationship sequence S ρ Corresponding social output features;

[0060] The social output feature fusion subunit is used to fuse the social output features corresponding to P social relationship sequences to obtain the social information features corresponding to the object to be predicted.

[0061] Optionally, the item recommendation module includes:

[0062] a fourth output unit, configured to combine the cascade features to be predicted corresponding to the N moments into a cascade feature sequence, sequentially input the cascade features to be predicted contained in the cascade feature sequence into a fourth neural network layer in the recommendation model, and output network output features corresponding to the N cascade features to be predicted through the fourth neural network layer;

[0063] A unidirectional cyclic encoding unit is used to sequentially input the N network output features into the second cyclic neural network in the recommendation model, and perform forward cyclic encoding on the network output features according to the second cyclic neural network to obtain network implicit features;

[0064] A time series migration feature acquisition unit is used to acquire the time series migration feature corresponding to the cascade feature sequence according to the network implicit feature and the hidden layer weight corresponding to the second recurrent neural network;

[0065] The recommendation status determination unit is used to obtain the predicted preference value of the object to be predicted for the item to be predicted based on the time series migration feature, and determine the recommendation status of the item to be predicted for the object to be predicted based on the predicted preference value.

[0066] Optionally, the number of items to be predicted is L, where L is a positive integer;

[0067] The recommended state determination unit includes:

[0068] A sorting subunit is used to obtain the predicted preference values ​​of the subject to be predicted for L items to be predicted, and sort the predicted preference values ​​associated with the L items to be predicted in descending order to obtain the sorted L items to be predicted;

[0069] The recommended item selection subunit is used to determine the recommendation status corresponding to the first y items to be predicted among the sorted L items to be predicted as the recommendation success status; y is a positive integer less than or equal to L;

[0070] The item push subunit is used to push y recommended items that are in the recommendation success state to the object to be predicted.

[0071] In one aspect, an embodiment of the present application provides a computer device, including a memory and a processor, wherein the memory is connected to the processor, the memory is used to store a computer program, and the processor is used to call the computer program so that the computer device executes the method provided in the above aspect of the embodiment of the present application.

[0072] On one hand, an embodiment of the present application provides a computer-readable storage medium, in which a computer program is stored. The computer program is suitable for being loaded and executed by a processor, so that a computer device with a processor executes the method provided in the above aspect of the embodiment of the present application.

[0073] According to one aspect of the present application, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method provided in the above aspect.

[0074] In the embodiment of the present application, the items held by the object to be predicted at time t can be fused with item information, that is, analyzed at the item level to obtain item information fusion features; the objects holding the items to be predicted at time t can be fused with object information, that is, analyzed at the object level to obtain object information fusion features; the social information of the socially associated objects associated with the object to be predicted at time t can be obtained, that is, analyzed at the social relationship chain level to obtain social information features. Then, the item information fusion features, object information fusion features, social information features, object to be predicted features, and item to be predicted features obtained at N moments can be cascaded to obtain cascade features to be predicted. The temporal migration features between the cascade features to be predicted corresponding to N moments (including moment t) can be obtained. According to the temporal migration features, the recommendation status of the item to be predicted for the object to be predicted can be determined. In other words, in the item recommendation process, the present application can introduce the social relationships of the object to be predicted and the migration of the object's interest preferences / social relationships over time, which can enhance the matching degree between the object to be predicted and the recommended items, thereby improving the recommendation accuracy of the items. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0076] Figure 1 This is a schematic diagram of a network architecture provided by an embodiment of the present application;

[0077] Figure 2 This is a schematic diagram of an item recommendation scenario provided by an embodiment of the present application;

[0078] Figure 3 This is a flow chart of a data recommendation method provided in an embodiment of the present application;

[0079] Figure 4 This is a schematic diagram of a process of converting classification features into one-hot encoding provided by an embodiment of the present application;

[0080] Figure 5 This is a flow chart of a data recommendation method provided in an embodiment of the present application;

[0081] Figure 6 This is a flowchart of an item recommendation method provided in an embodiment of the present application;

[0082] Figures 7a to 7dThis is a schematic diagram of a network structure of a recommendation model provided in an embodiment of the present application;

[0083] Figure 8 This is a schematic diagram of a game prop recommendation scenario provided by an embodiment of the present application;

[0084] Figure 9 This is a structural diagram of a data recommendation device provided in an embodiment of the present application;

[0085] Figure 10 It is a structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0086] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0087] This application relates to artificial intelligence (AI), blockchain, and cloud technologies. AI is the theory, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive field of computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to perceive, reason, and make decisions. AI is an interdisciplinary discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Basic AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning. This application specifically relates to machine learning / deep learning within the AI ​​field.

[0088] Machine learning is a multidisciplinary field, encompassing probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to imbue computers with intelligence. Its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.

[0089] The machine learning / deep learning method specifically adopted in this application can be expressed as follows: according to the usage characteristics of the items, the preference value of the object for the items held by the object is calculated; according to the image characteristics of the object and the operation characteristics of the object on the item, the similarity between different objects is calculated; the similarity between different items can also be calculated according to the attribute characteristics of the item; and then the similarity of the object / item can be used as prior information to construct a recommendation model, and the social relationship sequence of the object and the preference value of the object for the items held by the object are used as input to predict the preference value of the object for the predicted item, and then the item is recommended to the object based on the predicted preference value.

[0090] The present application relates to the field of big data under cloud technology. Big data refers to a collection of data that cannot be captured, managed and processed by conventional software tools within a certain time frame. It is a massive, high-growth and diversified information asset that requires new processing models to have stronger decision-making power, insight discovery and process optimization capabilities. With the advent of the cloud era, big data has also attracted more and more attention. Big data requires special technologies to effectively process large amounts of data within a tolerable time. Technologies suitable for big data include large-scale parallel processing databases, data mining, distributed file systems, distributed databases, cloud computing platforms, the Internet and scalable storage systems. In the item recommendation scenario involved in this application, since the number of objects and items is very large, it is necessary to involve the calculation of information such as object portraits, object operation characteristics on items, and item usage characteristics, in order to achieve accurate recommendations for different objects.

[0091] Blockchain is a new application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, and encryption algorithm. Blockchain is essentially a decentralized database, a string of data blocks (also called blocks) generated using cryptography. Each data block is linked through random hashing (also called hash algorithm). The subsequent block contains the encrypted hash of the previous block, the corresponding timestamp, and transaction data (usually represented by the hash value calculated by the Merkle tree algorithm). This design makes the block content difficult to tamper with. The distributed ledger connected by blockchain technology allows both parties to effectively record transactions and permanently verify these transactions. The data involved in this application (for example, object data, object operation data on items, object held item data, etc.) can be stored in the blockchain.

[0092] See Figure 1 , Figure 1 This is a schematic diagram of a network architecture provided by an embodiment of the present application. Figure 1 As shown, the network architecture may include a server 10d and a user terminal cluster, and the user terminal cluster may include one or more user terminals, and the number of user terminals is not limited here. Figure 1 As shown, the user terminal cluster may specifically include user terminal 10a, user terminal 10b and user terminal 10c, etc. Among them, server 10d may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. User terminals 10a, user terminals 10b and user terminals 10c, etc. may include: smart phones, tablet computers, laptop computers, PDAs, mobile internet devices (MID), wearable devices (such as smart watches, smart bracelets, etc.) and smart TVs, etc., which have video / image playback functions. As Figure 1 As shown, the user terminal 10a, the user terminal 10b and the user terminal 10c can respectively establish a network connection with the server 10d, so that each user terminal can exchange data with the server 10d through the network connection.

[0093] Taking user terminal 10a as an example, when subject 1 triggers an application client installed on user terminal 10a to launch the application client on user terminal 10a, the application client on user terminal 10a can recommend items of interest to subject 1 based on the socially associated objects (which can be social friends) associated with subject 1 and subject 1's preferences for items they hold. For example, the application client can be an e-commerce client, a game client, an information client, etc., and the items in the application client can be merchandise in the e-commerce client, game props in the game client, or information content in the information client.

[0094] Taking the application client as a game client as an example, the items in this case can refer to game props in the game client. After object 1 enters the game client in user terminal 10a, user terminal 10a can obtain the historical item set held by object 1 at different time points (such as time 1, time 2, ..., time t). Based on object 1's preference value for each item in the historical item set and each item in the historical item set, the item information fusion feature corresponding to the historical item set can be obtained. User terminal 10a can also obtain the historical object set holding the item to be predicted. Based on each object in the historical object set's preference value for the item to be predicted and each object in the object item set, the object information fusion feature corresponding to the object item set can be obtained. The item to be predicted can refer to the remaining items in the game client excluding the historical item set. User terminal 10a can also construct a social relationship sequence corresponding to object 1 based on object 1's social association objects at different time points, and determine the social information feature corresponding to object 1 based on the social relationship sequence.

[0095] Furthermore, user terminal 10a can combine the item information fusion features, object information fusion features, social information features, object features corresponding to subject 1, and the item features corresponding to the item to be predicted at different times into a cascade feature to be predicted. The cascade features to be predicted at different times can be combined into a cascade feature sequence. Based on the cascade feature sequence, the temporal transition features between subject 1 and the item to be predicted can be determined. Based on the temporal transition features, subject 1's predicted preference value for the item to be predicted can be determined. The items to be predicted can be sorted in order of predicted preference values, and the first y (y is a positive integer, such as y can be 5) items to be predicted in the sorting can be recommended to subject 1. Because the item recommendation process incorporates the items held by subject 1, the objects holding the items to be predicted, subject 1's social relationship sequence, and subject 1's preference shifts and social relationship shifts at different times, the recommended items are those that subject 1 is interested in, which can improve the accuracy of item recommendations.

[0096] See Figure 2 , Figure 2This is a schematic diagram of an item recommendation scenario provided by an embodiment of the present application. Taking the game prop recommendation scenario as an example, the item recommendation solution proposed in the embodiment of the present application is described as follows: Figure 2 As shown, when object 1 triggers the game client in user terminal 20a, user terminal 20a can respond to the triggering operation on the game client, launch the game client in user terminal 20a, and enter the homepage of the game client, where a "Start Game" control and an item viewing control 20b can be displayed. Before starting the game, object 1 can trigger the item viewing control 20b to view its game items. At this time, user terminal 20a can respond to the triggering operation on the item viewing control 20b, generate an item viewing instruction, and send the item viewing instruction to the backend server corresponding to the game client. Optionally, object 1 can trigger the item viewing control 20b before starting the game, or trigger the item viewing control when it needs to use game items during the game.

[0097] After receiving the props viewing instruction sent by the user terminal 20a, the backend server corresponding to the game client can obtain the game props held by the object 1, such as Figure 2 As shown, the game props held by subject 1 are: prop 1, prop 2, and prop 3. Item usage information such as the first use time, the last use time, the total number of times the item is used, and the total number of times the same item type is used for the item held by subject 1 can be obtained from the object log of the game client. Then, object-item feedback data can be constructed based on the item usage information. Based on the object-item feedback data, the preference value of subject 1 for the item held can be calculated.

[0098] The backend server can combine the object, item and preference value into an evaluation triple, such as Figure 2 As shown, the evaluation triplet corresponding to prop 1 can be expressed as <object 1, prop 1, 0.6>. At this time, the evaluation triplet is used to indicate that the preference value of object 1 for prop 1 is 0.6; the evaluation triplet corresponding to prop 2 can be expressed as <object 1, prop 2, 0.8>. At this time, the evaluation triplet is used to indicate that the preference value of object 1 for prop 2 is 0.8; the evaluation triplet corresponding to prop 3 can be expressed as <object 1, prop 3, 0.3>. At this time, the evaluation triplet is used to indicate that the preference value of object 1 for prop 3 is 0.3.

[0099] The backend server can convert the props held by subject 1 and subject 1's preference value for each held item into vectors. In the evaluation triple <object 1, prop 1, 0.6>, prop 1 can be converted into prop vector 20c, and preference value 0.6 can be converted into preference vector 20d. Prop vector 20c and preference vector 20d are input into the multi-layer fully connected neural network 1 in the recommendation model. The multi-layer fully connected neural network 1 can then output a feature vector 20e corresponding to the evaluation triple <object 1, prop 1, 0.6>. Similarly, the item vector corresponding to item 2 in the evaluation triple <object 1, item 2, 0.8> and the preference vector corresponding to the preference value 0.8 are input into the multi-layer fully connected neural network 1 in the recommendation model. The multi-layer fully connected neural network 1 can output the feature vector 20f corresponding to the evaluation triple <object 1, item 2, 0.8>. The item vector corresponding to item 3 in the evaluation triple <object 1, item 3, 0.3> and the preference vector corresponding to the preference value 0.3 are input into the multi-layer fully connected neural network 1 in the recommendation model. The multi-layer fully connected neural network 1 can output the feature vector 20g corresponding to the evaluation triple <object 1, item 3, 0.3>. The backend server can obtain the item weight (also called item weight) corresponding to each item held by object 1. Based on feature vector 20e, feature vector 20f, feature vector 20g and the item weight, the item information fusion feature corresponding to the item held by object 1 is obtained. Among them, the prop vectors corresponding to the above-mentioned props 1, props 2 and props 3 can all be called historical item features, the preference vectors corresponding to the above-mentioned preference values ​​0.6, 0.8 and 0.3 can all be called first preference vectors, and the above-mentioned feature vectors 20e, 20f and 20g can all be called item output features.

[0100] Optionally, the background server can also obtain the game props to be predicted (also referred to as items to be predicted) corresponding to object 1. The game props to be predicted can be the remaining game props in the game client except the items held by object 1 (that is, the remaining game props except the above-mentioned props 1, props 2 and props 3). For example, the game props to be predicted of object 1 can be represented as: props 4, props 5, props 6, props 7,...; in the embodiment of the present application, the processing method for each game prop to be predicted is the same, so the following takes prop 4 as an example to describe the processing process of the game props to be predicted.

[0101] The backend server can get all objects holding props 4, such as Figure 2As shown, the objects holding item 4 may include: object 2, object 3, object 4, ..., object i (i is a positive integer); based on the objects holding item 4, we can obtain the evaluation triples <object 2, item 4, 0.7>, <object 3, item 4, 0.5>, ..., and <object i, item 4, 0.9>. Among them, the evaluation triple <object 2, item 4, 0.7> is used to indicate that the preference value of object 2 holding item 4 for item 4 is 0.7, the evaluation triple <object 3, item 4, 0.5> is used to indicate that the preference value of object 3 holding item 4 for item 4 is 0.5, and the evaluation triple <object i, item 4, 0.9> is used to indicate that the preference value of object i holding item 4 for item 4 is 0.9.

[0102] The backend server can convert the object holding prop 4 and the preference value of the object holding prop 4 to prop 4 into vectors, and based on the converted vectors, obtain the feature vectors associated with all objects holding prop 4. Among them, object 2 in the evaluation triple <object 2, prop 4, 0.7> can be converted into object vector 20h, and the preference value 0.7 can be converted into preference vector 20i. The object vector 20h and the preference vector 20i are input into the multi-layer fully connected neural network 2 in the recommendation model, and the multi-layer fully connected neural network 2 can output the feature vector 20j corresponding to the evaluation triple <object 2, prop 4, 0.7>; similarly, the object vector corresponding to object 3 in the evaluation triple <object 3, prop 4, 0.5> and the preference value 0.5 can be mapped to the feature vector 20j. The corresponding preference vector is input into the multi-layer fully connected neural network 2 in the recommendation model, which can output the feature vector 20m corresponding to the evaluation triple <object 3, item 4, 0.5>. The object vector corresponding to object i in the evaluation triple <object i, item 4, 0.9> and the preference vector corresponding to the preference value 0.9 can be input into the multi-layer fully connected neural network 2 in the recommendation model, which can output the feature vector 20n corresponding to the evaluation triple <object i, item 4, 0.9>. The backend server can obtain the object weights corresponding to all objects holding item 4 and obtain the object information fusion features corresponding to all objects holding item 4 based on feature vector 20j, feature vector 20m, feature vector 20n, and object weights. Among them, the object vectors corresponding to the above-mentioned objects 2, object 3,..., object i can also be called historical object features, the preference vectors corresponding to the above-mentioned preference values ​​0.7, 0.5 and 0.9 can also be called second preference vectors, and the above-mentioned feature vectors 20j, 20m and 20n can also be called object output features.

[0103] Optionally, the backend server can also obtain the game friends (also called social related objects) of object 1 in the game client. Figure 2As shown, the game friends of object 1 may include: object 2, object 7, object 8, ..., object i. By using the random walk method, starting from object 1, a friend sequence associated with object 1 (also called an associated object sequence) is constructed; wherein the associated objects included in the friend sequence may be the game friends and indirect game friends of object 1, such as Figure 2 As shown, the friend sequence can be represented as {object 2, object 12, object 18, ..., object M}, where M is a positive integer; indirect game friends in the embodiment of the present application can be represented as game friends of the game friends of object 1. For example, object 2 is a game friend of object 1, object 3 is a game friend of object 2, but object 3 is not a game friend of object 1. In this case, object 3 can be called an indirect game friend of object 1. Of course, in the embodiment of the present application, both game friends and indirect game friends can be referred to as socially associated objects associated with object 1, and indirect game friends and object 1 can have common friends.

