WO2021159776A1 - 基于人工智能的推荐方法、装置、电子设备及存储介质 - Google Patents

基于人工智能的推荐方法、装置、电子设备及存储介质 Download PDF

Info

Publication number
WO2021159776A1
WO2021159776A1 PCT/CN2020/126549 CN2020126549W WO2021159776A1 WO 2021159776 A1 WO2021159776 A1 WO 2021159776A1 CN 2020126549 W CN2020126549 W CN 2020126549W WO 2021159776 A1 WO2021159776 A1 WO 2021159776A1
Authority
WO
WIPO (PCT)
Prior art keywords
recommendation information
candidate recommendation
feature
candidate
recommended
Prior art date
Application number
PCT/CN2020/126549
Other languages
English (en)
French (fr)
Inventor
李天浩
Original Assignee
腾讯科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 腾讯科技（深圳）有限公司 filed Critical 腾讯科技（深圳）有限公司
Publication of WO2021159776A1 publication Critical patent/WO2021159776A1/zh
Priority to US17/699,421 priority Critical patent/US20220215032A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G06F16/2457—Query processing with adaptation to user needs
    • G06F16/24578—Query processing with adaptation to user needs using ranking
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G06F16/2455—Query execution
    • G06F16/24553—Query execution of query operations
    • G06F16/24554—Unary operations; Data partitioning operations
    • G06F16/24556—Aggregation; Duplicate elimination
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • G06N5/043—Distributed expert systems; Blackboards
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks

Definitions

  • This application relates to artificial intelligence technology, and in particular to a recommendation method, device, electronic device, and computer-readable storage medium based on artificial intelligence.
  • AI Artificial Intelligence
  • the embodiments of the present application provide an artificial intelligence-based recommendation method, device, electronic device, and computer-readable storage medium, which can realize accurate recommendation of information.
  • the embodiment of the present application provides a recommendation method based on artificial intelligence, including:
  • Multi-level mapping processing is performed on the fusion feature corresponding to each candidate recommendation information, and each candidate recommendation information is scored corresponding to multiple indicators;
  • An embodiment of the present application provides an artificial intelligence-based recommendation device, including:
  • the information acquisition module is used to acquire multiple candidate recommendation information corresponding to the object to be recommended;
  • the feature forming module is used to obtain the object feature of the object to be recommended and the candidate recommendation information feature of each candidate recommendation information, and combine the object feature with the candidate recommendation information feature of each candidate recommendation information, respectively, to Form a fusion feature corresponding to each candidate recommendation information;
  • the feature processing module is used to perform multi-level mapping processing on the fusion feature corresponding to each candidate recommendation information to obtain scores corresponding to multiple indicators for each candidate recommendation information;
  • the information sorting module is used to perform multi-index aggregation processing on the scores corresponding to the multiple indicators to obtain a comprehensive score of each candidate recommendation information, and to compare the multiple candidate recommendation information according to the comprehensive score of each candidate recommendation information Sort in descending order;
  • the recommendation module is configured to select at least one candidate recommendation information ranked higher in the result of the descending order, and perform a recommendation operation corresponding to the object to be recommended based on the selected candidate recommendation information.
  • the embodiment of the application provides an artificial intelligence-based recommendation model training method, including:
  • the candidate recommendation information prediction model obtained by the training is used for the recommendation system to perform multi-index aggregation processing and sorting, so as to determine the candidate recommendation information to be recommended according to the sorting result.
  • the embodiment of the present invention provides an artificial intelligence-based recommendation model training device, including:
  • the training sample set building module is used to preprocess the log of the recommendation system to build a training sample set
  • a model construction module for constructing a candidate recommendation information prediction model based on a weight generator corresponding to multiple indicators one-to-one, predictors corresponding to the multiple indicators one-to-one, and multiple expert networks;
  • a training module configured to perform multi-index training on the candidate recommendation information prediction model through the training sample set
  • the candidate recommendation information prediction model obtained by the training is used for the recommendation system to perform multi-index aggregation processing and sorting, so as to determine the candidate recommendation information to be recommended according to the sorting result.
  • An embodiment of the application provides an electronic device, including:
  • Memory used to store executable instructions
  • the processor is configured to implement the artificial intelligence-based recommendation method and the artificial intelligence-based recommendation model training method provided in the embodiments of the present application when executing the executable instructions stored in the memory.
  • the embodiment of the application provides a computer-readable storage medium storing executable instructions for causing a processor to execute, to implement the artificial intelligence-based recommendation method and the artificial intelligence-based recommendation model training method provided by the embodiment of the application .
  • the information to be recommended is selected by sorting the comprehensive scores of multi-level mapping processing of fusion features and multi-index aggregation processing, which is more comprehensive and objective than single-target prediction, thus achieving accurate personalized recommendation and improving the performance of the recommendation system.
  • FIG. 1A is a schematic diagram of an optional architecture of an artificial intelligence-based recommendation system 100-A provided by an embodiment of the present application;
  • FIG. 1B is a schematic diagram of an optional architecture of the artificial intelligence-based recommendation system 100-B provided by an embodiment of the present application;
  • FIG. 2A is a schematic structural diagram of a training server 200-A applying an artificial intelligence-based recommendation model training method provided by an embodiment of the present application;
  • FIG. 2B is a schematic structural diagram of an application server 200-B applying an artificial intelligence-based recommendation method provided by an embodiment of the present application;
  • FIG. 3 is a model architecture diagram in the artificial intelligence-based recommendation model training method provided by an embodiment of the present application.
  • Figure 4 is a schematic diagram of a novel activation function provided by an embodiment of the present application.
  • 5A-5C are schematic diagrams of an optional process of the artificial intelligence-based recommendation model training method provided by an embodiment of the present application.
  • 6A-6C are schematic diagrams of an optional flow of the artificial intelligence-based recommendation method provided by an embodiment of the present application.