[0104] For each associated object in the friend sequence, the game prop that is most similar to the game prop to be predicted (such as the above-mentioned prop 4) can be selected from the items held by the associated object, as well as the associated object's preference value for the most similar game prop, to construct a friend-prop sequence (also known as a social relationship sequence). Each element in the friend-prop sequence can be represented as an evaluation triple, such as Figure 2As shown, the friend-prop sequence may include: evaluation triple <object 2, prop 7, 0.5>, evaluation triple <object 12, prop 9, 0.1>, evaluation triple <object 18, prop 20, 0.8>, ..., evaluation triple <object M, prop 16, 0.4>. Among them, prop 7 in the above evaluation triplet <object 2, prop 7, 0.5> is the item most similar to prop 4 among the items held by object 2, and 0.5 is the preference value of object 2 for prop 7; prop 9 in the evaluation triplet <object 12, prop 9, 0.1> is the item most similar to prop 4 among the items held by object 12, and 0.1 is the preference value of object 12 for prop 9; prop 20 in the evaluation triplet <object 18, prop 20, 0.8> is the item most similar to prop 4 among the items held by object 18, and 0.8 is the preference value of object 18 for prop 20; prop 16 in the evaluation triplet <object M, prop 16, 0.4> is the item most similar to prop 4 among the items held by object M, and 0.4 is the preference value of object M for prop 16. The objects, props and preference values ​​in each element (i.e., evaluation triplet) contained in the friend-prop sequence are converted into vectors, and then the vectors corresponding to the objects, props and preference values ​​in each element can be input into the multi-layer fully connected neural network 3 in the recommendation model, and the sequence output features corresponding to each element can be output through the multi-layer fully connected neural network 3; the sequence output features corresponding to each element can be input into the bidirectional short-term memory network (bidirectional LSTM, also called the first recurrent neural network) in sequence according to the order of the elements in the friend-prop sequence, and the implicit state features corresponding to each element can be output through the bidirectional LSTM; the social weight corresponding to each associated object in the friend sequence is obtained, and the social information features corresponding to object 1 are determined based on the social weight and the implicit state features.

[0105] Furthermore, the background server can cascade the predicted object features corresponding to object 1, the predicted item features corresponding to props 4, the item information fusion features, the object information fusion features and the social information features to obtain cascade features. It should be noted that in order to learn how the friend relationships and interest preferences of the objects migrate over time, for object 1 and each game prop to be predicted, the item information fusion features, object information fusion features and social information features corresponding to N different moments can be obtained, and the predicted item features corresponding to the game props to be predicted and the predicted object features corresponding to object 1 are introduced to obtain the cascade features corresponding to N different moments; and then the cascade features at N different moments can be sequentially input into the one-way LSTM (also called the second recurrent neural network), and the predicted preference value of object 1 for the game props to be predicted is output through the one-way LSTM. As shown in FIG. Figure 2As shown, for object 1 and prop 4, N cascade features between object 1 and prop 4 can be obtained, and the N cascade features are input into the unidirectional LSTM. The unidirectional LSTM can output the predicted preference value of object 1 for prop 4 as: a1; similarly, the unidirectional LSTM can output the predicted preference value of object 1 for prop 5 as: a2, the unidirectional LSTM can output the predicted preference value of object 1 for prop 6 as: a3, the unidirectional LSTM can output the predicted preference value of object 1 for prop 7 as: a4, ...

[0106] The backend server can sort the predicted preference values ​​of subject 1 for different game props to be predicted, and recommend the game props corresponding to the top two largest predicted preference values ​​to subject 1 (here, the default number of game props recommended for each subject is 2). For example, the game props corresponding to the top two largest predicted preference values ​​are: props 7 and props 8. The backend server can push props 7 and props 18 to user terminal 20a. User terminal 20a can receive props 7 and props 18 pushed by the backend server and display the props already owned by subject 1 (including props 1, props 2, and props 3) and the game props recommended for subject 1 (including props 7 and props 18) on the prop display page of the game client. It should be noted that, for user terminal 20a, after subject 1 triggers the item view control 20b, user terminal 20a can respond to the triggering operation on the item view control 20b by directly switching from the game client's homepage to the item display page. The item display page then displays the game items already owned by subject 1 and the game items recommended for subject 1. If subject 1 is interested in the recommended game items, subject 1 can activate the recommended game items (e.g., purchase recommended items 7 and 18). After items 7 and 18 are successfully activated, subject 1 can directly use the activated items 7 and 18 during subsequent game play. Of course, if subject 1 is not interested in the recommended game items, subject 1 can disregard the recommended items on the item display page.

[0107] In an embodiment of the present application, the friend relationship of object 1 and the migration of the object's interest preferences over time can be introduced into the game prop recommendation scenario, which can enhance the matching degree between object 1 and the recommended game props, thereby improving the recommendation accuracy of the game props.

[0108] See Figure 3 , Figure 3 This is a flow chart of a data recommendation method provided by an embodiment of the present application. It is understandable that the data recommendation method can be executed by a computer device, which can be an independent server, or a user terminal, or a system consisting of a server and a user terminal, or a server cluster consisting of multiple servers, or a computer program application (including program code), which is not specifically limited here. Figure 3 As shown, the data recommendation method may include the following steps:

[0109] Step S101 : obtaining item information fusion features corresponding to a set of historical items held by an object to be predicted at time t, and obtaining object information fusion features corresponding to a set of historical objects holding an object to be predicted at time t.

[0110] Specifically, in the item recommendation scenario, the computer device can make personalized item recommendations for the object based on the object's interest preferences, social relationships, and usage characteristics of the items held. The computer device can obtain the object to be predicted (such as the above Figure 2 In the corresponding embodiment, object 1) has a set of historical items held at time t, which may include all items held by the object to be predicted in the application client; based on each item included in the historical item set and the preference value of the object to be predicted for each item in the historical item set, the item information fusion feature corresponding to the historical item set is obtained. The computer device may determine the remaining items in the application client other than the items held by the object to be predicted as the items to be predicted corresponding to the object to be predicted; for each item to be predicted corresponding to the object to be predicted, the computer device may obtain a set of historical objects that held the item to be predicted at time t, which may include all objects that held the item to be predicted; based on each object included in the historical object set and the preference value of each object in the historical object set for the item to be predicted, the object information fusion feature corresponding to the historical object set is obtained.

[0111] Here, time t can be time 1, time 2, time 3, ..., N (N is a positive integer greater than 1). Since the items held by the subject to be predicted may change over time, for example, the number of items held by the subject to be predicted was 3 a week ago, 5 three days ago, and 6 at the current moment, in order to more accurately capture the changes in the items held by the subject to be predicted, the computer device can obtain a historical set of items held by the subject to be predicted at different moments. For example, when N is 5, the N moments can include time 1, time 2, time 3, time 4, and time 5. Time 5 can represent the current moment when an item is recommended for the subject to be predicted, time 4 can represent 1 day ago, time 3 can represent 3 days ago, time 2 can represent 1 week ago, and time 1 can represent 1 month ago. Of course, for each item to be predicted corresponding to the subject to be predicted, the number of subjects holding the item to be predicted may also change over time. In order to more accurately capture the changes in the subjects holding the item to be predicted, the computer device can retain a historical set of objects for the item to be predicted at different moments. It should be noted that each moment may correspond to a historical item set and a historical object set, a historical item set corresponds to an item information fusion feature, and a historical object set corresponds to an object information fusion feature.

[0112] It should be noted that in the item recommendation scenario, due to the lack of explicit feedback data from the subject on the item, that is, the subject's specific rating of the item, it is necessary to construct subject-item feedback data. This object-item feedback data can be used to reflect the subject's preference for the item. The computer device can calculate the subject's preference value for the item it already holds based on information such as the item's usage characteristics, a timing factor (used to predict the probability of the subject using the item next time), a frequency factor (used to evaluate the frequency of the subject's use of the item per unit time), a weighting factor, a balance factor (used to represent the weighted value of N days within 30 days), and a decay factor (used to represent the factor that decays the subject's interest and activity based on the sliding time window). This preference value can be calculated using the object-item preference calculation model.

[0113] Among them, the specific calculation process of the preference value may include: the computer device can obtain the object's FIRST_DATE (used to indicate the time point of the first use of an item), LAST_DATE (used to indicate the time point of the most recent use of an item), TOTAL_CNT (used to indicate the total number of times an item is used), and PV_X (used to indicate the number of times an item is used per unit time, where X = 1, 3, 5, 7, 10, 15, 20, 30) from the object log corresponding to the application client, and abstract the object-item preference calculation model into four indicators: TI (time interval of use), FoU (frequency of use), TIFoU (item usage decay factor), and BF (balance factor).

[0114] Among them, TI = TODAY - LAST_DATE, which is used to represent the time interval since the subject last used an item. The closer the time of last use, the more preference the subject has for the item. The TI indicator can be used to characterize the degree of attention the subject pays to the item.

[0115] FoU = (TOTAL_CNT) / (TODAY - FIRST_DATE), which is used to indicate the frequency with which an object uses an item within a limited time. The more frequently an object uses an item, the higher its satisfaction with the item. This FoU indicator can be used to characterize the object's satisfaction with the item.

[0116] TIFoU = [(TODAY-LAST_DATE)*TOTAL_CNT] / (TODAY-FIRST_DATE). The TIFoU indicator is the product of the above-mentioned TI indicator and the FoU indicator. If x1 = (TODAY-LAST_DATE) and x2 = (LAST_DATE-FIRST_DATE), then (TODAY-FIRST_DATE) = (x1+x2), that is, the size of the TIF indicator depends on the size of x2 / x1; the larger the value of x2 / x1, the smaller the value of the TIFoU indicator, indicating that the longer the subject uses an item and the more recently it is used, the stronger the preference for the item.

[0117] BF=a1*PV_1+b1*(PV_3) / 3+c1*(PV_5) / 5+d1*(PV_7) / 7+e1*(PV_10) / 10+f1*(PV_15) / 15+g1*(PV_20) / 20+h1*(PV_30) / 30, where a1+b1+c1+d1+e1+f1+g1+h1=1.0; the BF indicator is used to represent the balance factor, which can be used to increase the weight value of recently used items. The larger the value of the BF indicator, the higher the object's recent preference for a certain item. The BF indicator can be used to measure the object's recent preference for an item; the above-mentioned weight values ​​a1, b1, c1, d1, e1, f1, g1 and h1 can be determined using the hierarchical analysis method. The Analytic Hierarchy Process (AHP) decomposes a decision problem into a hierarchical structure, sequentially following the overall objective, sub-objectives, evaluation criteria, and specific alternatives. The eigenvectors of the judgment matrix are then used to determine the priority of each element at each level relative to an element at the previous level. Finally, a weighted summation method is used to recursively combine the final weights of each alternative relative to the overall objective. The one with the highest final weight is considered the optimal solution. The "priority" mentioned above can be a relative measure, indicating the relative superiority of each alternative relative to a specific evaluation criterion or sub-objective, as well as the relative importance of each sub-objective relative to the objective at the previous level. The AHP is particularly suitable for decision-making problems involving systems with hierarchical and interleaved evaluation indicators, where the objective values ​​are difficult to quantitatively describe. Its application involves constructing a judgment matrix and finding its largest eigenvalue.

[0118] Furthermore, the computer device can also use the hierarchical analysis method to determine the weight values ​​of the above-mentioned TI indicator, FoU indicator, TIFoU indicator and BF indicator. For example, the weight value of the TI indicator is: a2, the weight value of the FoU indicator is: b2, the weight value of the TIFoU indicator is: c2, and the weight value of the BF indicator is: d2, that is,<TI,FoU,TIFoU,BF> The weight value can be expressed as<a2,b2,c2,d2> , at this time, the object's behavior data on the item can be abstracted as:<USER,ITEM,TOTAL_CNT,TI,FoU,TIFoU,BF> A seven-tuple, where USER refers to the object ID (identification number, also called object identification) of the object in the application client, and ITEM refers to the item ID (identification number, also called item identification) of the item in the application client. Based on the above seven-tuple, the computer device can determine the object's preference value ALL_SCORE for a certain item based on the weight values ​​corresponding to the TI indicator, FoU indicator, TIFoU indicator, and BF indicator, the weight of TOTAL_CNT, and TOTAL_CNT, TI, FoU, TIFoU, and BF in the seven-tuple, and convert the above seven-tuple into<USER,ITEM,ALL_SCORE> triple, that is, the triple<USER,ITEM,ALL_SCORE> ALL_SCORE is used to represent a user's preference for an item (the triplet here can also be called an evaluation triplet between the user and the item). The larger the ALL_SCORE value, the higher the user's preference for the item. Optionally, the computer device can normalize the preference value ALL_SCORE to obtain a normalized preference value. In this case, the normalized preference value can still be represented by ALL_SCORE. The normalization process can include, but is not limited to, normalization and maximum and minimum value normalization.

[0119] Using the preference value calculation process described above, we can obtain the preference value of each object for the items it holds. Let the object set be U = {object 1, object 2, ..., object A}, and the item set be V = {item 1, item 2, ..., item B}, where A in the object set U can be represented as the total number of objects contained in the application client, and B in the item set V can be represented as the total number of items contained in the application client. For A objects and B items in the application client, we can construct the object's preference matrix R for the items. The size of the preference matrix R can be A×B, and the elements R in the preference matrix R are abIt can be expressed as the preference value of object a for item b, where object a represents the ath object in the object set U and item b represents the bth item in the item set V. Here, A, B, a, and b are all positive integers. If object a already holds item b, the above preference value ALL_SCORE can be determined as the element R in the preference matrix R. ab ; If object a does not hold item b, then the element R in the preference matrix R can be ab =-1. More intuitively, let Λ = {<object a, item b, R ab >|R ab ≠-1} represents the known preference value ALL_SCORE, Φ={<object a, item b, R ab >|R ab =-1} indicates an unknown preference value, that is, there is a lack of usage characteristics of object a for item b, and the preference value of object a for item b cannot be calculated. In this case, item b can be used as the item to be predicted corresponding to object a. When obtaining the item information fusion features corresponding to the above historical item set, the preference value of the object to be predicted for each item in the historical item set can be directly obtained from the preference matrix R, that is, the row corresponding to the object to be predicted is obtained from the preference matrix R. For example, if the object to be predicted is object 1 (that is, the first object in the above object set U), then the first row of the preference matrix R is reduced to R. 1b = -1 (the value range of b is 1, 2, ..., B) can be expressed as the preference value of the object to be predicted (the above-mentioned object 1) for each item in the historical item set; when obtaining the object information fusion feature corresponding to the above-mentioned historical object set, the preference value of each object in the historical object set for the item to be predicted can be directly obtained from the preference matrix R, that is, the column corresponding to the item to be predicted is obtained from the preference matrix R. For example, when the item to be predicted is item 4 (that is, the fourth item in the above-mentioned item set V), the fourth column of the preference matrix R is excluding R a4 = -1 (the value range of a is 1, 2, ..., A) can be expressed as the preference value of each object in the historical object set for the item to be predicted (the above-mentioned item 4).

[0120] It should be noted that the objects to be predicted, objects contained in the historical object set, socially associated objects, associated objects, etc. involved in this application all belong to the above-mentioned object set U = {object 1, object 2, ..., object A}, and the items to be predicted, items contained in the historical item set, similar items, etc. involved in this application all belong to the above-mentioned item set V = {item 1, item 2, ..., item B}.

[0121] Step S102 : constructing a social relationship sequence corresponding to the object to be predicted based on the socially associated objects associated with the object to be predicted at time t, and determining the social information features corresponding to the object to be predicted based on the social relationship sequence.

[0122] Specifically, the computer device can obtain the social related objects associated with the object to be predicted at time t. The social related objects at this time can be the friends of the object to be predicted in the application client. Of course, the computer device can also obtain the social related objects of the social related objects, that is, the indirect social related objects of the object to be predicted. The computer device can construct a related object sequence corresponding to the object to be predicted based on the social related objects and the indirect social related objects (such as the above Figure 2 The social relationship sequence corresponding to the object to be predicted can be constructed based on the associated objects, items held by the associated objects, and the associated objects' preferences for the items contained in the associated object sequence. The associated object sequence includes the socially associated objects and indirect socially associated objects of the object to be predicted. Elements in the social relationship sequence can be used to represent the associated objects' preferences for items, expressed in the form of evaluation triples.

[0123] It should be noted that the computer device can obtain the object set U = {object 1, object 2, ..., object A} in the application client, and according to the social relationships between the objects included in the object set U, a social relationship chain matrix H∈R for A objects can be constructed. A×A , that is, the size of the social relationship chain matrix is ​​A×A, where if object a and object i are in a social association relationship (also called a friend relationship), then the element H in the social relationship chain matrix H is ai =1; if object a and object i are not in a social relationship (also known as not a friend relationship), then the element H in the social relationship chain matrix H ai = 0. Wherein, object i represents the i-th object in the object set U, and the value range of i is 1, 2, ..., A, where i and a are different positive integers.

[0124] Optionally, the computer device may obtain the social associated objects associated with the object to be predicted at time t, generate an object relationship chain including the object to be predicted and the social associated objects, and perform a random walk (RW) in the object relationship chain starting from the object to be predicted to obtain a sequence of associated objects; and then obtain the associated object u in the associated object sequence. [m] , in the associated object u [m] Among the items held, the item with the greatest item similarity to the item to be predicted is determined as the similar item corresponding to the item to be predicted; the associated object u [m] , similar items and related objects u [m]The preference values ​​for similar items are determined as social sequence elements, and the social sequence elements associated with each associated object in the associated object sequence are determined as the social relationship sequence corresponding to the object to be predicted. In other words, the computer device can use a random walk algorithm to construct social association objects for the object to be predicted. A random walk is a mathematical statistical model consisting of a series of trajectories, each of which is random. A random walk can be used to represent irregular changes, such as the random process recorded by a person walking around drunkenly.