  • FIG. 7 is a schematic diagram of an application scenario product of an artificial intelligence-based recommendation method provided by an embodiment of the present application.
  • FIG. 8 is a diagram of the training and application architecture of the candidate recommendation information prediction model provided by an embodiment of the present application.
  • first ⁇ second involved is only to distinguish similar objects, and does not represent a specific order for the objects. Understandably, “first ⁇ second” can be used if permitted.
  • the specific order or sequence is exchanged, so that the embodiments of the present application described herein can be implemented in a sequence other than those illustrated or described herein.
  • Multi-task learning uses the useful information contained in multiple learning tasks to obtain a more accurate learner for each learning task.
  • AUC Area Under Curve
  • the object to be recommended is the target of information recommendation. Since the medium of information presentation is the terminal, the target of information recommendation is the user who operates the corresponding terminal. Therefore, "object” and “user” are equivalently described below. Understandably, the user here may be a natural person who can operate the terminal, or a robot program that can simulate human behavior running in the terminal.
  • Candidate recommendation information that is, information that can be sent to the terminal for presentation to make recommendations to users of the corresponding terminal.
  • each expert network is a forward propagation network, its output is characteristic, corresponding to different tasks, a weight generation structure is introduced for each task, and the weight value of the expert network is output to make different comprehensive tasks Use the output of the expert network in different ways.
  • the embodiment of the application proposes a recommendation method based on artificial intelligence, which respectively design a multi-layer expert network, introduce a multi-index uncertainty factor, solve the classification imbalance problem, and improve model activation.
  • the candidate recommendation information prediction model of the personalized recommendation system has been deeply optimized and transformed.
  • the AUC and root mean square error of the offline evaluation of the model have been significantly improved.
  • the ranking target of the online recommendation system is configurable. The design of, makes the sorting results more flexible and controllable, and ultimately can positively drive the core indicators of online recommendations, and the proportion of user interactions and explicit behaviors has also been significantly improved.
  • the embodiments of this application provide an artificial intelligence-based recommendation method, device, electronic device, and computer-readable storage medium, which can positively drive various core indicators of online recommendation, and the proportion of user interaction and explicit behavior has also been significantly improved.
  • the electronic equipment provided by the embodiments of the application can be implemented as a notebook computer, a tablet computer, a desktop computer, a set-top box, a mobile device (for example, a mobile phone, a portable music player, and a personal computer).
  • Various types of user terminals, such as digital assistants, dedicated messaging devices, portable game devices, and on-board computers can also be implemented as servers. In the following, an exemplary application when the device is implemented as a server will be described.
  • FIG. 1A is a schematic diagram of an alternative architecture of an artificial intelligence-based recommendation system 100-A provided by an embodiment of the present application.
  • the terminal 400 is connected to the application server 200-B and the training server 200-A through the network 300, and the network 300 It can be a wide area network or a local area network, or a combination of the two.
  • the training server 200-A is responsible for offline training of the candidate recommendation information prediction model.
  • the training server 200-A includes a training sample set building module 2556, a model building module 2557, and a training module 2558 , And model online module 2559, training sample set building module 2556 includes training library generation unit 25561, sampling unit 25562, feature generation unit 25563, training module 2558 includes model training unit 25581, model online module 2559 includes: solidification graph unit 25591 and model push unit 25592, through the training library generating unit 25561 to effectively preprocess the original data, through the sampling unit 25562 to sample the positive and negative samples of the effectively preprocessed original data, and through the feature generating unit 25563 to further characterize the sample information Extract and process to obtain the sample data required for training.
  • the model training unit 25581 in the training module 2558 trains the candidate recommendation information prediction model generated by the model construction module 2557 based on the obtained sample data, and passes the solidified map in the model online module 2559
  • the unit 25591 and the model pushing unit 25592 upload the trained candidate recommendation information prediction model to the recommendation system in the application server 200-B.
  • the trained candidate recommendation information prediction model outputs the candidate recommendation information with the highest ranking, and returns the candidate recommendation information with the highest ranking to the terminal 400.
  • the training server 200-A may be an independent physical server, or a server cluster or a distributed system composed of multiple physical servers, or it may provide cloud services, cloud databases, cloud computing, cloud functions, Cloud servers for basic cloud computing services such as cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.
  • the terminal 400 may be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited to this.
  • the terminal and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in the embodiment of the present application.
  • FIG. 2A is a schematic structural diagram of a training server 200-A applying an artificial intelligence-based recommendation model training method provided by an embodiment of the present application.
  • the training server 200-A shown in FIG. 2A includes: at least one processor 210, Storage 250 and at least one network interface 220.
  • the various components in the terminal 200 are coupled together through the bus system 240.
  • the bus system 240 is used to implement connection and communication between these components.
  • the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for clear description, various buses are marked as the bus system 240 in FIG. 2A.
  • the processor 210 may be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP, Digital Signal Processor), or other programmable logic devices, discrete gates or transistor logic devices, or discrete hardware Components, etc., where the general-purpose processor may be a microprocessor or any conventional processor.
  • DSP Digital Signal Processor
  • the memory 250 may be removable, non-removable, or a combination thereof.
  • Exemplary hardware devices include solid-state memory, hard disk drives, optical disk drives, and so on.
  • the memory 250 optionally includes one or more storage devices that are physically remote from the processor 210.
  • the memory 250 includes volatile memory or non-volatile memory, and may also include both volatile and non-volatile memory.
  • the non-volatile memory may be a read only memory (ROM, Read Only Memory), and the volatile memory may be a random access memory (RAM, Random Access Memory).
  • ROM read only memory
  • RAM Random Access Memory
  • the memory 250 described in the embodiment of the present application is intended to include any suitable type of memory.
  • the memory 250 can store data to support various operations. Examples of these data include programs, modules, and data structures, or a subset or superset thereof, as illustrated below.
  • the operating system 251 includes system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks;
  • the network communication module 252 is used to reach other computing devices via one or more (wired or wireless) network interfaces 220.
  • Exemplary network interfaces 220 include: Bluetooth, Wireless Compatibility Authentication (WiFi), and Universal Serial Bus ( USB, Universal Serial Bus), etc.;
  • the artificial intelligence-based recommendation model training device provided in the embodiments of the present application can be implemented in software.
  • FIG. 2A shows the artificial intelligence-based recommendation model training device 255-A stored in the memory 250, which It can be software in the form of programs and plug-ins, including the following software modules: training sample collection building module 2556, model building module 2557, training module 2558, and model online module 2559. These modules are logical and therefore based on the implemented functions Any combination or further splitting can be carried out, and the functions of each module will be explained below.
  • the artificial intelligence-based recommendation model training device provided in the embodiment of the application may be implemented in hardware.
  • the artificial intelligence-based recommendation model training device provided in the embodiment of the application may be hardware decoding.
  • a processor in the form of a processor which is programmed to execute the artificial intelligence-based recommendation model training method provided in the embodiments of the present application.
  • a processor in the form of a hardware decoding processor may adopt one or more application-specific integrated circuits (ASICs). , Application Specific Integrated Circuit), DSP, Programmable Logic Device (PLD, Programmable Logic Device), Complex Programmable Logic Device (CPLD, Complex Programmable Logic Device), Field Programmable Gate Array (FPGA, Field-Programmable Gate Array) or Other electronic components.
  • ASICs application-specific integrated circuits
  • the first stage is the offline training stage of the candidate recommendation information prediction model.
  • FIG. 3 is a model architecture diagram in the artificial intelligence-based recommendation model training method provided by an embodiment of the present application.
  • the model here is a candidate recommendation information prediction model, user features in the form of discrete vectors (represented by Field0-Field5 in Figure 3), candidate recommendation information features in the form of discrete vectors (represented by Field5-Field12 in Figure 3), and discrete vectors
  • the formal environmental features represented by Field13-Field15 in Figure 3) are transformed into dense embedding vectors.
  • the user features, candidate recommendation information features, and environmental features in the form of dense embedding vectors are input to the pooling layer, and each The sum of all the pixel values of the feature map of the channel, so that each channel gets a real value, N channels will finally get a vector of length N, which is the result of the sum pooling, and the pooling result is input to the map
  • the multi-layer expert network shown in 3 (including expert networks 1-8) and the weight generator corresponding to each index.
  • the weight generator processes the result of pooling based on the following formula (1) to obtain the corresponding index Weight distribution of expert network:
  • g k (x) is the weight of the features output by each layer of the expert network corresponding to the kth index
  • W gk is the weight parameter of the weight generator corresponding to the kth index
  • x is the sum pooling of all input features
  • f k is the output of the multi-layer expert network corresponding to the k-th index
  • g k (x) i is the weight of the feature output of the i-th expert network corresponding to the k-th index
  • f i (x) is the i-th index Characteristics of the output of a network of experts.
  • the features obtained by the weighted summation of the features of each layer of the expert network are input into the fully connected layer.
  • the fully connected layer is composed of a new activation function (Swish(256)).
  • the original formula is shown in formula (3), and the deformation formula is shown in formula (4). ):
  • the Swish activation function here has the characteristics of being unsaturated, smooth, and non-monotonic.
  • ⁇ (beta) 1
  • the activation function has No upper bound has a lower bound, smooth and non-monotonic characteristics, which improves the phenomenon of gradient disappearance during training.
  • ⁇ 0.1
  • the activation function has the characteristics of no upper bound, lower bound, and smooth.
  • ⁇ 10
  • activation The function has the characteristics of no upper bound, lower bound and non-monotonic.
  • the features output by the fully connected layer are output by the Sigmoid activation function to obtain the prediction score corresponding to indicator 1 (click rate output Pctr in Figure 3).
  • the features output by the fully connected layer are passed through " lambda x: x" expression, returns the input value to the output value, and obtains the predicted score corresponding to indicator 2 (the duration output Pdur in Figure 3).
  • the final score corresponding to each indicator is obtained The process can be realized by formula (5):
  • f k is the output of the multi-layer expert network corresponding to the kth index
  • h k is the tower network corresponding to the kth index
  • the tower network is the fully connected layer shown in FIG. 3.
  • FIG. 5A is an optional flowchart of the artificial intelligence-based recommendation model training method provided by an embodiment of the present application, which will be described in conjunction with steps 101-103 shown in FIG. 5A.
  • step 101 the training server preprocesses the log of the recommendation system to construct a training sample set.
  • the training server here is a server used to train the above candidate recommendation information prediction model.
  • the log here includes user behavior data in the recommendation system.
  • User behavior data is data related to user behavior events.
  • the three elements of user behavior events include: operation, definition An operation action (such as click, drag and drop); parameter/attribute, the parameter can be any attribute related to the event, including the business information that triggered the event (person, time, location, equipment, operation information); attribute value, parameter /The value of the attribute, the training sample set is obtained by preprocessing the original data in the log
  • Figure 5C is an optional flow diagram of the artificial intelligence-based recommendation model training method provided by an embodiment of the present application.
  • the log of the recommendation system is preprocessed to construct training samples
  • the collection can be implemented through steps 1011 to 1013.
  • step 1011 at least one of the exposure log, the playback log, the information forward index, and the image feature log in the recommendation system is fused according to the user equipment identification to obtain sample data.