[0125] The process of constructing the associated object sequence using the random walk method may include: according to the above social relationship chain matrix H∈R A×A , obtain a set consisting of social associated objects corresponding to the object to be predicted. When the object to be predicted is object a (hereinafter referred to as object to be predicted a), the social associated objects of a can be formed into a set N(a). The set N(a) belongs to the above-mentioned object set U. According to the social association relationship between the social associated objects contained in the set N(a) and the object to be predicted a, an object relationship chain corresponding to the object to be predicted a can be generated; for any object aj∈N(a) in the set N(a), the transition probability from the object to be predicted a to the object aj can be expressed as P(u [1] = object aj) = 1 / |N(a)|, where |N(a)| is used to represent the number of socially associated objects of the object a to be predicted, and object aj is used to represent the jth object in the set N(a), where j is a positive integer less than or equal to |N(a)|. [1] = object a2, that is, moving from the object to be predicted a to the object a2 in the set N(a) (object a2 is the second object in the set N(a)), then according to the social relationship chain matrix H∈R A×A , obtain the set N2(a) consisting of the social associated objects of object a2, which also belongs to the above object set U, and add the objects in the set N2(a) to the above object relationship chain to update the above object relationship chain; for any object ad∈N2(a) in the set N2(a) (object ad is used to represent the d-th object in the set N2(a)), the transition probability from object a2 to object ad can be expressed as P(u [2] = object ad) = 1 / |N2(a)|, where |N2(a)| is used to represent the number of socially associated objects of object a2, and then u can be determined from the set N2(a) [2] , object ad is used to represent the d-th object in the set N2(a), where d is a positive integer less than or equal to |N2(a)|; by analogy, a sequence of associated objects {u [1] ,u [2] ,u [3] ,……,u [M]}, associated object sequence {u [1] ,u [2] ,u [3] ,……,u [M] Any object u in [m] They can all be called associated objects corresponding to the object to be predicted, u [m] It can be represented as the mth associated object in a sequence of associated objects.

[0126] For example, assuming that object 1 is the object to be predicted, if the socially associated objects of object 1 obtained from the object set U are: object 2, object 4, object 5, object 7, then objects 2, object 4, object 5, and object 7 can be combined into a set N(1) (i.e., the above set N(a)). The transition probability of moving from object 1 to any object in the set N(1) is 1 / 4. If u is determined [1] For object 4, we can obtain the socially associated objects of object 4 from the object set U, such as object 3, object 1, object 6, object 7, and object 8. Object 3, object 1, object 6, object 7, and object 8 form a set N(2) (i.e., the above set N2(a)). The transition probability of moving from object 4 to any object in the set N(2) is 1 / 5. Determine u in the set N(2) [2] is object 6; and so on, a sequence of associated objects {object 4, object 6, object 9, object 10} with a length of 4 (that is, the value of M above is 4) can be constructed.

[0127] Furthermore, for each associated object u in the associated object sequence [m] , computer devices can obtain the associated object u [m] The items held and in the associated object u [m] Select an item that is most relevant to the item to be predicted from the items held as a similar item to the item to be predicted (according to the similarity of the items, the items can be selected from the associated object u [m] Select similar items to the item to be predicted from the items held. The calculation formula for item similarity can refer to the following formula (5). The similar items at this time can be recorded as v [m] , and get the associated object u in the above preference matrix [m] For similar items v [m] The preference value at this time can be recorded as r [m] ; Then the associated object u [m] 、Similar itemsv [m] and the preference value r [m] Construct a social sequence element. According to the M associated objects in the associated object sequence, M social sequence elements can be obtained. The M social sequence elements are used to construct the social relationship sequence corresponding to the object to be predicted {(u [1] , v [1], r [1] ), (u [2] , v [2] , r [2] ), (u [3] , v [3] , r [3] ),……,(u [M] , v [M] , r [M] )}, as above Figure 2 The friend-prop sequence in the corresponding embodiment.

[0128] After constructing the social relationship sequence corresponding to the object to be predicted, the computer device can [m] 、Similar itemsv [m] and the associated object u [m] For similar items v [m] The preference value r [m] , obtain the social information features corresponding to the object to be predicted.

[0129] Optionally, for each object to be predicted, the above-mentioned random walk method can be used to construct multiple associated object sequences. For example, the number of associated object sequences can be P, where P is a positive integer, such as P can take the value of 1, 2, 3,..., and then P social relationship sequences can be constructed. According to the associated objects, similar items, and the preference values ​​of the associated objects for similar items contained in each social relationship sequence, the social output features corresponding to the P social relationship sequences are obtained, and the P social output features are subjected to feature fusion to obtain the social information features corresponding to the object to be predicted.

[0130] Step S103: Obtain the object features corresponding to the object to be predicted and the item features corresponding to the item to be predicted, and combine the item information fusion features, object information fusion features, social information features, object features to be predicted, and item features to be predicted into the cascade features to be predicted corresponding to time t.

[0131] Specifically, the computer device can obtain the item operation characteristics corresponding to the object to be predicted from the object log corresponding to the application client, wherein the item operation characteristics can refer to the item usage of the object to be predicted in the application client; for example, the 24 hours of a day can be divided into 12 time periods, each time period includes 2 hours, and the item operation characteristics at this time can be the item usage of the object to be predicted in different time periods. The object basic portrait corresponding to the object to be predicted is obtained in the application client, and the object characteristics corresponding to the object to be predicted are determined based on the item operation characteristics and the object basic portrait. When the object to be predicted is object a (hereinafter referred to as object to be predicted a), the object representation vector u corresponding to object to be predicted a can be determined based on the item operation characteristics and the object basic portrait corresponding to object to be predicted a a =[e1, e2, ..., e q ], where the object represents the vector u a It can include q-dimensional features, namely e f It can represent the f-th dimension feature of the object a to be predicted, where f is a positive integer less than or equal to q.

[0132] Since the dimensions of different features vary greatly, there are a large number of missing values ​​and outliers, and it is usually necessary to perform data preprocessing on the object to be predicted. Data preprocessing may include, but is not limited to, normalization, missing value completion, continuous feature discretization, one-hot encoding, and outlier removal; normalization may include, but is not limited to, normalization and maximum and minimum value normalization. For example, the object basic portrait corresponding to the object to be predicted a may include classification features (for example, gender features may be divided into three categories: male, female, and unknown). The computer device may use continuous feature discretization and one-hot encoding to convert the classification features in the object basic portrait into one-hot encoding to obtain the object representation vector u corresponding to the object to be predicted a. a .

[0133] Please also see Figure 4 , Figure 4 This is a schematic diagram of a process of converting classification features into one-hot encoding provided by an embodiment of the present application. Figure 4 As shown, the gender feature e in the basic portrait of the object may include: male, female, and unknown. The gender feature e is discretized into a continuous feature, and "male" is converted to 0, "female" is converted to 1, and "unknown" is converted to 2. The gender feature e' after the discretization of the continuous feature can then be one-hot encoded, and 0 is encoded as [0, 0, 1], 1 is encoded as [0, 1, 0], and 2 is encoded as [1, 0, 0] to obtain the one-hot encoded gender feature e".

[0134] The following takes the maximum and minimum normalization as an example to represent the object vector u a Perform maximum and minimum normalization, for the object representation vector ua Any one-dimensional eigenvalue e in f (also called target eigenvalue), you can get the object representation vector u a The maximum eigenvalue e in max and the minimum eigenvalue e min , and obtain the maximum eigenvalue e max and the minimum eigenvalue e min The difference between max -e min ), and the eigenvalue e f With the minimum eigenvalue e min The difference between f -e min ), the difference (e f -e min ) and the difference (e max -e min ) is determined as the normalized eigenvalue e f '. Among them, the object represents the vector u a The maximum and minimum normalization can be shown as the following formula (1):

[0135]

[0136] Among them, e in the above formula (1) f The object represents the vector u a The f-th dimension eigenvalue in, e f ' is the f-th dimension eigenvalue after maximum and minimum normalization; then, the object representation vector u can be expressed according to the above formula (1) a Perform maximum and minimum normalization, and then the object representation vector after maximum and minimum normalization can be input into the embedding layer. The object representation vector after maximum and minimum normalization is converted into a dense embedding vector through the embedding layer, and the embedding vector is determined as the object feature to be predicted of the object to be predicted a. At this time, the object feature to be predicted can be used a Optionally, the computer device can also directly determine the maximum and minimum normalized object representation vector as the object feature to be predicted of the object to be predicted a. a .

[0137] It can be understood that for any object in the above object set U = {object 1, object 2, ..., object A}, the above method can be used to determine the object features corresponding to each object, that is, the method used for the object features corresponding to each object is the same.

[0138] Optionally, the computer device may obtain attribute description information associated with the item to be predicted and convert the attribute description information into an initial attribute feature vector corresponding to the item to be predicted. For example, the categorical features in the attribute description information may be converted using continuous feature discretization and one-hot encoding, and the numerical values ​​in the attribute description information may be processed using methods such as missing value completion and outlier removal. For example, when the item to be predicted is a game prop in a game client, the attribute description information corresponding to the item to be predicted may include information such as the prop type, prop price, prop color, and prop skill; when the item to be predicted is a commodity in an e-commerce client, the attribute description information corresponding to the item to be predicted may include information such as the commodity type, commodity price, commodity color, commodity applicable population, and commodity style. The computer device can obtain the minimum eigenvalue and the maximum eigenvalue in the initial attribute feature vector corresponding to the item to be predicted, obtain the first difference between the maximum eigenvalue and the minimum eigenvalue, and the second difference between the target eigenvalue and the minimum eigenvalue, determine the ratio between the second difference and the first difference as the normalized target eigenvalue, and determine the item to be predicted feature corresponding to the item to be predicted based on the normalized target eigenvalue. The normalization here can be the maximum and minimum normalization shown in the above formula (1). The process of obtaining the item to be predicted feature corresponding to the item to be predicted is similar to the process of obtaining the object to be predicted feature corresponding to the object to be predicted. The item to be predicted feature can also refer to the embedding vector obtained by the embedding layer; of course, the item to be predicted feature can also refer to the initial attribute feature vector after maximum and minimum normalization.

[0139] It can be understood that for any item in the above item set V = {item 1, item 2, ..., item B}, the above method can be used to determine the item features corresponding to each item, that is, the method used to determine the item features corresponding to each item in the item set V is the same.

[0140] Furthermore, the computer device may combine the features of the object to be predicted, the features of the item to be predicted, the item information fusion features obtained at time t, the object information fusion features obtained at time t, and the social information features obtained at time t into a cascade feature to be predicted corresponding to time t, where t = 1, 2, ..., N. In other words, the computer device may obtain the item information fusion features, object information fusion features, and social information features corresponding to N different time periods. The item information fusion features, object information fusion features, and social information features corresponding to the N different time periods may be the same or different, and the features of the object to be predicted and the features of the item to be predicted corresponding to the N different time periods are all the same. Therefore, the computer device may compose the cascade features to be predicted corresponding to N different time periods based on the features of the object to be predicted, the features of the item to be predicted, the item information fusion features, object information fusion features, and social information features corresponding to the N different time periods.

[0141] Step S104: Determine the temporal migration characteristics between the object to be predicted and the item to be predicted based on the cascade feature sequence, and determine the recommended status of the item to be predicted for the object to be predicted based on the temporal migration characteristics; the cascade feature sequence includes cascade features to be predicted corresponding to N moments, where N moments include moment t, and N is a positive integer greater than 1.

[0142] Specifically, the computer device can form a cascade feature sequence with N cascade features to be predicted, and then can input the N cascade features to be predicted in the cascade feature sequence into a recurrent neural network in chronological order (the recurrent neural network in this case can be a unidirectional LSTM, or can also be called a second recurrent neural network), and output the temporal transition features between the N cascade features to be predicted through the recurrent neural network. According to the temporal transition features, the predicted preference value of the object to be predicted for the item to be predicted can be determined, and the recommendation status of the item to be predicted for the object to be predicted can be determined according to the predicted preference value. Among them, the recommendation status can include a recommendation success status and a recommendation failure status. The recommendation success status can be used to indicate that the item to be predicted is pushed to the object to be predicted, and the recommendation failure status can be used to indicate that the item to be predicted cannot be pushed to the object to be predicted.

[0143] Optionally, the number of items to be predicted may be L, where L is a positive integer, such as 1, 2, ..., and the items to be predicted may refer to the remaining items in the item set V excluding the items held by the subject to be predicted, or the items to be predicted may refer to the remaining items in the item set V excluding the items held by the subject to be predicted. According to the steps described above, the predicted preference values ​​of the subject to be predicted for each of the L items to be predicted may be obtained, and the predicted preference values ​​associated with the L items to be predicted may be sorted in descending order to obtain the sorted L items to be predicted. Among the sorted L items to be predicted, the recommendation status corresponding to the first y items to be predicted is set to a successful recommendation status, where y is a positive integer less than or equal to L. The y items to be predicted in the successful recommendation status are pushed to the subject to be predicted.

[0144] Optionally, in a product recommendation scenario, the set of historical items held by the object to be predicted may include products purchased by the object to be predicted in an e-commerce client, the set of historical objects holding the items to be predicted may include objects that purchased the items to be predicted, and the socially associated objects corresponding to the object to be predicted may be friends of the object to be predicted in the e-commerce client, friends in instant messaging applications, objects of concern, etc.

[0145] In an embodiment of the present application, the items held by the object to be predicted at time t can be fused with item information, that is, analyzed at the item level to obtain item information fusion features; the objects holding the items to be predicted at time t can be fused with object information, that is, analyzed at the object level to obtain object information fusion features; and the social information of the socially associated objects associated with the object to be predicted at time t can be obtained, that is, analyzed at the social relationship chain level to obtain social information features. Then, the item information fusion features, object information fusion features, social information features, object to be predicted features, and item to be predicted features obtained at N moments can be cascaded to obtain cascade features to be predicted. The temporal transition features between the cascade features to be predicted corresponding to N moments (including time t) can be obtained. Based on the temporal transition features, the recommendation status of the item to be predicted for the object to be predicted can be determined. In other words, in the item recommendation process, the present application can introduce the influence of indirect social associated objects (indirect friends) on the preferences of the object to be predicted through random walks, and introduce the migration of the interest preferences of the object to be predicted over time, which can enhance the matching degree between the object to be predicted and the recommended items, thereby improving the accuracy of item recommendations.

[0146] See Figure 5 , Figure 5 This is a flowchart of an item recommendation method provided by an embodiment of the present application. It is understandable that the data recommendation method can be executed by a computer device, which can be an independent server, or a user terminal, or a system consisting of a server and a user terminal, or a server cluster consisting of multiple servers, or a computer program application (including program code), which is not specifically limited here. Figure 5 As shown, the data recommendation method may include the following steps:

[0147] Step S201: obtain the historical item set held by the object to be predicted at time t, obtain the first evaluation triplet between the object to be predicted and the items contained in the historical item set; obtain the historical item feature corresponding to item k in the first evaluation triplet, and convert the preference value in the first evaluation triplet into a first preference vector associated with item k.

[0148] Specifically, the computer device can obtain the object to be predicted (such as the above Figure 2In the corresponding embodiment, the historical item set held by object 1) at time t can include all items held by the target object in the application client. The historical item set belongs to the item set V = {item 1, item 2, ..., item B} in the application client. For ease of description, the following detailed description uses the target object as object a in the object set U = {object 1, object 2, ..., object A}, i.e., the ath object in object set U. In this case, the historical item set held by target object a can be denoted as C(a).

[0149] For the object to be predicted a, the set {<object a, item k, R ak >|item k∈C(a)}, where item k is used to represent the kth item in the historical item set C(a), and the value of k is less than or equal to the number of items contained in the historical item set C(a). Set {<object a, item k, R ak Each element in >|item k∈C(a)}<object a, item k, R ak > can all be called the first evaluation triplet, the first evaluation triplet <object a, item k, R ak >It can be expressed as the preference value of the object a to be predicted for item k is R ak .

[0150] Among them, the first evaluation triple <object a, item k, R ak >The acquisition process may include: the computer device may obtain the object identifier corresponding to the object a to be predicted, obtain the item identifier corresponding to the item k in the historical item set, and then obtain the usage time interval (which may be called TI indicator) and usage frequency (which may be called FoU indicator) of the object a to be predicted for the item k, and determine the product of the usage time interval and the usage frequency as the item usage attenuation factor (which may be called TIFoU indicator) corresponding to the item k; determine the balance factor (which may be called BF indicator) corresponding to the item k according to the unit usage number of the object to be predicted for the item k, and determine the preference value R of the object to be predicted for the item k according to the usage time interval (TI indicator), the usage frequency (FoU indicator), the item usage attenuation factor (TIFoU indicator) and the balance factor (BF indicator). ak ; The object identifier, item identifier and the preference value R of the object a to be predicted for item k ak , determined as the first evaluation triplet between the object to be predicted a and item k <object a, item k, R ak >. Among them, the preference value R ak The specific calculation process can be found in the above Figure 3The step S101 in the corresponding embodiment will not be described here in detail. It is understandable that the computer device can pre-calculate the preference value of each object in the object set U for the items it holds, and form a preference matrix R (also see the above Figure 3 In the corresponding embodiment, step S101), the preference value R ak It can be directly obtained from the preference matrix R.

[0151] Furthermore, the computer device can obtain the historical item features corresponding to item k (as described above Figure 2 The prop vector 20c in the corresponding embodiment, the acquisition process of the historical item feature may include: the computer device may obtain the attribute description information associated with the item k, and convert the attribute description information of the item k into the initial attribute feature vector corresponding to the item k; obtain the minimum eigenvalue and the maximum eigenvalue in the initial attribute feature vector, obtain the first difference between the maximum eigenvalue and the minimum eigenvalue, and the second difference between the target eigenvalue and the minimum eigenvalue, wherein the target eigenvalue is any eigenvalue in the initial attribute feature vector; determine the ratio between the second difference and the first difference as the normalized target eigenvalue, and determine the historical item feature corresponding to the item k according to the normalized target eigenvalue. After normalizing each eigenvalue in the initial attribute feature vector, a normalized initial attribute feature vector can be obtained, and the normalized initial attribute feature vector can be converted into a dense embedding vector through the embedding layer. The embedding vector at this time can be called the historical item feature corresponding to the item k. In the embodiment of the present application, the historical item feature corresponding to the item k can be represented by item k The above normalization can be the maximum and minimum normalization shown in the above formula (1), that is, the historical item feature item corresponding to any item k in the historical item set C(a) k The acquisition process is the same as above Figure 3 The process of acquiring the features of the item to be predicted in the corresponding embodiments is the same.

[0152] For the first evaluation triplet value < object a, item k, R ak >The preference value R in ak , whose value range is [0,1]. In order to reduce the amount of calculation and improve the recommendation efficiency, the preference value R can be retained. ak One decimal place in the value, that is, the preference value R ak Execute round(R ak,1), at this time, the value range of the preference value with only one decimal place is {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0}, each value in {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0} can be passed through the embedding layer to obtain a dense embedding vector, retaining the preference value R of one decimal place. ak The corresponding embedding vector can be expressed as r k , that is, the first evaluation triple <object a, item k, R ak >The preference value R in ak Converted to the first preference vector r associated with item k k .

[0153] In step S202 , the first cascade feature between the historical item feature and the first preference vector is input into the first neural network layer in the recommendation model, and the item output feature corresponding to the first cascade feature is output through the first neural network layer.

[0154] Specifically, the computer device can obtain a recommendation model, which can be used to predict the predicted preference value of the predicted object for the predicted item. In order to better integrate the historical item features and the first preference vector r k , the historical item features and the first preference vector can be cascaded to obtain the first cascade feature, and then the first cascade feature is input into the first neural network layer in the recommendation model. The first neural network layer can be a multi-layer fully connected neural network (such as the above Figure 2 The multi-layer fully connected neural network 1) in the corresponding embodiment can output the item output feature corresponding to the first cascade feature through the first neural network layer. ak Each first evaluation triplet of <object a, item k, R>|item k∈C(a)} ak >, you can perform the following operations in the first neural network layer:

[0155]

[0156] Among them, g in formula (2) v It can be represented as the first neural network layer in the recommendation model, which can include more than two fully connected layers. It can represent the output result of the first neural network layer (that is, the output features of the above items), item k can be expressed as the historical item features corresponding to item k, and ⊕ can be expressed as the cascade operation of two vectors (or matrices). In other words, the computer device can calculate the set {< object a, item k, R akEach first evaluation triplet of <object a, item k, R>|item k∈C(a)} ak >The corresponding item output features, that is, one item in the historical item set C(a) corresponds to one item output feature.

[0157] Step S203 : obtaining the item weight corresponding to each item in the historical item set, and determining the item information fusion feature corresponding to the historical item set according to the item weight and the item output feature associated with each item in the historical item set.

[0158] Specifically, since the similarity between each item in the historical item set C(a) and the item to be predicted is different, each item in the historical item set C(a) can be assigned a different item weight, that is, the computer device can obtain the item weight corresponding to each item in the historical item set C(a); perform a weighted sum based on the item weight and the item output feature, and output the item information fusion feature corresponding to the historical item set C(a) through the first fully connected layer and the first activation layer. The item information fusion feature can be expressed as the following formula (3):

[0159]

[0160] Among them, O in the above formula (3) item It can be expressed as the above-mentioned item information fusion feature, σ is the sigmoid activation function (i.e. the first activation layer mentioned above), α ak It can be expressed as the item weight corresponding to item k in the historical item set C(a), W item is the coefficient matrix of the first fully connected layer, b item is the bias of the first fully connected layer, and can be expressed as a product operation between two vectors (or matrices).

[0161] Among them, the item weight α in the above formula (3) ak The item weight α can be calculated by the first attention network layer in the recommendation model. ak The calculation process may include: the computer device may obtain the item operation characteristics corresponding to the object to be predicted a in the object log, determine the object characteristics corresponding to the object to be predicted a according to the item operation characteristics and the object basic portrait of the object to be predicted a; and then determine the item characteristics corresponding to the object to be predicted and the historical item characteristics corresponding to the item k. k , obtain the item similarity between the item to be predicted and item k. The computer device can kAnd the item similarity is input into the first attention network layer in the recommendation model. According to the network parameter matrix of the first attention network layer, the features of the object to be predicted and the features of the item to be predicted are analyzed. k And the feature matrix determined by the item similarity is multiplied to obtain the item weight α corresponding to each item in the historical item set ak Among them, the item weight α ak The mathematical expression of is shown in the following formula (4):

[0162]

[0163] The first attention network layer can refer to a two-layer attention network (including the first network layer and the second network layer) that introduces prior information (such as item similarity), w 1,item 、b 1,item They are respectively represented as the network parameters and bias of the first network layer contained in the first attention network layer, w 2,item 、b 2,item is the network parameter and bias of the second network layer contained in the first attention network layer. At this time, w 1,item and w 2,item It can be called the network parameter matrix of the first attention network layer, It can be expressed as the above w 2,item The transposed matrix of ; σ is the sigmoid activation function, is the output result of the first neural network layer, ⊕ can be expressed as a cascade operation of two vectors (or matrices); the item weight corresponding to item k needs to be continuously iterated in the first attention network layer, and the intermediate iteration value in the iterative process can be used Indicates that the final item weight of item k is α ak Indicates; USER in formula (4) can be expressed as the feature of the object to be predicted corresponding to the object to be predicted (in the embodiment of the present application, USER is the object feature user corresponding to the object to be predicted a a , the feature of the object to be predicted is the embedding vector obtained by the embedding layer of the object to be predicted a), ITEM can represent the feature of the item to be predicted corresponding to the item to be predicted (it can refer to the embedding vector obtained by the embedding layer of the item to be predicted), ITEMCORR is the item similarity between the item to be predicted and the item k in the historical item set C(a), and ITEMCORR can be used as the prior information of the first attention network layer.

[0164] As shown in the above formula (4), the first attention network layer can be used to output the item weight corresponding to each item in the historical item set C(a). The input of the first attention network layer is the feature matrix obtained by cascading the item output features output by the above first neural network layer, the object features corresponding to the object to be predicted, the item features corresponding to the item to be predicted, and the item similarity between the item to be predicted and item k. USER⊕ITEM⊕ITEMCORR, the feature matrix USER⊕ITEM⊕ITEMCORR and the network parameter w of the first network layer 1,item Perform a product operation and add the result of the product operation to the bias b of the first network layer 1,item By adding, the output result of the first network layer can be obtained; the output result of the first network layer is nonlinearly processed by the sigmoid activation function, and then the output result after nonlinear processing can be input into the second network layer. According to the network parameter w of the second network layer 2,item and bias b 2,item , outputting the item weight corresponding to each item in the historical item set C(a). By introducing the first attention network layer, we can capture the influence of the historical item set held by the target on its preference for the target, thereby improving the accuracy of the predicted preference value of the target for the predicted item.

[0165] It should be noted that the process of obtaining the features of the object to be predicted, USER, and the features of the item to be predicted, ITEM, can be found in the above Figure 3 The description of step S103 of the corresponding embodiment will not be repeated here. The item similarity can be calculated by a similarity algorithm, which may include but is not limited to: Euclidean metric, Pearson correlation coefficient, cosine similarity. Taking cosine similarity as an example, the item similarity between the item to be predicted and the item k in the historical item set C(a) is calculated; if the item to be predicted is the b-th item in the item set V, and b is a positive integer less than or equal to B, then the item similarity between the item to be predicted b and the item k in the historical item set C(a) can be shown as the following formula (5):

[0166]

[0167] Among them, sim bk Indicates the item similarity between the item to be predicted (item b) and item k in the historical item set C(a), item k It can be expressed as the item feature corresponding to item k (i.e. historical item feature), item bIt can be expressed as the item feature corresponding to the b-th item in the item set V (that is, the item feature ITEM corresponding to the item to be predicted), ||item b || can be expressed as the feature item of the item to be predicted b The L2 norm of ||item k || can be represented as a historical item feature item k The L2 norm of .

[0168] Optionally, when calculating the item similarity between the item to be predicted and item k in the historical item set C(a), the item representation vector of the item to be predicted and the item representation vector of item k can also be used to calculate the item similarity between the two. The item representation vector here can refer to a representation vector that has been preprocessed but not converted by the embedding layer.

[0169] Step S204: obtain a set of historical objects that hold the item to be predicted at time t, obtain a second evaluation triplet between each object in the historical object set and the item to be predicted; obtain the historical object features corresponding to object i in the historical object set, and convert the preference value in the second evaluation triplet into a second preference vector corresponding to object i.

[0170] Specifically, the computer device may obtain a historical set of objects that held the item to be predicted at time t. This historical set of objects may include all objects in the application client that hold the item to be predicted. The historical set of objects belongs to the object set U = {object 1, object 2, ..., object A} in the application client. For ease of description, the following detailed description assumes that the item to be predicted is item b in the item set V = {item 1, item 2, ..., item B}, i.e., the bth object in item set V. The historical set of objects that hold item b in this case can be denoted as Z(b).

[0171] For the item b to be predicted, the set between the historical object set Z(b) and the item b to be predicted {<object i, item b, R ib >|object i∈Z(b)}, where object i is used to represent the i-th object in the historical object set Z(b), and the value of i is less than or equal to the number of objects contained in the historical object set Z(b). Set {<object i, item b, R ib >|Every element in object i∈Z(b)}<object i, item b, R ib > can be called the second evaluation triple, element <object i, item b, R ib >It can be expressed as the preference value R for object i to predict item b ib Among them, the second evaluation triple <object i, item b, R ibThe acquisition process of > is the same as that of the first evaluation triplet. For details, please refer to the above Figure 3 Step S101 in the corresponding embodiment will not be described in detail here.

[0172] Furthermore, the computer device can obtain the historical object features corresponding to the object i in the historical object set Z(b) (as described above). Figure 2 The object vector 20h in the corresponding embodiment, wherein the acquisition process of the historical object features is the same as the acquisition process of the above-mentioned object features to be predicted, and the specific description can be referred to the above-mentioned Figure 3 The step S103 in the corresponding embodiment will not be described here in detail. In the embodiment of the present application, the historical object feature corresponding to the object i can be recorded as user i , historical object features user i It can refer to the embedding vector output by the embedding layer.

[0173] For the second evaluation triple <object i, item b, R ib >The preference value R in ib , whose value range is [0,1]; in order to reduce the amount of calculation and improve the efficiency of recommendation, only the preference value R can be retained. ib One decimal place in the preference value R ib Execute round(R ib ,1), at this time, the value range of the preference value with only one decimal place is {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0}, and the preference value R with only one decimal place will be retained through the embedding layer ib Converted to a dense embedding vector, the dense embedding vector can be recorded as r i , that is, the second evaluation triple <object i, item b, R ib >The preference value R in ib Converted to the second preference vector associated with object i.

[0174] Step S205: input the second cascade feature between the historical object feature and the second preference vector into the second neural network layer in the recommendation model, and output the object output feature corresponding to the second cascade feature through the second neural network layer.

[0175] Specifically, in order to better integrate historical object features user i and the second preference vector r i , you can use the historical object features user i and the second preference vector r iCascade to obtain the second cascade feature, and then the second cascade feature can be input into the second neural network layer in the recommendation model. The second neural network layer can be a multi-layer fully connected neural network (such as the above Figure 2 The multi-layer fully connected neural network 2) in the corresponding embodiment can output the object output feature corresponding to the second cascade feature through the second neural network layer. ib Each second evaluation triplet <object i, item b, R>|object i∈Z(b)} ib >, the following operations can be performed in the second neural network layer:

[0176]

[0177] Among them, g in formula (6) u It can be represented as the second neural network layer in the recommendation model, which can include more than two fully connected layers. It can represent the output result of the second neural network layer (that is, the output features of the above objects), user i It can be expressed as the historical object feature corresponding to object i in the historical object set Z(b), and ⊕ can be expressed as the cascade operation of two vectors (or matrices). In other words, the computer device can calculate the set {< object i, item b, R ib Each second evaluation triplet <object i, item b, R>|object i∈Z(b)} ib >The corresponding object output features, that is, one object in the historical object set Z(b) corresponds to one object output feature.

[0178] Step S206 , obtaining the object weight corresponding to each object in the historical object set, and determining the object information fusion feature corresponding to the historical object set according to the object weight and the object output feature associated with each object in the historical object set.

[0179] Specifically, since the similarity between each object in the historical object set Z(b) and the object to be predicted a is different, a different object weight can be assigned to each object in the historical object set Z(b), that is, the computer device can obtain the object weight corresponding to each object in the historical object set Z(b); perform weighted summation based on the object weight and the object output feature, and output the object information fusion feature corresponding to the historical object set Z(b) through the second fully connected layer and the second activation layer. The object information fusion feature can be shown as the following formula (7):

[0180]

[0181] Among them, O in the above formula (7)user It can be expressed as the above object information fusion feature, σ is the sigmoid activation function, γ ib It can be expressed as the object weight corresponding to object i in the historical object set Z(b), W user is the coefficient matrix of the second fully connected layer, b user is the bias of the second fully connected layer, and can be expressed as a product operation between two vectors (or matrices).

[0182] Among them, the object weight γ in the above formula (7) ib The object weight γ can be calculated by the second attention network layer in the recommendation model. ib The calculation process may include: the computer device may calculate the predicted object features corresponding to the predicted object and the historical object features user corresponding to the object i. i , obtain the object similarity between the object to be predicted and object i; input the features of the object to be predicted, the features of the item to be predicted corresponding to the item to be predicted, and the object similarity into the second attention network layer in the recommendation model; then, according to the network parameter matrix of the second attention network layer, the feature matrix determined by the features of the object to be predicted, the features of the item to be predicted, and the item similarity can be multiplied to obtain the object weight corresponding to each object in the historical object set Z(b). Among them, the object weight γ ib The mathematical expression of is shown in the following formula (8):

[0183]

[0184] The second attention network layer can refer to a two-layer attention network (including the third network layer and the fourth network layer) that introduces prior information (such as object similarity), w 1,user 、b 1,user They are respectively represented as the network parameters and bias of the third network layer contained in the second attention network layer, w 2,user 、b 2,user is the network parameter and bias of the fourth network layer contained in the second attention network layer. At this time, w 1,user and w 2,user It can be called the network parameter matrix of the second attention network layer, It can be expressed as the above w 2,user The transposed matrix of ; σ is the sigmoid activation function, is the output result of the second neural network layer (i.e., the object output feature), ⊕ can be expressed as a cascade operation of two vectors (or matrices); the object weight corresponding to object i needs to be continuously iterated in the second attention network layer, and the intermediate iteration value in the iterative process can be used Indicates that the object weight finally obtained by object i is γib Indicates; USER in formula (8) can be expressed as the feature of the object to be predicted corresponding to the object to be predicted (in the embodiment of the present application, USER is the feature of the object to be predicted corresponding to the object to be predicted a, user a , the feature of the object to be predicted is the embedding vector obtained by the embedding layer of the object to be predicted), ITEM can represent the feature of the item to be predicted corresponding to the item to be predicted (in the embodiment of this application, ITEM is the item feature item corresponding to the item to be predicted b b , the feature of the item to be predicted is the embedding vector obtained by the embedding layer of the item to be predicted b), USERCORR is the object similarity between the object to be predicted and the object i in the historical object set Z(b), USERCORR can be used as the prior information of the second attention network layer.

[0185] As shown in the above formula (8), the second attention network layer can be used to output the object weight corresponding to each object in the historical object set Z(b). The input of the second attention network layer is the feature matrix obtained by cascading the object output features output by the above second neural network layer, the object features corresponding to the object to be predicted, the item features corresponding to the item to be predicted, and the object similarity between the object to be predicted and object i. USER⊕ITEM⊕USERCORR, the feature matrix USER⊕ITEM⊕USERCORR and the network parameter w of the third network layer 1,user Perform a product operation and add the result of the product operation to the bias b of the third network layer 1,user By adding, the output result of the third network layer can be obtained; the output result of the third network layer is nonlinearly processed by the sigmoid activation function, and then the output result after nonlinear processing can be input into the fourth network layer. According to the network parameter w of the fourth network layer 2,user and bias b 2,user , outputting the object weight corresponding to each object in the historical object set Z(b). By introducing a second attention network layer, we can capture the influence of the historical object set holding the to-be-predicted item b on the to-be-predicted object’s preference, thereby improving the accuracy of the predicted preference value of the to-be-predicted object towards the predicted item.

[0186] The object similarity is calculated in the same way as the item similarity. For example, the above formula (5) can be used to calculate the predicted object feature user corresponding to the predicted object a. a The user corresponding to object i iCalculate and obtain the object similarity between the object to be predicted a and object i in the historical object set Z(b). Optionally, when calculating the object similarity between the object to be predicted a and object i, the object representation vector of the object to be predicted a and the object representation vector of object i can also be used to calculate the item similarity between the two. The object representation vector here can refer to the representation vector that has been preprocessed but not transformed by the embedding layer.

[0187] Step S207: Construct a social relationship sequence of the object to be predicted based on the social associated objects associated with the object to be predicted at time t, obtain associated object features corresponding to the associated objects in the social relationship sequence, and similar item features corresponding to similar items in the social relationship sequence, and convert the preference values ​​in the social relationship sequence into a third preference vector.

[0188] Specifically, the computer device can obtain the social associated objects associated with the object to be predicted a at time t, and use the random walk method to process the social associated objects associated with the object to be predicted a at time t, and can construct an associated object sequence of length M; for each associated object in the associated object sequence, the item most relevant to the item to be predicted b can be selected from the items held by the associated object, and the most relevant item can be determined as a similar item to the item to be predicted b. Based on the associated objects, similar items, and the associated object's preference value for similar items contained in the associated object sequence, a social relationship sequence corresponding to the object to be predicted is constructed. Among them, similar items can refer to items held by the associated object that have the greatest item similarity with the item to be predicted b. The specific construction process of the social relationship sequence can be referred to above. Figure 3 The corresponding embodiment describes step S102.