  • first extract and analyze data such as exposure, click, conversion, and stay time in the buried point log such as associating various operations based on the exposure sequence number, and analyze the buried point parameters (for example, the real-time features recorded in the log) , Analyze context features, etc., and then filter the obtained sample data, such as filtering malicious user samples, filtering invalid exposure samples, etc., for example, for the case where the same candidate recommendation information is exposed to the same user multiple times at different times, training In the sample set, there will be situations where the same user clicks and does not click on the same candidate recommendation information. If the interval between multiple exposures is very short, consider using only one of the exposure data, or, to avoid highly active users from affecting the loss function The impact of extracting the same amount of raw data for each user in the log.
  • step 1012 positive sample sampling and negative sample sampling are performed on the obtained sample data, so that the positive samples and negative samples obtained by sampling are kept in a standard ratio.
  • feature extraction is performed on the filtered sample data to generate feature samples.
  • Feature engineering is mainly performed from the two dimensions of user and candidate recommendation information, and also from the three dimensions of user, candidate recommendation information, and environmental information. Perform feature engineering, and then sample the positive and negative samples according to a certain ratio of positive and negative samples, so that the positive and negative samples obtained by sampling are maintained as the standard ratio.
  • the standard ratio is a set ratio, which can be determined according to the prior data of the recommendation system.
  • the positive samples and negative samples here are for each indicator.
  • the indicators can be click-through rate and duration.
  • the positive sample can be various sample data corresponding to the predicted click-through rate, and the sample in the sample data
  • the label is high click-through rate (high click-through rate here is a relative concept, it can be preset to be higher than the click-through rate threshold to be high click-through rate), negative samples can be various sample data corresponding to the predicted click-through rate, and the sample data
  • the sample label of is low click-through rate (the low click-through rate here is a relative concept, and it can be preset to be lower than the click-through rate threshold to be low click-through rate), then the number of negative samples may be a lot, if it is directly used for training, it is not Reasonable, so it is necessary to ensure the ratio of positive and negative samples before training.
  • the original data may have the following problems: the number of samples is too small, and the amount of data is too small to represent the entire sample.
  • the number of samples must be as representative of the entire data as possible; sampling deviation, the number of samples must be guaranteed, and the quality must be guaranteed.
  • some algorithms are very sensitive to the balance of the data, and have high requirements for feature distribution or feature segmentation. For such a model, if you neglect to pay attention to the ratio between positive and negative samples, then Will lead to significant performance degradation; data that does not meet business requirements, in a specific business scenario, there are already very few positive samples. At this time, the trained parameter expression ability is limited. Will the model not meet the business recall requirements?
  • the method of the imbalance problem is to carry out sampling processing, including sampling to obtain positive samples and sampling to obtain negative samples, so that the positive samples and negative samples obtained by sampling are finally kept in a set ratio.
  • step 1013 perform feature extraction processing on the positive and negative samples obtained after sampling to obtain features that match each indicator, and determine the set of true results including the features and corresponding indicators as the training candidate recommendation information prediction model The required training sample collection.
  • feature extraction is performed on the positive sample to obtain at least one feature corresponding to the positive sample and the true result (label) of the indicator corresponding to the positive sample.
  • the indicator is click-through rate
  • the real result of can be a high click-through rate, and the real result of the feature and the corresponding indicator is used as a training sample.
  • the training server constructs a candidate recommendation information prediction model based on a weight generator corresponding to a plurality of indicators, a predictor corresponding to a plurality of indicators, and a plurality of expert networks.
  • the candidate recommendation information prediction model includes a feature input part. Features are classified according to fields. First, they are divided into user features, candidate recommendation information features, and environmental features. The user features pass through the regions according to different types of features. (field) to distinguish, the features are input to the pooling layer, and the pooling layer is input to multiple expert networks (expert networks 1-8).
  • the candidate recommendation information prediction model also includes two weight generators, as shown in Figure 3. The weight generator on the left corresponds to indicator 1, and the weight generator on the right in Figure 3 corresponds to indicator 2.
  • the candidate recommendation information prediction model also includes two tower networks, and each tower network includes a fully connected layer and an activation function. Each predictor corresponds to different indicators (indicator 1 and indicator 2).
  • step 103 the training server performs multi-index training on the candidate recommendation information prediction model through the training sample set, where the candidate recommendation information prediction model obtained by training is used for the recommendation system to perform multi-index aggregation processing and sorting according to the sorting As a result, the candidate recommendation information to be recommended is determined.
  • step 103 multi-index training is performed on the candidate recommendation information prediction model through the training sample set, which can be implemented by the following technical solutions, initialize the candidate recommendation information prediction model, and initialize the loss function corresponding to multiple indicators.
  • the function includes the feature sample and the score of the corresponding feature sample; the following processing is performed during each iteration of the candidate recommendation information prediction model: through the candidate recommendation information prediction model, the feature samples included in the training sample set are scored to obtain the corresponding feature sample The score for each indicator; substitute the true results and scores of the corresponding feature samples into the loss function to determine the corresponding candidate recommendation information prediction model parameters when the loss function obtains the minimum value; update the candidate recommendation based on the determined candidate recommendation information prediction model parameters Information prediction model.
  • the above step of scoring feature samples included in the training sample set by using the candidate recommendation information prediction model to obtain the scores of the corresponding feature samples for each indicator can be achieved through the following technical solutions, through the candidate recommendation information prediction
  • Multiple expert networks in the model map the feature samples to the feature space corresponding to each expert network to obtain the sub-features of the feature dimensions associated with each expert network; based on the sub-feature weight distribution corresponding to the sub-features obtained by each expert network
  • the weight of the feature is to perform weighting processing on the associated sub-features obtained by each expert network to obtain the indicator features of the corresponding feature samples;
  • the predictor corresponding to each indicator included in the candidate recommendation information prediction model is combined with the indicator features to perform the feature sample
  • the score prediction process obtains the score of the characteristic sample based on each index.
  • the feature samples include: user features, candidate recommendation information features, and environmental features.
  • the feature samples are input to a multi-layer expert network and a weight generator corresponding to each index.
  • the weight generator is based on the above formula (1) and
  • the pooled result is processed, and the weight distribution of the expert network for the corresponding index is obtained.
  • the indicator corresponding to the weight generator on the left in Figure 3 is indicator 1, which can be click-through rate, and the indicator corresponding to the weight generator on the right in Figure 3 is indicator 2, and indicator 2 can be the duration.
  • the weight generator on the side can generate 8 weights assigned to expert networks 1-8, and each expert network outputs a sub-feature, which is multiplied by the corresponding weights and then summed to obtain indicators for prediction
  • the detailed process can be found in formula (2).
  • the index features obtained by the weighted summation of the features of each layer of the expert network are input into the fully connected layer in the predictor (tower network).
  • the fully connected layer is composed of a new activation function (Swish(256)).
  • the original formula is shown in the formula (3), see formula (4) for the deformation formula.
  • the above-mentioned step of substituting the true results and scores of the corresponding feature samples into the loss function to determine the corresponding candidate recommendation information prediction model parameters when the loss function obtains the minimum value can be implemented by the following technical solutions, and the corresponding feature samples
  • the actual results and scores for each indicator are substituted into the loss function corresponding to each indicator; combined with the loss weight of each loss function, the weighted summation processing is performed on each loss function, and the aggregate loss function of the corresponding candidate recommendation information prediction model is obtained;
  • the aggregation loss function is minimized, and the candidate recommendation information prediction model parameters corresponding to the minimum value of the aggregation loss function are obtained; among them, the candidate recommendation information prediction model parameters include the structural parameters of the corresponding candidate recommendation information model and the loss of the corresponding aggregation loss function Weights.
  • the aggregate loss function is:
  • ⁇ t is a balance factor constant used to balance positive and negative samples
  • ⁇ is an adjustment constant used to adjust the weight reduction rate of simple samples
  • P t is the probability value of the corresponding score.
  • the weight of the loss function corresponding to each indicator is also different.
  • the aggregate loss function for:
  • L 1 (W) is the first loss function corresponding to the first index
  • L 2 (W) is the second loss function corresponding to the second index
  • ⁇ 1 is used to characterize the uncertainty of the prediction score for the first indicator
  • ⁇ 2 is used to characterize the uncertainty of the prediction score for the second index.
  • the aggregate loss function is:
  • L 1 (W) is the first loss function corresponding to the first index
  • L 2 (W) is the second loss function corresponding to the second index
  • ⁇ 1 is used to characterize the uncertainty of the prediction score for the first indicator
  • ⁇ 2 is used to characterize the uncertainty of the prediction score for the second index.
  • steps 104-105 may be performed. See FIG. 5B.
  • FIG. 5B is an optional flowchart of the artificial intelligence-based recommendation model training method provided by the embodiment of the present application.
  • step 104 the training server converts the parameters of the candidate recommendation information prediction model into constants, and solidifies them in the candidate recommendation information prediction model to generate a solidified binary model file.
  • step 105 the training server pushes the binary model file to the recommendation system, so that the candidate recommendation information prediction model used in the recommendation system is consistent with the candidate recommendation information prediction model obtained through training.
  • the parameters are solidified in the model network structure by converting the model parameters into constants, so as to ensure that the network structure of the offline training model and the online prediction model are consistent. Consistency, the online recommendation system can obtain the network structure and model training parameters at the same time by loading the solidified model file, thereby ensuring consistency.
  • the final binary model file is passed through the timing task according to the preset frequency, by day or by It is pushed to the online recommendation system within hours, and at the same time, the file content generation information summary algorithm is encoded for subsequent model verification. At this point, the entire offline training process is completed.
  • the second stage is the application stage of the candidate recommendation information prediction model, which is applied to the application server 200-B.
  • FIG. 1B is a schematic diagram of an optional architecture of the artificial intelligence-based recommendation system 100-B provided by an embodiment of the present application.
  • the terminal 400 is connected to the application server 200-B and the training server 200-A through the network 300, and the network 300 It can be a wide area network or a local area network, or a combination of the two.
  • the application server 200-B is responsible for sorting the candidate recommendation information through the candidate recommendation information prediction model.
  • the application server 200-B includes an information acquisition module 2551, a feature formation module 2552, The feature processing module 2553, the information ranking module 2554, and the recommendation module 2555.
  • the information acquisition module 2551 acquires candidate recommendation information corresponding to the user request from the recall system 500, and applies The server 200-B sorts the candidate recommendation information through the trained candidate recommendation information prediction model obtained from the training server 200-A, and pushes the candidate recommendation information corresponding to the ranking result to the terminal 400 used by the user according to the ranking result.
  • the application server 200-B may be an independent physical server, or a server cluster or a distributed system composed of multiple physical servers, or it may provide cloud services, cloud databases, cloud computing, cloud functions, Cloud servers for basic cloud computing services such as cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.
  • the terminal 400 may be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited to this.
  • the terminal and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in the embodiment of the present invention.
  • FIG. 2B is a schematic structural diagram of an application server 200-B applying an artificial intelligence-based recommendation method according to an embodiment of the present application.
  • the structure of the application server 200-B shown in FIG. The structure is the same, except that the application server 200-B includes a recommendation device 255-B based on artificial intelligence instead of the recommendation model training device 255-A based on artificial intelligence.
  • the artificial intelligence-based recommending device 255-B provided in the embodiments of the present application can be implemented in software.
  • FIG. 2B shows the artificial intelligence-based recommending device 255-B stored in the memory 250, which can It is software in the form of programs and plug-ins, including the following software modules: information acquisition module 2551, feature formation module 2552, feature processing module 2553, information ranking module 2554, and recommendation module 2555. These modules are logical, so according to the implementation The functions can be combined or split further, and the functions of each module will be described below.
  • the artificial intelligence-based recommendation device provided in the embodiment of the application may be implemented in hardware.