[0189] For the social relationship sequence {(u [1] , v [1] , r [1] ), (u [2] , v [2] , r [2] ),……,(u [M] , v [M] , r [M] )} in any social sequence element (u [m] , v [m] , r [m] ), where m is a positive integer less than or equal to M. The computer device can obtain the associated object u in the social relationship sequence [m] The corresponding associated object feature y′ [m] 、Similar itemsv [m] Corresponding similar item features v′ [m]The acquisition process of the associated object features can refer to the acquisition process of the features of the object to be predicted, and the acquisition process of the similar item features is the same as the acquisition process of the features of the item to be predicted, which will not be repeated here. [m] It can be converted into a dense embedding vector through the embedding layer. At this time, the dense embedding vector can be called the third preference vector r′ [m] It should be noted that the associated object u [m] 、Similar itemsv [m] and the preference value r [2] The corresponding embedding layers can remain consistent in the recommendation model; of course, the embedding layers corresponding to any object in the object set U, any item in the item set V, and any preference value in the preference matrix can all remain consistent in the recommendation model.

[0190] Optionally, for each object to be predicted, P social relationship sequences can be constructed according to the above method, where P is a positive integer, such as 1, 2, 3, ..., and the P social relationship sequences corresponding to the object to be predicted can be expressed as in represents the ρth social relationship sequence corresponding to the object to be predicted, Associated with the object to be predicted and the item to be predicted, it can also be simplified to S ρ .

[0191] In step S208, the third cascade feature among the associated object feature, the similar item feature, and the third preference vector is input into the third neural network layer in the recommendation model, and the sequence output feature corresponding to the third cascade feature is output through the third neural network layer.

[0192] Specifically, the computer device can cascade the associated object feature, the similar item feature and the third preference vector to obtain a third cascade feature, and then input the third cascade feature into the third neural network layer in the recommendation model. The third neural network layer can be a multi-layer fully connected neural network (such as the above Figure 2 The multi-layer fully connected neural network 3) in the corresponding embodiment can output the sequence output feature corresponding to the third cascade feature through the third neural network layer. The following operations can be performed in the third neural network layer:

[0193]

[0194] Among them, for the social relationship sequence Each social relationship element in can be calculated by the above formula (9) to obtain the sequence output features corresponding to each social relationship sequence; g in the above formula (9)s It can be represented as the third neural network layer in the recommendation model, which can include more than two fully connected layers; u′ [m] Can be represented as an associated object u [m] The corresponding associated object feature, v′ [m] Can be expressed as similar items v [m] Corresponding similar item features, r′ [m] It can be expressed as the preference value r [m] The corresponding third preference vector,⊕,can be expressed as the concatenation operation of two vectors (or matrices); It can be expressed as the ρth social relationship sequence The mth social relationship element (u [m] , v [m] , r [m] ), that is, each social relationship element in the social relationship sequence corresponds to a sequence output feature.

[0195] In step S209, the sequence output features are input into the first recurrent neural network in the recommendation model, and bidirectional recurrent encoding is performed on the sequence output features according to the first recurrent neural network to obtain forward encoding features and reverse encoding features, which are then concatenated into implicit state features.

[0196] Specifically, in order to extract the abstract features of P social relationship sequences, the M sequence output features corresponding to each social relationship sequence can be sequentially input into the first recurrent neural network in the recommendation model. The first recurrent neural network can be a bidirectional LSTM. According to the bidirectional LSTM, the M sequence output features are bidirectionally cyclically encoded to obtain forward encoding features and reverse encoding features. The forward encoding features and reverse encoding features are spliced ​​into implicit state features. The splicing here can be understood as a cascade operation. Among them, the bidirectional LSTM is used to encode the ρth social relationship sequence. The processing of the corresponding M sequence output features is as follows:

[0197]

[0198] Among them, the above formula (10) is the calculation result of the above formula (9), It can be expressed as the ρth social relationship sequence The mth social relationship element (u [m] , v [m] , r [m] ) forward encoding features in the bidirectional LSTM, It can be expressed as the ρth social relationship sequence The mth social relationship element (u[m] , v [m] , r [m] ) reverse encoding features in bidirectional LSTM; It can be expressed as the ρth social relationship sequence The mth social relationship element (u [m] , v [m] , r [m] ) corresponds to the hidden state feature; ⊕ can be expressed as a cascade operation between two vectors (or matrices).

[0199] Step S210 , obtaining the social weight corresponding to each associated object in the social relationship sequence, and determining the social information feature corresponding to the object to be predicted based on the social weight and the implicit state feature.

[0200] Specifically, since socially associated objects with different degrees of intimacy or interest similarity have different influences on the preferences of the predicted objects, different social weights can be assigned to the associated objects in the social relationship sequence. That is, the computer device can obtain the social weight corresponding to each associated object in the P social relationship sequences. Then, a weighted sum can be performed based on the social weight and the implicit state feature, and the social output features corresponding to the P social relationship sequences can be output through the third fully connected layer and the third activation layer. The social output features can be shown as the following formula (11):

[0201]

[0202] Among them, the above formula (11) It can be expressed as the ρth social relationship sequence The corresponding social output features, σ is the sigmoid activation function, β m,ρ It can be expressed as the mth associated object in the ρth social relationship sequence (i.e., associated object u [m] ) corresponding to the social weight, is the coefficient matrix of the third fully connected layer above, is the bias of the third fully connected layer, and can be expressed as a product operation between two vectors (or matrices).

[0203] Among them, the social weight β in the above formula (11) m,ρ The social weight β can be calculated by the third attention network layer in the recommendation model. m,ρThe calculation process may include: the computer device may obtain the social similarity between the object to be predicted and each associated object based on the associated object features corresponding to each associated object in the social relationship sequence and the predicted object features corresponding to the object to be predicted; and then determine the social intimacy between the object to be predicted and each associated object based on the social parameters between the object to be predicted and each associated object, wherein the social parameters may include but are not limited to: social time, number of common social objects, interaction time and interaction frequency. The computer device may input the features of the object to be predicted, social similarity, social intimacy and the features of the item to be predicted corresponding to the item to be predicted into the third attention network layer in the recommendation model, and perform a product operation on the feature matrix determined by the features of the object to be predicted, social similarity, social intimacy and the features of the item to be predicted according to the network parameter matrix of the third attention network layer to obtain the social weight corresponding to each associated object. Among them, the social weight β corresponding to the mth associated object in the ρth social relationship sequence is m,ρ The mathematical expression of is shown in the following formula (12):

[0204]

[0205] The third attention network layer can refer to a two-layer attention network (including the fifth network layer and the sixth network layer) that introduces prior information (such as social similarity and intimacy), w 1,social 、b 1,social They are respectively represented as the network parameters and bias of the fifth network layer contained in the third attention network layer, w 2,social 、b 2,social is the network parameter and bias of the sixth network layer contained in the third attention network layer. At this time, w 1,social and w 2,social It can be called the network parameter matrix of the third attention network layer, It can be expressed as the above w 2,social The transposed matrix of ; tanh is the tanh activation function, is the output result (i.e., social output feature) of the above-mentioned bidirectional LSTM (i.e., the first recurrent neural network), ⊕ can be expressed as a cascade operation of two vectors (or matrices); the associated object u [m] The corresponding social weights need to be iterated continuously in the third attention network layer, and the intermediate iteration values ​​in the iterative process can be used Indicates that the associated object u [m] The final social weight is β m,ρ Indicates; USER in formula (12) can be expressed as the feature of the object to be predicted corresponding to the object to be predicted (in the embodiment of the present application, USER is the feature of the object to be predicted corresponding to the object to be predicted a, user a, the feature of the object to be predicted is the embedding vector obtained by the embedding layer of the object to be predicted), ITEM can represent the feature of the item to be predicted corresponding to the item to be predicted (in the embodiment of this application, ITEM is the item feature item corresponding to the item to be predicted b b , the feature of the item to be predicted is the embedding vector obtained by the embedding layer of the item to be predicted b), USERCORR is the associated object u in the ρth social relationship sequence between the object to be predicted and [m] The social similarity between the object to be predicted and the associated object u in the ρth social relationship sequence is CLOSE. [m] The social intimacy between them, USERCORR and CLOSE at this time can be used as prior information of the third attention network layer.

[0206] As shown in the above formula (12), the third attention network layer can be used to output the social weight corresponding to each associated object in the P social relationship sequence. The input of the third attention network layer is the feature matrix obtained by cascading the social output features of the above bidirectional LSTM output, the predicted object features corresponding to the predicted object, the predicted item features corresponding to the predicted item, the social similarity between the predicted object and the associated object, and the social intimacy between the predicted object and the associated object. The feature matrix and the network parameter w of the fifth network layer 1,social Perform a product operation and add the result of the product operation to the bias b of the fifth network layer 1,social By adding, the output result of the fifth network layer can be obtained; the output result of the fifth network layer is nonlinearly processed by the tanh activation function, and then the output result after nonlinear processing can be input into the sixth network layer. According to the network parameter w of the sixth network layer 2,social and bias b 2,social , outputting the social weights corresponding to each associated object in the P social relationship sequences. By introducing the third attention network layer, we can capture the influence of social relationships on the preferences of the predicted objects, thereby improving the accuracy of the predicted preferences of the predicted objects for the predicted items.

[0207] It is understandable that the direct friends (also called social associated objects) and indirect friends (also called indirect social associated objects) of the object to be predicted in the application client can both have an impact on the preference of the object to be predicted. The more distant the social relationship, the lower the influence on the preference of the object to be predicted, and the closer the social relationship, the greater the influence on the preference of the object to be predicted. Therefore, the computer device can obtain the social parameters between the object to be predicted and each associated object in the social relationship sequence. For example, the social parameters may include the social duration S1 (such as the duration of being friends), the number of common social objects D2 (such as the number of common friends), the interaction duration D3 (such as the duration of common games), and the interaction frequency D4 (such as the frequency of chatting). Then, the social duration D1, the number of common social objects D2, the interaction duration D3, and the interaction frequency D4 can be normalized to the maximum and minimum, and the weight corresponding to the social duration D1 can be calculated through the hierarchical analysis method. The weight corresponding to the number of common social objects D2 The weight corresponding to the interaction time D3 And the weight corresponding to the interaction frequency D4 The computer device can calculate the social intimacy between the object to be predicted and the associated objects in the social relationship sequence through weighted summation: This application does not limit the parameter types and number of parameters included in the social parameters. It should be noted that the social similarity between the predicted object and each associated object is obtained in the same process as the aforementioned object similarity, which will not be repeated here.

[0208] Furthermore, for Each social relationship sequence in The social output features can be calculated by the above formulas (11) and (12): ρ = 1, 2, ..., P. The computer device can fuse the social output features corresponding to the P social relationship sequences to obtain the social information features corresponding to the object to be predicted. The implementation process is as follows: the P social output features can be fused through a multi-layer fully connected neural network 4. The specific mathematical representation is shown in the following formula (13):

[0209]

[0210] Among them, g′ s It can be represented as the above multi-layer fully connected neural network 4, It can be represented as the social information features corresponding to the object to be predicted. In other words, the P social output features can be cascaded and input into the multi-layer fully connected neural network 4. The multi-layer fully connected neural network 4 can then output the social information features corresponding to the object to be predicted.

[0211] In step S211 , the item information fusion feature, the object information fusion feature, the social information feature, the object feature to be predicted corresponding to the object to be predicted, and the item feature to be predicted corresponding to the item to be predicted are combined into a cascade feature to be predicted corresponding to time t.

[0212] The specific implementation process of step S211 can be found in the above Figure 3 The description of step S103 in the corresponding embodiment will not be repeated here.

[0213] In step S212, the cascade features to be predicted corresponding to the N moments are combined into a cascade feature sequence, the cascade features to be predicted contained in the cascade feature sequence are sequentially input into the fourth neural network layer in the recommendation model, and the network output features corresponding to the N cascade features to be predicted are output through the fourth neural network layer.

[0214] Specifically, the computer device can combine the cascade features to be predicted corresponding to N moments into a cascade feature sequence. For any cascade feature to be predicted corresponding to moment t in the cascade feature sequence (t=1, 2, ..., N), each can be input into the fourth neural network layer in the recommendation model. The fourth neural network layer can be a multi-layer fully connected neural network. The fourth neural network layer can output the network output features corresponding to the N cascade features to be predicted corresponding to the cascade feature to be predicted. Among them, the cascade feature to be predicted corresponding to moment t can be formed by cascading the item information fusion feature obtained at moment t, the object information fusion feature obtained at moment t, the social information feature obtained at moment t, the object feature to be predicted, and the item feature to be predicted. For the cascade feature to be predicted corresponding to moment t, the following operations can be performed in the fourth neural network layer:

[0215]

[0216] Among them, g in the above formula (14) t It can be represented as the fourth neural network layer in the recommendation model, which can include more than two fully connected layers; USER⊕ITEM can be expressed as the cascade features to be predicted corresponding to time t, where It can represent the fusion features of item information obtained at time t, It can be expressed as the social information features obtained at time t, It can be represented as the object information fusion feature obtained at time t, and USER can be represented as the object feature to be predicted corresponding to the object to be predicted (in the embodiment of the present application, USER can be the object feature user corresponding to object a). a, the feature of the object to be predicted is the embedding vector obtained by the embedding layer of the object to be predicted), ITEM can represent the feature of the item to be predicted corresponding to the item to be predicted (in the embodiment of the present application, ITEM can be the item feature item corresponding to item b) a , the feature of the item to be predicted is the embedding vector obtained by the item to be predicted through the embedding layer); ⊕ can be expressed as a cascade operation of two vectors (or matrices); O t It can be expressed as the network output feature corresponding to time t.

[0217] In step S213, the N network output features are sequentially input into the second recurrent neural network in the recommendation model, and the network output features are forward cyclically encoded according to the second recurrent neural network to obtain network implicit features.

[0218] Specifically, in order to learn the rules of interest migration and social object relationship chain changes of the predicted object, the computer device can input the network output features corresponding to time t (t = 1, 2, ..., N) into the second recurrent neural network in the recommendation model. The second recurrent neural network can be a unidirectional LSTM. The network output features corresponding to time t are forward-loop encoded according to the second recurrent neural network, and the network implicit features can be output. The network implicit features can be expressed by the following formula (15):

[0219]

[0220] Among them, O in the above formula (15) t It can be expressed as the network output feature output by the fourth neural network layer at time t. It can be expressed as the network implicit feature obtained by using the second recurrent neural network to perform forward cyclic encoding on the network output feature at time t (that is, the network implicit feature output at time t, which can also be called the hidden layer output result in the second recurrent neural network). The network implicit features corresponding to N moments can be calculated using the above formula (15). The network implicit feature output at the previous moment can be used as the input information of the second recurrent neural network at the next moment. For example, the network implicit feature obtained at time t-1 can be used as the input information of the second recurrent neural network at time t.

[0221] In step S214, based on the network implicit features and the hidden layer weights corresponding to the second recurrent neural network, the temporal transition features corresponding to the cascade feature sequence are obtained; based on the temporal transition features, the predicted preference value of the object to be predicted for the item to be predicted is obtained, and based on the predicted preference value, the recommendation status of the item to be predicted for the object to be predicted is determined.

[0222] Specifically, since the object interests and social object relationship chains at different times have different effects on the preferences of the predicted objects, different weights can be assigned to the hidden layers in the second recurrent neural network, that is, the hidden layer weights corresponding to the first recurrent neural network are obtained; a global sum is performed based on the network hidden features and the hidden layer weights, and the temporal migration features corresponding to the cascade feature sequence are output through the fourth fully connected layer and the fourth activation layer, that is, the temporal migration features between N network hidden features; the temporal migration features are input into the multi-layer neural network in the recommendation model, and the predicted preference value of the object to be predicted for the predicted item can be output through the multi-layer neural network. The predicted preference value can be shown as the following formula (16):

[0223]

[0224] Among them, the above formula (16) It can be expressed as the network implicit feature output by the second recurrent neural network at time t, θ t It can be expressed as the hidden layer weight in the second recurrent neural network, σ is the sigmoid activation function (i.e. the fourth activation layer mentioned above), W is the coefficient matrix of the fourth fully connected layer, b o is the bias of the fourth fully connected layer. It can be expressed as the temporal migration feature between N network implicit features; g o It can be represented as a multi-layer neural network in a multi-layer neural network recommendation model, and O can be represented as the predicted preference value of the object to be predicted to the item to be predicted.

[0225] Among them, the hidden layer weight θ in the above formula (16) is t It can be calculated by the fourth attention network layer in the recommendation model, and the hidden layer weight θ t The mathematical expression of is shown in the following formula (17):

[0226]

[0227] The fourth attention network layer can be a double-layer attention network (including the seventh network layer and the eighth network layer), and w in the above formula (17) is 1,t 、b 1,t They are respectively represented as the network parameters and bias of the seventh network layer contained in the fourth attention network layer, w 2,t 、b 2,t is the network parameter and bias of the seventh network layer contained in the fourth attention network layer. At this time, w 1,t and w 2,t It can be called the network parameter matrix of the fourth attention network layer, It can be expressed as the above w 2,tThe transposed matrix of ; tanh is the tanh activation function, · represents the product operation between two vectors (or matrices); the hidden layer weights of the second recurrent neural network need to be continuously iterated in the fourth attention network layer, and the intermediate iteration values ​​in the iterative process can be used Indicates that the final hidden layer weight of the second recurrent neural network is recorded as θ t By introducing the fourth attention layer, the hidden layer weights in the second recurrent neural network can be output, taking into account the influence of object interest migration and social object relationship chain on the preference of the predicted object, which can improve the accuracy of the predicted preference value of the predicted object for the predicted item.