  • the artificial intelligence-based recommendation device provided in the embodiment of the application may be in the form of a hardware decoding processor.
  • a processor which is programmed to execute the artificial intelligence-based recommendation method provided by the embodiments of the present application.
  • a processor in the form of a hardware decoding processor may adopt one or more application specific integrated circuits (ASIC, Application Specific Integrated Circuit) , DSP, Programmable Logic Device (PLD, Programmable Logic Device), Complex Programmable Logic Device (CPLD, Complex Programmable Logic Device), Field Programmable Gate Array (FPGA, Field-Programmable Gate Array) or other electronic components.
  • ASIC Application Specific Integrated Circuit
  • DSP Programmable Logic Device
  • PLD Programmable Logic Device
  • CPLD Complex Programmable Logic Device
  • FPGA Field-Programmable Gate Array
  • FIG. 6A is an optional flowchart of the artificial intelligence-based recommendation method provided by an embodiment of the present application, and will be described in conjunction with steps 201-203 shown in FIG. 6A.
  • step 201 the application server obtains multiple candidate recommendation information corresponding to the user to be recommended.
  • obtaining multiple candidate recommendation information corresponding to the user to be recommended in step 201 can be achieved through the following technical solutions to obtain at least one type of candidate recommendation information: similar to the content of the historical browsing information corresponding to the user to be recommended , And multiple candidate recommendation information whose content similarity is not less than the content similarity threshold; and multiple candidate recommendation information whose behavior is similar to the historical behavior information corresponding to the user to be recommended, and whose behavior similarity is not less than the behavior similarity threshold.
  • multiple candidate recommendation information can be obtained through the recall module in the recommendation system, where multiple candidate recommendation information is obtained in response to a user request of the user to be recommended, and the user request here may be a query carrying a specific target.
  • a request can also be a request to initialize an application.
  • the behavior similarity here refers to the similarity between the user's historical behavior information and the candidate recommendation information
  • the content similarity refers to the similarity between the user's historical browsing information and the candidate recommendation information.
  • step 202 the application server obtains the user characteristics of the user to be recommended and the candidate recommendation information characteristics of each candidate recommendation information, and combines the user characteristics with the candidate recommendation information characteristics of each candidate recommendation information to form a corresponding candidate recommendation information.
  • the fusion feature of the candidate recommendation information is not limited to the candidate recommendation information.
  • obtaining the user characteristics of the user to be recommended and the candidate recommendation information characteristics of each candidate recommendation information in step 202 may be implemented by the following technical solutions to obtain at least one of the following user characteristics corresponding to the user to be recommended: Basic attribute characteristics used to characterize the basic information of users to be recommended; social relationship characteristics used to characterize users' social relationships; interactive behavior characteristics used to characterize user interaction behaviors; reading psychological characteristics used to characterize users' reading preferences; obtain corresponding candidates
  • the candidate recommendation information feature of at least one of the following recommendation information the category feature used to characterize the candidate recommendation information category; the label feature used to characterize the candidate recommendation information content; the time feature used to characterize the release time of the candidate recommendation information; used to characterize The release feature of the candidate recommendation information source; the length feature used to characterize the length of the candidate recommendation information.
  • the basic information may be basic attributes such as the user’s gender, age, long-term residence, etc.
  • the social relationship may be whether the marriage or job position has social attributes, and the user’s interactive behavior may be like, forward, or favorite, etc.
  • the reading preference can be reading interest, the point of interest can be entertainment gossip or international news, etc.
  • the candidate recommended information category can be the information display carrier category, for example, video information, image information or text information, and the content of the candidate recommended information can be It is the content topic, such as education topic or entertainment topic, etc.
  • the user characteristics are combined with the candidate recommendation information characteristics of each candidate recommendation information respectively to form a fusion feature corresponding to each candidate recommendation information, which can be achieved through the following technical solutions to obtain the corresponding to-be-recommended
  • At least one of the following environmental characteristics of the user the time characteristic of pushing to the user to be recommended; the user location characteristics of the user to be recommended; the device characteristics of the user to be recommended; the network characteristics of the device used by the user to be recommended;
  • the environmental characteristics, user characteristics, and candidate recommendation information characteristics of the user to be recommended are combined into a fusion feature corresponding to each candidate recommendation information.
  • environmental characteristics also have an impact on index prediction.
  • the time to push to the user to be recommended will affect whether the user to be recommended is free to view the candidate recommendation information.
  • the location of the user to be recommended represents the current life of the user to be recommended. Scenes, different life scenes have an impact on the prediction of indicators. For example, when the location feature indicates that the user to be recommended is in a movie theater, the prediction results for duration and click-through rate and the location feature indicate the prediction results of the user to be recommended in the study room. There will be a big difference, and the network where the device used by the user to be recommended is located will affect whether the user to be recommended wants to receive candidate recommendation information that requires a lot of network resources, such as videos.
  • step 203 the application server performs multi-level mapping processing on the fusion feature corresponding to each candidate recommendation information, and obtains the scores of multiple indicators corresponding to each candidate recommendation information.
  • Fig. 6B is an optional flowchart of the artificial intelligence-based recommendation method provided by an embodiment of the present application.
  • step 203 multi-level mapping processing is performed on the fusion feature corresponding to each candidate recommendation information.
  • Obtaining the scores corresponding to multiple indicators for each candidate recommendation information can be implemented through steps 2031-2032.
  • step 2031 through multiple expert networks, multi-level mapping processing is performed on the fusion feature corresponding to each candidate recommendation information to obtain multiple sub-features corresponding to multiple expert networks one-to-one, and the multiple sub-features are weighted. Obtain the indicator characteristics corresponding to each indicator.
  • Figure 6C is an optional flowchart of the artificial intelligence-based recommendation method provided by an embodiment of the present application.
  • multiple expert networks are used to determine the fusion features corresponding to each candidate recommendation information.