[0228] It is understandable that for a certain object to be predicted, the number of items to be predicted is usually multiple (e.g., the object to be predicted has L items to be predicted). Through the above steps S201 to S214, the predicted preference value between the object to be predicted and each item to be predicted can be calculated. The computer device can sort the L items to be predicted according to the predicted preference value and push a more accurate list of recommended items to the object to be predicted based on the sorting result. The process of determining the recommended item list can be represented by the following formula (18):

[0229] Y1=sort(f(μ)) (18)

[0230] Here, f(·) represents the aforementioned recommendation model, which is essentially a deep learning network model. sort(·) represents the descending sorting of the predicted preference values ​​output by the recommendation model by object group. μ represents the social object relationship chain information obtained from the application client and the object's preference value for the item it holds. Y1 in the above formula (18) represents the system recommendation result determined from the sorted items to be predicted. In other words, the recommendation status of the items to be predicted included in the system recommendation result can be determined as a successful recommendation status, and the items to be predicted included in the system recommendation result can be pushed to the object to be predicted.

[0231] It's important to note that before a recommendation model can be used in an app client to recommend items, it must be trained. This trained model can then be used to predict the preference values ​​of the predicted object for the item being recommended. The following describes the training process for a recommendation model, using object and item information in a game client as an example.

[0232] The computer device can obtain the object set (the object set in this case is the player set in the game client) and the item set (the item set in this case is the prop set in the game client) in the game client, and then calculate the actual preference value of each object for the items it holds, and construct the preference matrix R based on the calculated preference values ​​(as shown above). Figure 3The computer device may also obtain an object portrait corresponding to each object in the object set, and obtain item operation characteristics corresponding to each object in the object set from the object log corresponding to the game client. Each object may be structured based on the object portrait and item operation characteristics to obtain an object feature vector corresponding to each object. Similarly, the computer device may obtain attribute description information corresponding to each item in the item set, structure each item, and obtain an item feature vector corresponding to each item. The computer device may calculate object similarity between different objects based on the object feature vectors, and item similarity between different items based on the item feature vectors. The preference matrix R, object feature vectors, item feature vectors, object similarity, and item similarity calculated above serve as preparatory work before model training.

[0233] Furthermore, the computer device can obtain an initialized recommendation model (hereinafter referred to as the initial recommendation model), and use the <object, item> in the game client as the sample data pair of the initial recommendation model, and the object in each sample data pair holds the item in the sample data pair. Since the above recommendation model is a trained initial recommendation model, that is, the network structure of the initial recommendation model is the same as that of the above recommendation model, during the training of the initial recommendation model, the processing of the sample data pairs is the same as the processing of the aforementioned <object to be predicted, item to be predicted>, and the processing of the sample data pairs will not be repeated here; only after the initial recommendation model outputs the sample predicted preference value of the object to the item in the sample data pair, the network parameters of the initial recommendation model can be adjusted by the error between the sample predicted preference value of the object to the item in the sample data pair and the actual preference value. In other words, the computer device can construct a loss function Loss for the initial recommendation model by the error between the sample predicted preference value corresponding to the sample data pair and the actual deviation value. The loss function Loss can be expressed as:

[0234]

[0235] Among them, the above formula (19) It can be expressed as the object and item in the sample data pair, R′ ij It can be expressed as the sample prediction preference value of object i to item j in the sample data pair <object i, item j>, R ijIt can be expressed as the actual preference value of object i to item j in the sample data pair <object i, item j>. In the embodiment of the present application, for the time t=1, 2, ..., N during the training process, the network parameters of the initial recommendation model are the same; the training process of the initial recommendation model can use an optimization algorithm to optimize the loss function Loss, correct the network parameters in the initial recommendation model by minimizing the loss function Loss, save the network parameters when the loss function Loss is at the minimum value, and determine the initial recommendation model at this time as the recommendation model, which can be applied to the prop recommendation scenario of the game client. Among them, the optimization algorithm may include but is not limited to: Adam optimization algorithm (an extension of the stochastic gradient descent algorithm), stochastic gradient descent algorithm (Stochastic Gradient Descent, SGD), Quasi-Newton Methods, Conjugate Gradient Method.

[0236] See Figure 6 , Figure 6 This is a flow chart of an item recommendation method provided in an embodiment of the present application. Figure 6 As shown, the implementation process of the item recommendation method may include the following steps S1 to S5.

[0237] Step S1: Obtain system flow data. The computer device can obtain system flow data from the application client. The system flow data may include basic object information, item attribute information, the time when the object first used an item (FIRST_DATE), the time when the object last used an item (LAST_DATE), the total number of times the item was used (LAST_DATE), and the number of times the item was used per unit time (PV_X, where X = 1, 3, 5, 7, 10, 15, 20, 30).

[0238] Step S2: Extracting and structuring portrait attributes: The computer device can extract portrait attributes based on the system flow data and perform structured processing on the extracted portrait attributes to obtain the basic object portrait corresponding to each object.

[0239] Step S3, object-item preference calculation model. The computer device can use the object-item preference calculation model to calculate the object's preference value for the item it holds. The calculation method of the preference value can refer to the above Figure 3 The corresponding step S101 in the embodiment.

[0240] Step S4, object portrait / item feature data preprocessing. The computer device can perform data preprocessing on the object basic portrait, the object item operation features and the item attribute features of the object. For example, data preprocessing may include methods such as continuous feature discretization, one-hot encoding and maximum and minimum normalization. The computer device can use continuous feature discretization and one-hot encoding to convert the item operation features corresponding to the object and the classification features (such as gender features) contained in the object basic portrait, to construct the object representation vector, and then use maximum and minimum normalization to normalize the object representation vector to obtain the preprocessed object representation vector; similarly, continuous feature discretization and one-hot encoding can be used to convert the classification features (such as item style features) contained in the item attribute information to construct the item representation vector, and then use maximum and minimum normalization to normalize the item representation vector to obtain the preprocessed item representation vector. Among them, the data preprocessing process can refer to the above Figure 3 The corresponding step S103 in the embodiment.

[0241] Step S5: Calculate object / item similarity. The computer device can use a similarity algorithm (e.g., a cosine similarity algorithm) to calculate the preprocessed object representation vectors to obtain object similarity between different objects. Furthermore, the same similarity algorithm can be used to calculate the preprocessed item representation vectors to obtain object similarity between different items.

[0242] Step S6: Obtain a deep learning model. The computer device can use object similarity and item similarity as prior information to construct a deep learning model (also known as a recommendation model). The specific network structure of the deep learning model is as follows: Figures 7a to 7d shown.

[0243] Step S7: Model prediction and recommendation result sorting. The computer device can use the preference value of the object to be predicted for the item it holds, the preference value of all objects holding the item to be predicted for the item to be predicted, and the social object relationship chain of the object to be predicted (i.e., the above social relationship sequence) as input to the deep learning model, and use the deep learning model to predict the predicted preference value of the object to be predicted for the item to be predicted. The application process of the deep learning model can be seen in the above Figure 5 Description in the corresponding embodiment.

[0244] Step S8, item recommendation results. The computer device can sort all items to be predicted for the prediction target based on the predicted preference values ​​output by the deep learning model, determine the recommended items corresponding to the prediction target based on the first y (e.g., y is 5) items after sorting, and recommend the first y items to the prediction target.

[0245] Please also see Figures 7a to 7d , Figures 7a to 7d This is a schematic diagram of the network structure of a recommendation model provided by an embodiment of the present application. Figure 7d As shown, the recommendation model can include four parts: item aggregation module, object aggregation module, social information module and interest / relationship chain migration module.

[0246] like Figure 7a As shown, the item aggregation module may include a first neural network layer g v (multi-layer fully connected neural network), a two-layer attention network 50d (i.e., the first attention network layer), a first fully connected layer, and a first activation layer. Figure 7a As shown, the items held by the object to be predicted may include item 1, item 2, item 3, ...; the preference value of the object to be predicted for item 1 is r1, the preference value of the object to be predicted for item 2 is r2, the preference value of the object to be predicted for item 3 is r3, .... The input of the item aggregation module is the historical item features of the items held by the object to be predicted (such as the item features corresponding to item 1, the item features corresponding to item 2, the item features corresponding to item 3, etc.) and the preference value of the object to be predicted for the items held (such as preference value r1, preference value r2, preference value r3, etc.). The historical item features and preference values ​​are input into the item aggregation module and first input into the first neural network layer g v , through the first neural network g v The output features of the items can be output (such as feature 50a determined by item 1 and preference value r1, feature 50b determined by item 2 and preference value r2, feature 50c determined by item 3 and preference value r3, etc.); the item similarity between the item held by the object to be predicted and the item to be predicted, the object feature to be predicted corresponding to the object to be predicted, the item feature to be predicted corresponding to the item to be predicted, and the first neural network g v The output item output features are input to the double-layer attention network 50d, and the item weights corresponding to the items held by the object to be predicted are output through the double-layer attention network 50d, such as the item weight corresponding to item 1 is α1, the item weight corresponding to item 2 is α2, the item weight corresponding to item 3 is α3, and so on; the item output features and the item weights are weighted and summed, and the item information fusion features 50e corresponding to the items held by the object to be predicted are output through the first fully connected layer and the first activation layer (such as the above formula (3)).

[0247] like Figure 7b As shown, the social information module may include a third neural network layer g s (multi-layer fully connected neural network), the first recurrent neural network (bidirectional LSTM), a two-layer attention network 50i (i.e., the third attention network layer), a third fully connected layer, and a third activation layer. Figure 7bAs shown, a social relationship sequence corresponding to the object to be predicted can be constructed through random walks. Each social relationship element in the social relationship sequence can include: the associated object of the object to be predicted, the similar items held by the associated object that are most relevant to the object to be predicted, and the associated object's preference value for the similar items. The input of the social information module is the associated object features corresponding to the associated object, the similar item features of the similar items, and the associated object's preference value for the similar items. The associated object features, similar item features, and preference values ​​are input into the social information module and first input into the third neural network layer g s , through the third neural network layer g s The social output features (such as feature 50f, feature 50g, feature 50h, etc.) can be output; the social output features corresponding to each social relationship element in the social relationship sequence are input into the first recurrent neural network (i.e., bidirectional LSTM, which can be forward encoded and reverse encoded, such as the above formula (10)), and the implicit state features corresponding to each social relationship element in the social relationship sequence are output through the bidirectional LSTM; the social similarity between the associated object and the object to be predicted (equivalent to the object similarity between different objects), the social intimacy between the associated object and the object to be predicted, the object features corresponding to the object to be predicted, the item features corresponding to the item to be predicted, and the implicit state features output by the bidirectional LSTM are input into the double-layer attention network 50i, and the double-layer attention network is used to generate the corresponding implicit state features. 50i outputs the social weight corresponding to each associated object in the social relationship sequence, such as the social weight corresponding to associated object 1 in the social relationship sequence is β1, the social weight corresponding to associated object 2 in the social relationship sequence is β2, the social weight corresponding to associated object 3 in the social relationship sequence is β3, and so on; the implicit state feature and the social weight are weighted and summed, and the social output feature corresponding to the social relationship sequence (such as the social output feature 50j) is output through the third fully connected layer and the third activation layer (such as the above formula (11)); when the number of social relationship sequences constructed by random walk is P, the social output features corresponding to P (here P is defaulted to be greater than 1) social relationship sequences can be obtained in the social information module. At this time, the social information module can also include a multi-layer fully connected neural network g′ s , through a multi-layer fully connected neural network g′ s Output the social information features 50k corresponding to the P social output features. It can be understood that when the number of social relationship sequences is one, the social output features corresponding to the social relationship sequence can be determined as the social information features corresponding to the object to be predicted.

[0248] like Figure 7c As shown, the object aggregation module may include a second neural network layer g u(multi-layer fully connected neural network), a two-layer attention network 50q (i.e., the second attention network layer), a second fully connected layer, and a second activation layer. Figure 7c As shown, the objects holding the items to be predicted may include object 1, object 2, object 3, ...; the preference value of object 1 for the item to be predicted is r4, the preference value of object 2 for the item to be predicted is r5, the preference value of object 3 for the item to be predicted is r6, .... The input of the object aggregation module is the historical object features corresponding to the objects holding the items to be predicted (such as the object features corresponding to object 1, the object features corresponding to object 2, the object features corresponding to object 3, etc.) and the preference values ​​of the objects holding the items to be predicted for the items to be predicted (such as preference value r4, preference value r5, preference value r6, etc.). The historical object features and preference values ​​are input into the object aggregation module and first input into the second neural network layer g u , through the second neural network layer g u The object output features (such as feature 50m determined by object 1 and preference value r4, feature 50n determined by object 2 and preference value r5, feature 50p determined by object 3 and preference value r6, etc.) can be output; the object similarity between the object holding the item to be predicted and the object to be predicted, the object feature to be predicted corresponding to the object to be predicted, the item feature to be predicted corresponding to the item to be predicted, and the second neural network layer g u The output object output feature is input to the double-layer attention network 50q, and the object weight corresponding to the object holding the item to be predicted is output through the double-layer attention network 50q, such as the object weight corresponding to object 1 is γ1, the object weight corresponding to object 2 is γ2, the object weight corresponding to object 3 is γ3, and so on; the object output feature and the object weight are weightedly summed, and the object information fusion feature 50r corresponding to the object holding the item to be predicted is output through the second fully connected layer and the second activation layer (such as the above formula (7)).

[0249] like Figure 7d As shown, the interest / relationship chain migration module may include a fourth neural network layer g t (multi-layer fully connected neural network), the second recurrent neural network (unidirectional LSTM), a two-layer attention network 50u (i.e., the fourth attention network layer), a fourth fully connected layer, a fourth activation layer, and a multi-layer neural network g o .like Figure 7dAs shown, the item information fusion feature 50e output by the item aggregation module, the object information fusion feature 50r output by the object aggregation module, and the social information feature 50k output by the social information module are cascaded to obtain the cascade feature 50s, that is, the input of the interest / relationship chain transfer module is the cascade feature 50s, the object feature to be predicted corresponding to the object to be predicted, and the item feature to be predicted corresponding to the item to be predicted; the cascade feature 50s, the object feature to be predicted, and the item feature to be predicted are cascaded and input into the interest / relationship chain transfer module, and first input into the fourth neural network layer g t , through the fourth neural network layer g t The network output feature 50t is obtained. Assuming that the item information fusion feature 50e is the output result of the item clustering module at time t (t = 1, 2, ..., N), the object information fusion feature 50r is the output result of the object clustering module at time t, and the social information feature 50k is the output result of the social information module at time t, the network output feature 50t is the fourth neural network layer g t The output result at time t, the network output feature 50x is the fourth neural network layer g t The output result at time t-1, the network output feature 50w is the fourth neural network layer g t Output result at time N; input the network output feature into the second recurrent neural network (i.e., unidirectional LSTM, as shown in the above formula (15)), output the network implicit feature through the second recurrent neural network, input the network implicit feature output by the second recurrent neural network into the double-layer attention network 50u, and output the hidden layer weight in the second recurrent neural network through the double-layer attention network 50u, such as the hidden layer weight in the unidirectional LSTM The corresponding hidden layer weight is θ2, and the hidden layer The corresponding hidden layer weight is θ1, and the hidden layer The corresponding hidden layer weight is θ3, etc. The hidden layer weight is weighted and summed with the network implicit features, and the temporal migration feature 50v between the N network implicit features is output through the fourth fully connected layer and the fourth activation layer; the temporal migration feature 50v can then be input into the multi-layer neural network g o , through the multi-layer neural network g o Output the predicted preference value of the object to be predicted for the item to be predicted (as shown in the above formula (16)).

[0250] Please also see Figure 8 , Figure 8 This is a schematic diagram of a game prop recommendation scenario provided by an embodiment of the present application. For objects and game props (items) in the game client, the above Figures 7a to 7d The recommendation model shown in , predicts the object’s predicted preference value for the game props that it does not own. Figure 8As shown, the predicted preference values ​​output by the recommendation model are added to the preference value set 60a. Assuming that object 1, object 2 and object 3 do not own props 1 to props 10, the predicted preference values ​​of object 1 for props 1 to props 10 output by the recommendation model are: 0.3, 0.2, 0.2, 0.6, 0.3, 0.8, 1.0, 0.6, 0.1, 0.9; the predicted preference values ​​of object 2 for props 1 to props 10 are: 0.7, 0.5, 0.9, 0.3, 0.1, 0.2, 0.4, 0.9, 1.0, 0.5; the predicted preference values ​​of object 3 for props 1 to props 10 are: 0.9, 0.4, 0.6, 0.8, 1.0, 0.1, 0.7, 0.3, 0.5, 0.1. Through sort(·), the predicted preference values ​​output by the recommendation model can be sorted from large to small by object grouping to obtain a sorted preference value set 60b. In the preference value set 60b, after sorting from large to small according to the predicted preference values, it can be determined that the prop sorting order for object 1 is: prop 7, prop 10, prop 6, prop 4, prop 8, prop 1, prop 5, prop 2, prop 3, prop 9; the prop sorting order for object 2 is determined to be: prop 9, prop 3, prop 8, prop 1, prop 2, prop 10, prop 7, prop 4, prop 6, prop 5; the prop sorting order for object 3 is determined to be: prop 5, prop 1, prop 4, prop 7, prop 3, prop 9, prop 2, prop 8, prop 6, prop 10. If the number of items recommended for the object is 3, then according to the preference value set 60b, it can be determined that the recommended props for object 1 are: prop 7, prop 10, prop 6, the recommended props for object 2 are: prop 9, prop 3, prop 8, and the recommended props for object 3 are: prop 5, prop 1, prop 4; then, props 7, prop 10, prop 6 can be pushed to the user terminal 60c used by object 1, and props 7, prop 10, prop 6 can be displayed in the game client installed on the user terminal 60c. Props 9, prop 3, prop 8 can be pushed to the user terminal 60d used by object 2, and props 7, prop 10, prop 6 can be displayed in the game client installed on the user terminal 60d. Props 5, prop 1, prop 4 can be pushed to the user terminal 60e used by object 3, and props 5, prop 1, prop 4 can be displayed in the game client installed on the user terminal 60e.