  • Perform multi-level mapping processing to obtain multiple sub-features corresponding to multiple expert networks one-to-one, and perform weighting processing on multiple sub-features to obtain indicator features corresponding to each indicator, which can be implemented through step 20311 to step 20313.
  • step 20311 through multiple expert networks in the candidate recommendation information prediction model, the fusion features are respectively mapped to the feature space corresponding to each expert network to obtain the sub-features of the feature dimension associated with each expert network.
  • each layer of expert networks can be seen as multiple fully connected layers, and each expert network is usually a fully connected layer with a relatively small scale.
  • step 20312 the fusion feature is subjected to maximum likelihood processing through the one-to-one corresponding weight generator of the multiple indicators in the candidate recommendation information prediction model to obtain the sub-feature weight distribution corresponding to each indicator.
  • the weight generator is used to select the signal proportion of each expert network.
  • Each expert network has its own prediction direction. Finally, they work together to correspond to the indicators of each weight generator.
  • Each here refers to multiple weight generators. Each of the meanings of each in the full text represents each of the multiple objects.
  • step 20313 based on the weights of the sub-features corresponding to the sub-features obtained by each expert network in the sub-feature weight distribution, weighting processing is performed on the associated sub-features obtained by each expert network to obtain the index features corresponding to each index.
  • a predictor corresponding to multiple indicators in the candidate recommendation information prediction model is used to perform scoring prediction processing on the fusion feature in combination with the indicator feature corresponding to each indicator to obtain the score of the fusion feature based on each indicator.
  • each weight generator as a weighted sum pooling operation. If the weight generator is replaced with the largest selection operation, x is the input, the expert network corresponding to the largest output sub-feature component of each layer of expert network is unique Select to pass the signal upwards.
  • step 204 the application server performs multi-index aggregation processing on the scores corresponding to multiple indicators to obtain a comprehensive score of each candidate recommendation information, and sorts the multiple candidate recommendation information in descending order according to the comprehensive score of each candidate recommendation information .
  • step 204 multi-index aggregation processing is performed on the scores corresponding to multiple indicators to obtain a comprehensive score of each candidate recommendation information, which can be implemented by the following technical solutions to obtain aggregation rules corresponding to multiple indicators;
  • the operators included in the aggregation rule perform calculation processing on the scores respectively corresponding to multiple indicators to obtain a comprehensive score corresponding to multiple indicators for each candidate recommendation information.
  • the aggregation rules of multiple indicators may correspond to different multi-indicator aggregation processing methods.
  • the multiple indicators here may be at least two of the indicators such as duration, click-through rate, number of favorites, number of forwardings, etc., and the aggregation rule may It is the addition, multiplication or other calculation rules with parameters.
  • the two indicators of duration and click-through rate can be multiplied as a multi-index aggregation processing method.
  • step 205 the application server selects at least one candidate recommendation information ranked higher among the results of the descending order, and performs a recommendation operation corresponding to the user to be recommended based on the selected candidate recommendation information.
  • the top N candidate recommendation information are selected in the descending order as the candidate recommendation information to be recommended to the user to be recommended, where N is a positive integer and is based on the selected candidate recommendation information Perform the recommendation operation corresponding to the user to be recommended.
  • the recommendation operation can be directly pushed to the user to be recommended, or it can be re-ranked based on the selected candidate recommendation information.
  • the re-ranking here refers to the aggregation of different multi-index processing methods.
  • the selected candidate recommendation information is reordered to obtain recommendation information that the user to be recommended likes from a more comprehensive perspective.
  • the recommendation operation corresponding to the user to be recommended is performed based on the selected candidate recommendation information.
  • the candidate recommendation information generated here as the recommendation result is based on the result of multi-index prediction, rather than a single item.
  • the result of the indicator prediction so the recommendation accuracy rate is high, and it is easy to arouse the interest of the user to be recommended.
  • the user to be recommended will read, click, bookmark, and forward the candidate recommendation information as the recommendation result, thereby increasing the proportion of interactive behaviors.
  • the operations that the user will read, click, bookmark, and forward will be archived as the original information of the user portrait, and used to train the model at time intervals to continuously improve the prediction performance of the model.
  • FIG. 7 is a schematic diagram of the application scenario product of the artificial intelligence-based recommendation method provided by the embodiment of the present application.
  • the interface shows Three sorted candidate recommendation information.
  • the sorted candidate recommendation information is obtained by sorting through the candidate recommendation information prediction model.
  • the application product can be a news client, which effectively utilizes the implicit and explicit behaviors of the user history , Based on the multi-level improved multi-task learning training method to train the candidate recommendation information prediction model in the recommendation system, and at the same time make configurable design for the ranking target of the online recommendation system to provide users with accurate and personalized news recommendation.
  • Figure 8 is a candidate recommendation information prediction model training and application architecture diagram provided by an embodiment of this application.
  • the architecture diagram shown in Figure 8 includes an offline model training part and an online recommendation system part. After the user requests, the personalized recommendation information list can be returned.
  • the units (1)-(6) that directly interact with the central control of the recommendation service are all online services, and (7) is the offline part, which will be explained separately below.
  • the service access part (1) receives user requests and obtains the relevant characteristics of the users to be recommended.
  • the relevant characteristics here can be user exposure, user feedback, and browsing time.
  • the recall part (3) obtains personalized information in response to user requests.
  • Candidate recommendation information through the online ranking part (4) to sort the obtained candidate recommendation information, through the re-ranking part (5) to re-rank the candidate recommendation information obtained after sorting based on different strategies, through the user portrait part (2)
  • the user portrait is acquired, feature extraction is performed through the feature acquisition part (6), and the model is trained offline through the offline model training part (7).
  • Service access part (1) Receive user requests sent by the user's client, and obtain user characteristics of the user based on the user request.
  • the user characteristics here may include preference characteristics, behavior characteristics, etc.