[0251] Optionally, the recommendation model in the embodiment of the present application can also be a tree model, support vector machine model or other models. In scenarios with limited computing power or rapid verification, a tree model or a support vector machine model can be used to output the predicted preference value of the object to be predicted for the item to be predicted, which can reduce the computational complexity and thereby improve the efficiency of item recommendation.

[0252] In an embodiment of the present application, the items held by the object to be predicted at time t can be subjected to item information fusion, that is, analysis is performed at the item level, and an attention network with item similarity as prior information is introduced to obtain item information fusion features; object information fusion is performed on the object holding the item to be predicted at time t, that is, analysis is performed at the object level, and an attention network with object similarity as prior information is introduced to obtain object information fusion features; social information is obtained based on the socially associated objects associated with the object to be predicted at time t, that is, analysis is performed at the social object relationship chain level, and an attention network with object similarity and object intimacy as prior information and random walks are introduced to obtain social information features. Then, the item information fusion features, object information fusion features, social information features, object to be predicted features, and item to be predicted features obtained at N times (including time t) can be cascaded to obtain the cascaded features to be predicted. The temporal migration features between the N cascaded features to be predicted are output through LSTM, and the predicted preference value of the object to be predicted for the item to be predicted is output. In other words, in the process of item recommendation, this application introduces the influence of indirect social related objects (indirect friends) on the preference of the predicted object through random walk, obtains the influence of different social related objects on the preference of the predicted object through the attention network, and learns the interests and social relationship chain migration of the object to be predicted through LSTM, which can enhance the matching degree between the object to be predicted and the recommended item, and thus improve the accuracy of item recommendation.

[0253] It is understandable that in the specific implementation of this application, user-related information may be involved (for example, user gender information, user object identification and object portrait in the application client, user friends in the application client, items held by the user, and user operation data on items, etc.). When the above embodiments of this application are applied to specific products or technologies, the user's permission or consent must be obtained before the user's relevant information can be obtained, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0254] See Figure 9 , Figure 9 This is a structural diagram of a data recommendation device provided in an embodiment of the present application. Figure 9 As shown, the data recommendation device can be a computer program (including program code) applied to a computer device, for example, the data recommendation device is an application software, and the data recommendation device can be used to perform the corresponding steps of the method provided in the embodiment of the present application. The data recommendation device 1 may include: a fusion feature acquisition module 11, a social information acquisition module 12, a feature cascade module 13, and an item recommendation module 14;

[0255] The fusion feature acquisition module 11 is used to obtain the item information fusion features corresponding to the historical item set held by the object to be predicted at time t, and obtain the object information fusion features corresponding to the historical object set holding the item to be predicted at time t;

[0256] A social information acquisition module 12 is configured to construct a social relationship sequence corresponding to the object to be predicted based on the socially associated objects associated with the object to be predicted at time t, and determine the social information features corresponding to the object to be predicted based on the social relationship sequence;

[0257] The feature cascade module 13 is used to obtain the object features to be predicted corresponding to the object to be predicted and the item features to be predicted corresponding to the item to be predicted, and combine the item information fusion feature, the object information fusion feature, the social information feature, the object features to be predicted, and the item features to be predicted into the cascade features to be predicted corresponding to time t;

[0258] The item recommendation module 14 is used to determine the temporal migration characteristics between the object to be predicted and the item to be predicted based on the cascade feature sequence, and determine the recommendation status of the item to be predicted for the object to be predicted based on the temporal migration characteristics; the cascade feature sequence includes the cascade features to be predicted corresponding to N moments, where N moments include moment t, and N is a positive integer greater than 1.

[0259] The specific functional implementation of the fusion feature acquisition module 11, the social information acquisition module 12, the feature cascade module 13, and the item recommendation module 14 can be found in the above Figure 3 Steps S101 to S104 in the corresponding embodiment will not be described in detail here.

[0260] In some feasible implementations, the fusion feature acquisition module 11 may include: a first acquisition unit 111 , a second acquisition unit 112 , a first output unit 113 , and an item information fusion unit 114 ;

[0261] A first acquisition unit 111 is configured to acquire a set of historical items held by the subject to be predicted at time t, and to acquire a first evaluation triplet between the subject to be predicted and the items contained in the set of historical items; the first evaluation triplet includes the subject to be predicted, item k in the set of historical items, and the subject to be predicted's preference value for item k, where k is a positive integer;

[0262] The second acquisition unit 112 is configured to acquire historical item features corresponding to item k and convert the preference value contained in the first evaluation triplet into a first preference vector associated with item k;

[0263] A first output unit 113 is configured to input a first cascade feature between the historical item feature and the first preference vector into a first neural network layer in the recommendation model, and output an item output feature corresponding to the first cascade feature through the first neural network layer;

[0264] The item information fusion unit 114 is configured to obtain an item weight corresponding to each item in the historical item set, and determine an item information fusion feature corresponding to the historical item set based on the item weight and the item output feature associated with each item in the historical item set.

[0265] The specific functional implementation of the first acquisition unit 111, the second acquisition unit 112, the first output unit 113, and the item information fusion unit 114 can be found in the above Figure 5 Step S203 to step S203 in the corresponding embodiment will not be described in detail here.

[0266] In some feasible implementations, the first acquisition unit 111 may include: an identification acquisition subunit 1111, an indicator acquisition subunit 1112, a preference value determination subunit 1113, and a triplet determination subunit 1114;

[0267] The identification acquisition subunit 1111 is used to obtain the object identification corresponding to the object to be predicted and obtain the item identification corresponding to item k in the historical item set;

[0268] The indicator acquisition subunit 1112 is used to obtain the usage time interval and usage frequency of the object to be predicted for the item k, and determine the product of the usage time interval and the usage frequency as the item usage attenuation factor corresponding to the item k;

[0269] The preference value determination subunit 1113 is configured to determine a balance factor corresponding to item k based on the number of unit usages of item k by the subject to be predicted, and to determine the preference value of the subject to be predicted for item k based on the usage time interval, usage frequency, item usage attenuation factor, and balance factor;

[0270] The triplet determination subunit 1114 is configured to determine the object identifier, the item identifier, and the preference value of the object to be predicted for the item k as a first evaluation triplet between the object to be predicted and the item k.

[0271] The specific functional implementation of the identification acquisition subunit 1111, the indicator acquisition subunit 1112, the preference value determination subunit 1113, and the triple determination subunit 1114 can be found in the above Figure 5 Step S201 in the corresponding embodiment will not be described in detail here.

[0272] In some feasible implementations, the second acquisition unit 112 may include: a feature conversion subunit 1121, a feature preprocessing subunit 1122, and a first item feature determination subunit 1123;

[0273] The feature conversion subunit 1121 is used to obtain attribute description information associated with item k and convert the attribute description information into an initial attribute feature vector corresponding to item k;

[0274] The feature preprocessing subunit 1122 is configured to obtain the minimum eigenvalue and the maximum eigenvalue in the initial attribute feature vector, obtain a first difference between the maximum eigenvalue and the minimum eigenvalue, and obtain a second difference between the target eigenvalue and the minimum eigenvalue; the target eigenvalue is any eigenvalue in the initial attribute feature vector;

[0275] The first item feature determination subunit 1123 is configured to determine the ratio between the second difference and the first difference as a normalized target feature value, and determine the historical item feature corresponding to the item k according to the normalized target feature value.

[0276] The specific functional implementation of the feature conversion subunit 1121, the feature preprocessing subunit 1122, and the first item feature determination subunit 1123 can be found in the above Figure 3 Step S103 in the corresponding embodiment will not be described in detail here.

[0277] In some feasible implementations, the item information fusion unit 114 may include: a first object feature determination subunit 1141, an item similarity acquisition subunit 1142, a first priori information input subunit 1143, and an item weight acquisition subunit 1144;

[0278] The first object feature determination subunit 1141 is configured to obtain item operation features corresponding to the object to be predicted from the object log, and determine the object features corresponding to the object to be predicted based on the item operation features and the basic object profile of the object to be predicted;

[0279] The item similarity acquisition subunit 1142 is configured to acquire the item similarity between the item to be predicted and item k based on the item features corresponding to the item to be predicted and the historical item features corresponding to item k;

[0280] The first prior information input subunit 1143 is used to input the characteristics of the object to be predicted, the characteristics of the item to be predicted, and the item similarity into the first attention network layer in the recommendation model;

[0281] The item weight acquisition subunit 1144 is used to perform a product operation on the feature matrix determined by the features of the object to be predicted, the features of the item to be predicted, and the item similarity according to the network parameter matrix of the first attention network layer, to obtain the item weight corresponding to each item in the historical item set.

[0282] The specific functional implementation of the first object feature determination subunit 1141, the item similarity acquisition subunit 1142, the first prior information input subunit 1143, and the item weight acquisition subunit 1144 can be found in the above Figure 5 Step S203 in the corresponding embodiment will not be described in detail here.

[0283] In some feasible implementations, the fusion feature acquisition module 11 may include: a third acquisition unit 115 , a fourth acquisition unit 116 , a second output unit 117 , and an object information fusion unit 118 ;

[0284] The third acquisition unit 115 is configured to acquire a set of historical objects that hold the item to be predicted at time t, and acquire a second evaluation triplet between each object in the set of historical objects and the item to be predicted; the second evaluation triplet includes object i in the set of historical objects, the item to be predicted, and the preference value of object i for the item to be predicted, where i is a positive integer;

[0285] The fourth obtaining unit 116 is configured to obtain the historical object features corresponding to the object i and convert the preference value included in the second evaluation triplet into a second preference vector corresponding to the object i;

[0286] A second output unit 117 is configured to input a second cascade feature between the historical object feature and the second preference vector into a second neural network layer in the recommendation model, and output an object output feature corresponding to the second cascade feature through the second neural network layer;

[0287] The object information fusion unit 118 is configured to obtain an object weight corresponding to each object in the historical object set, and determine an object information fusion feature corresponding to the historical object set based on the object weight and the object output feature associated with each object in the historical object set.

[0288] The specific functional implementation of the third acquisition unit 115, the fourth acquisition unit 116, the second output unit 117, and the object information fusion unit 118 can be found in the above Figure 5 Steps S204 to S206 in the corresponding embodiment will not be described in detail here.

[0289] In some feasible implementations, the object information fusion unit 118 may include: an object similarity acquisition subunit 1181 , a second priori information input subunit 1182 , and an object weight acquisition subunit 1183 ;

[0290] The object similarity acquisition subunit 1181 is used to acquire the object similarity between the object to be predicted and object i based on the object features corresponding to the object to be predicted and the historical object features corresponding to object i;

[0291] The second prior information input subunit 1182 is used to input the features of the object to be predicted, the features of the item to be predicted corresponding to the item to be predicted, and the object similarity into the second attention network layer in the recommendation model;

[0292] The object weight acquisition subunit 1183 is used to perform a product operation on the feature matrix determined by the features of the object to be predicted, the features of the item to be predicted, and the item similarity according to the network parameter matrix of the second attention network layer, to obtain the object weight corresponding to each object in the historical object set.

[0293] The specific functional implementation of the object similarity acquisition subunit 1181, the second prior information input subunit 1182, and the object weight acquisition subunit 1183 can be found in the above Figure 5 Step S206 in the corresponding embodiment will not be described in detail here.

[0294] In some feasible implementations, the social information acquisition module 12 may include: a random walk unit 121, a similar item determination unit 122, a social relationship sequence determination unit 123, a similar item feature acquisition unit 124, a third output unit 125, a bidirectional cyclic encoding unit 126, and a social information fusion unit 127;

[0295] The random walk unit 121 is configured to obtain the socially associated objects associated with the object to be predicted at time t, generate an object relationship chain including the object to be predicted and the socially associated objects, and perform a random walk in the object relationship chain starting from the object to be predicted to obtain a sequence of associated objects.

[0296] Similar item determination unit 122, used to obtain the associated object u in the associated object sequence [m] , in the associated object u [m] Among the items held, the item with the greatest item similarity to the item to be predicted is determined as the similar item corresponding to the item to be predicted;

[0297] The social relationship sequence determination unit 123 is used to associate the object u [m] , similar items and related objects u [m]The preference values ​​for similar items are determined as social sequence elements, and the social sequence elements associated with each associated object in the associated object sequence are determined as the social relationship sequence corresponding to the object to be predicted.

[0298] Similar item feature acquisition unit 124, used to acquire the associated object u [m] The corresponding associated object features, and the similar item features corresponding to the similar items, convert the preference values ​​in the social relationship sequence into a third preference vector;

[0299] A third output unit 125 is configured to input a third cascade feature among the associated object feature, the similar item feature, and the third preference vector into a third neural network layer in the recommendation model, and output a sequence output feature corresponding to the third cascade feature through the third neural network layer;

[0300] a bidirectional cyclic encoding unit 126 for inputting the sequence output features into a first recurrent neural network in the recommendation model, and performing bidirectional cyclic encoding on the sequence output features according to the first recurrent neural network to obtain forward encoding features and reverse encoding features;

[0301] The social information fusion unit 127 is used to splice the forward coding features and the reverse coding features into implicit state features, obtain the social weight corresponding to each associated object in the social relationship sequence, and determine the social information features corresponding to the object to be predicted based on the social weight and the implicit state features.

[0302] The specific functional implementation of the random walk unit 121, the similar item determination unit 122, the social relationship sequence determination unit 123, the similar item feature acquisition unit 124, the third output unit 125, the bidirectional cyclic encoding unit 126, and the social information fusion unit 127 can be found in the above Figure 3 Step S102 in the corresponding embodiment will not be described in detail here.

[0303] In some feasible implementations, the social information fusion unit 127 may include: a social similarity acquisition subunit 1271 , an intimacy determination subunit 1272 , a third priori information input subunit 1273 , and a social weight acquisition subunit 1274 ;

[0304] The social similarity acquisition subunit 1271 is configured to acquire the social similarity between the object to be predicted and each associated object based on the associated object features corresponding to each associated object in the social relationship sequence and the predicted object features corresponding to the object to be predicted;

[0305] The intimacy determination subunit 1272 is configured to determine the social intimacy between the object to be predicted and each associated object based on the social parameters between the object to be predicted and each associated object; the social parameters include social duration, number of common social objects, interaction duration, and interaction frequency;

[0306] The third prior information input sub-unit 1273 is used to input the characteristics of the object to be predicted, the social similarity, the social intimacy, and the characteristics of the item to be predicted corresponding to the item to be predicted into the third attention network layer in the recommendation model;

[0307] The social weight acquisition subunit 1274 is used to perform a product operation on the feature matrix determined by the features of the object to be predicted, social similarity, social intimacy and the features of the item to be predicted according to the network parameter matrix of the third attention network layer to obtain the social weight corresponding to each associated object.

[0308] The specific functional implementation of the social similarity acquisition subunit 1271, the intimacy determination subunit 1272, the third prior information input subunit 1273, and the social weight acquisition subunit 1274 can be found in the above Figure 5 Step S210 in the corresponding embodiment will not be described in detail here.

[0309] In some feasible implementations, the number of social relationship sequences is P, and the P social relationship sequences include social relationship sequences S ρ , P is a positive integer, ρ is a positive integer less than P;

[0310] The social information fusion unit 127 may include: a social output feature determination subunit 1275 , a social output feature fusion subunit 1276 ;

[0311] The social output feature determination subunit 1275 is used to determine the social relationship sequence S ρ The social weight corresponding to the associated object in the social relationship sequence S ρ The implicit state features of the association determine the social relationship sequence S ρ Corresponding social output features;

[0312] The social output feature fusion subunit 1276 is used to fuse the social output features corresponding to the P social relationship sequences to obtain the social information features corresponding to the object to be predicted.

[0313] The specific functional implementation of the social output feature determination subunit 1275 and the social output feature fusion subunit 1276 can be found in the above Figure 5 Step S210 in the corresponding embodiment will not be described in detail here.

[0314] In some feasible implementations, the item recommendation module 14 may include: a fourth output unit 141, a unidirectional cyclic encoding unit 142, a temporal transition feature acquisition unit 143, and a recommendation status determination unit 144;

[0315] The fourth output unit 141 is configured to combine the cascade features to be predicted corresponding to the N moments into a cascade feature sequence, sequentially input the cascade features to be predicted contained in the cascade feature sequence into the fourth neural network layer of the recommendation model, and output network output features corresponding to the N cascade features to be predicted through the fourth neural network layer;

[0316] A unidirectional cyclic encoding unit 142 is used to sequentially input the N network output features into the second cyclic neural network in the recommendation model, and perform forward cyclic encoding on the network output features according to the second cyclic neural network to obtain network implicit features;

[0317] The time series transition feature acquisition unit 143 is used to acquire the time series transition feature corresponding to the cascade feature sequence according to the network implicit feature and the hidden layer weight corresponding to the second recurrent neural network;

[0318] The recommendation status determination unit 144 is configured to obtain a predicted preference value of the object to be predicted for the item to be predicted based on the temporal transition feature, and determine a recommendation status of the item to be predicted for the object to be predicted based on the predicted preference value.

[0319] The specific functional implementation of the fourth output unit 141, the unidirectional cyclic encoding unit 142, the temporal transition feature acquisition unit 143, and the recommended state determination unit 144 can be found in the above Figure 5 Steps S212 to S214 in the corresponding embodiment will not be described in detail here.

[0320] In some feasible implementations, the number of items to be predicted is L, where L is a positive integer;

[0321] The recommendation status determination unit 144 may include: a sorting subunit 1441, a recommended item selection subunit 1442, and an item push subunit 1443;

[0322] The sorting subunit 1441 is used to obtain the predicted preference values ​​of the predicted object for L items to be predicted, and sort the predicted preference values ​​associated with the L items to be predicted in descending order to obtain the sorted L items to be predicted;

[0323] The recommended item selection subunit 1442 is configured to determine the recommendation status corresponding to the first y items to be predicted among the sorted L items to be predicted as a recommended success status; y is a positive integer less than or equal to L;

[0324] The item push subunit 1443 is configured to push y items to be recommended that are in a successful recommendation state to the object to be predicted.

[0325] The specific functional implementation of the sorting subunit 1441, the recommended item selection subunit 1442, and the item push subunit 1443 can be found in the above Figure 3 Step S104 in the corresponding embodiment will not be described in detail here.

[0326] In an embodiment of the present application, the items held by the object to be predicted at time t can be subjected to item information fusion, that is, analysis is performed at the item level, and an attention network with item similarity as prior information is introduced to obtain item information fusion features; object information fusion is performed on the object holding the item to be predicted at time t, that is, analysis is performed at the object level, and an attention network with object similarity as prior information is introduced to obtain object information fusion features; social information is obtained based on the socially associated objects associated with the object to be predicted at time t, that is, analysis is performed at the social object relationship chain level, and an attention network with object similarity and object intimacy as prior information and random walks are introduced to obtain social information features. Then, the item information fusion features, object information fusion features, social information features, object to be predicted features, and item to be predicted features obtained at N times (including time t) can be cascaded to obtain the cascaded features to be predicted. The temporal migration features between the N cascaded features to be predicted are output through LSTM, and the predicted preference value of the object to be predicted for the item to be predicted is output. In other words, in the process of item recommendation, this application introduces the influence of indirect social related objects (indirect friends) on the preference of the predicted object through random walk, obtains the influence of different social related objects on the preference of the predicted object through the attention network, and learns the interests and social relationship chain migration of the object to be predicted through LSTM, which can enhance the matching degree between the object to be predicted and the recommended item, and thus improve the accuracy of item recommendation.

[0327] See Figure 10 , Figure 10 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. Figure 10As shown, the computer device 1000 may include: a processor 1001, a network interface 1004 and a memory 1005. In addition, the above-mentioned computer device 1000 may also include: a user interface 1003, and at least one communication bus 1002. The communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. Optionally, the network interface 1004 may include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a non-volatile memory (non-volatile memory), such as at least one disk storage. Optionally, the memory 1005 may also be at least one storage device located away from the aforementioned processor 1001. As Figure 10 As shown, the memory 1005 as a computer-readable storage medium may include an operating system, a network communication module, a user interface module, and a device control application.

[0328] In such Figure 10 In the computer device 1000 shown, the network interface 1004 can provide network communication functions; the user interface 1003 is mainly used to provide an interface for user input; and the processor 1001 can be used to call the device control application stored in the memory 1005 to achieve:

[0329] Obtain the item information fusion features corresponding to the historical item set held by the object to be predicted at time t, and obtain the object information fusion features corresponding to the historical object set holding the item to be predicted at time t;

[0330] According to the social associated objects associated with the object to be predicted at time t, a social relationship sequence corresponding to the object to be predicted is constructed, and the social information features corresponding to the object to be predicted are determined according to the social relationship sequence;

[0331] Obtain the object features to be predicted corresponding to the object to be predicted and the item features to be predicted corresponding to the item to be predicted, and combine the item information fusion features, the object information fusion features, the social information features, the object features to be predicted, and the item features to be predicted into the cascade features to be predicted corresponding to time t;

[0332] The temporal migration characteristics between the object to be predicted and the item to be predicted are determined based on the cascade feature sequence, and the recommendation status of the item to be predicted for the object to be predicted is determined based on the temporal migration characteristics; the cascade feature sequence includes the cascade features to be predicted corresponding to N moments, where N moments include moment t, and N is a positive integer greater than 1.

[0333] It should be understood that the computer device 1000 described in the embodiment of the present application can execute the above Figure 3 and Figure 4 The description of the data recommendation method in any corresponding embodiment can also be performed as described above. Figure 9 The description of the data recommendation device 1 in the corresponding embodiment will not be repeated here. In addition, the description of the beneficial effects of adopting the same method will not be repeated here either.

[0334] In addition, it should be noted that: the embodiment of the present application also provides a computer-readable storage medium, and the computer-readable storage medium stores a computer program executed by the data recommendation device 1 mentioned above, and the computer program includes program instructions. When the processor executes the program instructions, it can execute the above-mentioned Figure 3 and Figure 4 The description of the data recommendation method in any corresponding embodiment will not be repeated here. In addition, the description of the beneficial effects of adopting the same method will not be repeated. For technical details not disclosed in the computer-readable storage medium embodiment involved in this application, please refer to the description of the method embodiment of this application. As an example, program instructions can be deployed on a computing device for execution, or on multiple computing devices located at one location, or on multiple computing devices distributed at multiple locations and interconnected by a communication network. Multiple computing devices distributed at multiple locations and interconnected by a communication network can constitute a blockchain system.

[0335] In addition, it should be noted that: the embodiment of the present application also provides a computer program product or computer program, which may include computer instructions, which may be stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor may execute the computer instructions, so that the computer device performs the above Figure 3 and Figure 4 The description of the data recommendation method in any corresponding embodiment will not be repeated here. In addition, the description of the beneficial effects of using the same method will not be repeated. For technical details not disclosed in the computer program product or computer program embodiment involved in this application, please refer to the description of the method embodiment of this application.

[0336] It should be noted that for the aforementioned various method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this application.

[0337] The steps in the method of the embodiment of the present application can be adjusted in order, combined and deleted according to actual needs.

[0338] The modules in the device of the embodiment of the present application can be merged, divided and deleted according to actual needs.

[0339] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0340] The above disclosure is only a preferred embodiment of the present application, and certainly cannot be used to limit the scope of rights of the present application. Therefore, equivalent changes made according to the claims of the present application are still within the scope covered by the present application.

Claims

1. A data recommendation method, characterized in that: include: Obtaining item information fusion features corresponding to the set of historical items held by the object to be predicted at time t, and obtaining object information fusion features corresponding to the set of historical objects holding the item to be predicted at said time t; Constructing a social relationship sequence corresponding to the object to be predicted based on the socially associated objects associated with the object to be predicted at time t, and determining the social information features corresponding to the object to be predicted based on the social relationship sequence; Obtaining a to-be-predicted object feature corresponding to the to-be-predicted object and a to-be-predicted item feature corresponding to the to-be-predicted item, and combining the item information fusion feature, the object information fusion feature, the social information feature, the to-be-predicted object feature, and the to-be-predicted item feature into a to-be-predicted cascade feature corresponding to the time t; Combining the cascade features to be predicted corresponding to N moments into a cascade feature sequence, and outputting the network output features corresponding to the N cascade features to be predicted in the cascade feature sequence through the fourth neural network layer in the recommendation model; the N moments include the moment t, and N is a positive integer greater than 1; According to the second recurrent neural network in the recommendation model, N network output features are forward cyclically encoded to obtain network implicit features; Obtaining a temporal transition feature corresponding to the cascade feature sequence according to the network implicit feature and the hidden layer weight corresponding to the second recurrent neural network; A predicted preference value of the object to be predicted for the item to be predicted is obtained according to the time series transition feature, and a recommendation status of the item to be predicted for the object to be predicted is determined according to the predicted preference value.

2. The method according to claim 1, characterized in that The step of obtaining the item information fusion features corresponding to the historical item set held by the object to be predicted at time t includes: Obtain a set of historical items held by the subject to be predicted at time t, and obtain a first evaluation triplet between the subject to be predicted and the items contained in the historical item set; the first evaluation triplet includes the subject to be predicted, item k in the historical item set, and the subject to be predicted's preference value for item k, where k is a positive integer; Obtaining historical item features corresponding to the item k, and converting the preference value included in the first evaluation triplet into a first preference vector associated with the item k; Inputting a first cascade feature between the historical item feature and the first preference vector into a first neural network layer in a recommendation model, and outputting an item output feature corresponding to the first cascade feature through the first neural network layer; An item weight corresponding to each item in the historical item set is obtained, and an item information fusion feature corresponding to the historical item set is determined based on the item weight and an item output feature associated with each item in the historical item set.

3. The method according to claim 2, characterized in that The obtaining of a first evaluation triplet between the object to be predicted and the items included in the historical item set includes: Obtaining an object identifier corresponding to the object to be predicted, and obtaining an item identifier corresponding to item k in the historical item set; Obtaining a usage time interval and a usage frequency of the item k by the subject to be predicted, and determining the product of the usage time interval and the usage frequency as an item usage attenuation factor corresponding to the item k; Determine a balance factor corresponding to item k based on the unit usage count of item k by the subject to be predicted, and determine a preference value of the subject to be predicted for item k based on the usage time interval, the usage frequency, the item usage attenuation factor, and the balance factor; The object identifier, the item identifier, and the preference value of the object to be predicted for the item k are determined as a first evaluation triplet between the object to be predicted and the item k.

4. The method according to claim 2, characterized in that The obtaining of the historical item features corresponding to the item k includes: Acquire attribute description information associated with the item k, and convert the attribute description information into an initial attribute feature vector corresponding to the item k; Obtaining a minimum eigenvalue and a maximum eigenvalue in the initial attribute feature vector, obtaining a first difference between the maximum eigenvalue and the minimum eigenvalue, and a second difference between a target eigenvalue and the minimum eigenvalue; the target eigenvalue is any eigenvalue in the initial attribute feature vector; The ratio between the second difference and the first difference is determined as a normalized target feature value, and the historical item feature corresponding to the item k is determined according to the normalized target feature value.

5. The method according to claim 2, characterized in that The obtaining of the item weight corresponding to each item in the historical item set includes: Obtaining item operation characteristics corresponding to the object to be predicted from the object log, and determining the object characteristics corresponding to the object to be predicted based on the item operation characteristics and the object basic profile of the object to be predicted; Obtaining an item similarity between the item to be predicted and the item k based on the item feature to be predicted corresponding to the item to be predicted and the historical item feature corresponding to the item k; Inputting the features of the object to be predicted, the features of the item to be predicted, and the item similarity into the first attention network layer in the recommendation model; According to the network parameter matrix of the first attention network layer, a product operation is performed on a feature matrix determined by the features of the object to be predicted, the features of the item to be predicted, and the item similarity to obtain an item weight corresponding to each item in the historical item set.

6. The method according to claim 1, characterized in that The acquiring of the object information fusion features corresponding to the set of historical objects holding the item to be predicted at time t includes: Obtain a set of historical objects that hold the item to be predicted at time t, and obtain a second evaluation triplet between each object in the set of historical objects and the item to be predicted; the second evaluation triplet includes object i in the set of historical objects, the item to be predicted, and the preference value of object i for the item to be predicted, where i is a positive integer; Obtaining historical object features corresponding to the object i, and converting the preference value included in the second evaluation triplet into a second preference vector corresponding to the object i; Inputting a second cascade feature between the historical object feature and the second preference vector into a second neural network layer in the recommendation model, and outputting an object output feature corresponding to the second cascade feature through the second neural network layer; An object weight corresponding to each object in the historical object set is obtained, and an object information fusion feature corresponding to the historical object set is determined based on the object weight and the object output feature associated with each object in the historical object set.

7. The method according to claim 6, characterized in that The obtaining of the object weight corresponding to each object in the historical object set includes: Obtaining an object similarity between the object to be predicted and the object i according to the object features corresponding to the object to be predicted and the historical object features corresponding to the object i; Inputting the features of the object to be predicted, the features of the item to be predicted corresponding to the item to be predicted, and the object similarity into the second attention network layer in the recommendation model; According to the network parameter matrix of the second attention network layer, a product operation is performed on the feature matrix determined by the features of the object to be predicted, the features of the item to be predicted, and the item similarity to obtain the object weight corresponding to each object in the historical object set.

8. The method according to claim 1, characterized in that The step of constructing a social relationship sequence corresponding to the object to be predicted based on the socially associated objects associated with the object to be predicted at time t includes: Obtaining the socially associated objects associated with the object to be predicted at time t, generating an object relationship chain including the object to be predicted and the socially associated objects, and performing a random walk in the object relationship chain starting from the object to be predicted to obtain a sequence of associated objects; Get the associated object u in the associated object sequence [m] , in the associated object u [m] Among the items held, the item with the greatest item similarity to the item to be predicted is determined as the similar item corresponding to the item to be predicted; The associated object u [m] , the similar items and the associated object u [m] The preference value of the similar item is determined as a social sequence element, and the social sequence element associated with each associated object in the associated object sequence is determined as the social relationship sequence corresponding to the object to be predicted.

9. The method according to claim 8, characterized in that The determining, based on the social relationship sequence, the social information features corresponding to the object to be predicted includes: Get the associated object u [m] corresponding associated object features, and similar item features corresponding to the similar items, converting the preference values ​​in the social relationship sequence into a third preference vector; Inputting a third cascade feature between the associated object feature, the similar item feature, and the third preference vector into a third neural network layer in the recommendation model, and outputting a sequence output feature corresponding to the third cascade feature through the third neural network layer; Inputting the sequence output features into a first recurrent neural network in the recommendation model, and performing bidirectional recurrent encoding on the sequence output features according to the first recurrent neural network to obtain forward encoding features and reverse encoding features; The forward coding features and the reverse coding features are spliced ​​into implicit state features, the social weight corresponding to each associated object in the social relationship sequence is obtained, and the social information features corresponding to the object to be predicted are determined according to the social weight and the implicit state features.

10. The method according to claim 9, characterized in that The obtaining of the social weight corresponding to each associated object in the social relationship sequence includes: Obtaining the social similarity between the object to be predicted and each associated object according to the associated object features corresponding to each associated object in the social relationship sequence and the to-be-predicted object features corresponding to the to-be-predicted object; Determining the social intimacy between the object to be predicted and each of the associated objects based on the social parameters between the object to be predicted and each of the associated objects; the social parameters include social duration, number of common social objects, interaction duration, and interaction frequency; Inputting the features of the object to be predicted, the social similarity, the social intimacy, and the features of the item to be predicted corresponding to the item to be predicted into the third attention network layer in the recommendation model; According to the network parameter matrix of the third attention network layer, a product operation is performed on the feature matrix determined by the features of the object to be predicted, the social similarity, the social intimacy and the features of the item to be predicted to obtain the social weight corresponding to each associated object.

11. The method according to claim 9, characterized in that The number of the social relationship sequences is P, and the P social relationship sequences include a social relationship sequence S ρ , P is a positive integer, ρ is a positive integer less than P; The determining, based on the social weight and the implicit state feature, the social information feature corresponding to the object to be predicted includes: According to the social relationship sequence S ρ The social weight corresponding to the associated object in the social relationship sequence S ρ The associated implicit state features determine the social relationship sequence S ρ Corresponding social output features; The social output features corresponding to the P social relationship sequences are fused to obtain the social information features corresponding to the object to be predicted.

12. The method according to claim 1, characterized in that The number of items to be predicted is L, where L is a positive integer; The determining, according to the predicted preference value, the recommendation status of the item to be predicted for the object to be predicted includes: Obtaining predicted preference values ​​of the subject to be predicted for L items to be predicted, and sorting the predicted preference values ​​associated with the L items to be predicted in descending order to obtain sorted L items to be predicted; Among the L items to be predicted after the sorting, the recommendation status corresponding to the first y items to be predicted is determined as the recommendation success status; y is a positive integer less than or equal to L; Push the y recommended items that are in a successful recommendation state to the object to be predicted.

13. A data recommendation device, characterized in that: include: A fusion feature acquisition module is used to obtain the item information fusion features corresponding to the historical item set held by the object to be predicted at time t, and obtain the object information fusion features corresponding to the historical object set holding the item to be predicted at time t; A social information acquisition module, configured to construct a social relationship sequence corresponding to the object to be predicted based on the socially associated objects associated with the object to be predicted at time t, and determine the social information features corresponding to the object to be predicted based on the social relationship sequence; a feature cascade module, configured to obtain the object feature to be predicted corresponding to the object to be predicted and the item feature to be predicted corresponding to the item to be predicted, and combine the item information fusion feature, the object information fusion feature, the social information feature, the object feature to be predicted, and the item feature to be predicted into a cascade feature to be predicted corresponding to the time t; a fourth output unit, configured to combine the cascade features to be predicted corresponding to N moments into a cascade feature sequence, and output, through a fourth neural network layer in the recommendation model, network output features corresponding to the N cascade features to be predicted in the cascade feature sequence; the N moments include the moment t, and N is a positive integer greater than 1; a unidirectional cyclic encoding unit, configured to perform forward cyclic encoding on N network output features according to the second cyclic neural network in the recommendation model to obtain network implicit features; A time series transition feature acquisition unit, configured to acquire a time series transition feature corresponding to the cascade feature sequence according to the network implicit feature and the hidden layer weight corresponding to the second recurrent neural network; A recommendation status determination unit is configured to obtain a predicted preference value of the object to be predicted for the item to be predicted based on the temporal transition feature, and determine a recommendation status of the item to be predicted for the object to be predicted based on the predicted preference value.

14. A computer device, characterized in that: including memory and processor; The memory is connected to the processor, the memory is used to store a computer program, and the processor is used to call the computer program so that the computer device executes the method according to any one of claims 1 to 12.

15. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which is suitable for being loaded and executed by a processor, so that a computer device having the processor executes the method according to any one of claims 1 to 12.

16. A computer program product, characterized in that The computer program product comprises computer instructions, and when the computer instructions are executed by a processor, the method according to any one of claims 1 to 12 is implemented.

Images