  • the central control of the recommended service responds to the user Request, obtain candidate recommendation information matching the user’s request from the recall part, and use the candidate recommendation information prediction model provided by the embodiment of this application to perform multi-index aggregation sorting based on the click-through rate on the candidate recommendation information obtained in the recall part, and recommend The result is returned to the client through the service access part.
  • Figure 7 shows the recommendation result returned to the client.
  • the three candidate recommendation information shown here are all recommended results that meet the user characteristics and can be obtained through the service access part User exposure data, user feedback data, and browsing time data.
  • Recall part (3) Trigger as many correct results as possible from the full information collection, and return the results to the online sorting part (4).
  • recall There are many ways to recall, including collaborative filtering, topic models, content recalls, and hotspot recalls.
  • the recommendation system the user will not provide a clear search term input, so the recommendation system needs to recommend content that may be of interest to the user based on various information such as user portraits and content portraits.
  • the feature generator is used to obtain the characteristics of the recommendation task, and the click rate of each candidate recommendation information (click rate result) is predicted based on the obtained features and the candidate recommendation information prediction model .
  • the source of feature acquisition can be user logs, news details, user conversations, and real-time features.
  • Reordering part (5) The difference between sorting and reordering is that the result of sorting is the result of a certain aggregation rule, but a single aggregation rule has limitations, so it needs to be reordered.
  • the process of sorting use aggregation rules that are different from the aggregation rules on which the sorting is based. For example, in the sorting phase, the product of click rate and duration is used as the aggregation method, while the reordering phase uses the product of the forwarding rate and duration as the aggregation method. , To re-rank the sorted candidate recommendation information.
  • the offline model training part can include: training sample collection building module, model building module, training module, and model online module.
  • the training sample collection building module includes training library generation unit and sampling unit , Feature generation unit, training module includes model training unit, model online module includes: solidification graph unit and model push unit.
  • the training library generation unit is used to effectively preprocess the original data, which is the prerequisite for improving the accuracy of the training model.
  • the user’s exposure log, playback log, information forward index and portrait feature log are merged according to the user’s device number, and data with missing valid single features are deleted at the same time to ensure that the valid single feature information of each sample is not missing.
  • the sampling unit is used for Sample positive and negative samples based on the original training library obtained in the training library generation unit, so that the positive and negative samples are kept in a reasonable ratio (set ratio), while filtering the samples whose playback duration is below a certain threshold, and for different information Time length and viewing time adopt a segmented threshold positive and negative sample setting scheme.
  • the feature generation unit feature generator
  • the feature generation unit is used to further extract the sample information obtained by the sampling unit, and select appropriate features in combination with specific prediction targets to generate offline
  • the model training unit is used to use the sample data generated by the feature generation unit to perform offline model training based on the multi-level improved multi-task learning model.
  • the candidate recommendation information prediction model uses multiple layers Expert networks and multi-layer expert networks promote better sharing of underlying features for multiple tasks.
  • An independent threshold network for each task is used to determine the degree of use of the results of different expert networks.
  • Both the expert network and the threshold network are three-dimensional tensors.
  • the expert network can It is a simple fully connected layer structure, and the candidate recommendation information prediction model is based on the following formula to output the score of each indicator:
  • y k is the output corresponding to the kth index
  • h k is the network of the predictor corresponding to the kth index
  • f k is the output of the multi-layer expert network corresponding to the kth index
  • g k (x) i is The weight of the feature output by the i-th expert network corresponding to the k-th index
  • f i (x) is the output feature of the i-th expert network
  • W gk is the weight parameter of the threshold network corresponding to the k-th index
  • x is the input
  • W i is the weight corresponding to the i-th index loss function weights
  • L i is the corresponding loss function i-th index
  • L total polymerization loss function multitasking although simple network search mode is valid, but the weight adjustment It is time-consuming, therefore, uncertainty is introduced to measure the weight of the loss function among multiple tasks.
  • the cognitive uncertainty explains the model. Parameter uncertainty, cognitive uncertainty can be eliminated by adding training data; accidental uncertainty, if there is a relatively large labeling error during data labeling, this error is not brought by the model, but the data itself , The greater the bias in the data set, the greater the chance uncertainty. Among them, chance uncertainty can be subdivided into two categories: (1) Data-dependent or heteroscedastic uncertainty.
  • Task-dependent or homoscedastic uncertainty does not depend on the input data, nor the model output result, but the same constant for all input data
  • task-dependent uncertainty because in multi-task learning, task uncertainty indicates relative confidence and reflects the inherent uncertainty in regression and classification problems. Therefore, the same variance uncertainty is used as noise to optimize the weights in multi-task learning.
  • the multi-task loss function uses the same variance uncertainty to maximize the Gaussian likelihood estimation.
  • f W is the output of the neural network
  • x is the input data
  • W is the weight.
  • the output is usually pushed into the activation function, see the following formula:
  • y 1 is the output of each subtask in the multi-task. Therefore, the maximum likelihood estimation can be shown by the following formula (14), the maximum likelihood estimation is proportional to the norm in formula (14), where , ⁇ is the standard deviation of the Gaussian distribution and the noise of the model.
  • the next task is to maximize the likelihood distribution according to W and ⁇ :
  • L 1 (W) is the first loss function corresponding to the first index
  • L 2 (W) is the second loss function corresponding to the second index
  • ⁇ 1 is used to characterize the uncertainty of the prediction score for the first indicator
  • ⁇ 2 is used to characterize the uncertainty of the prediction score of the second index
  • W is the structural weight corresponding to each index in the candidate recommendation information prediction model.
  • the task of training is to minimize this maximum likelihood estimate, so when ⁇ (noise) increases, the corresponding weight will decrease. On the other hand, as the noise ⁇ decreases, the corresponding weight will increase. .
  • the loss function for independent output needs to be modified on the basis of the two-category cross-entropy loss function.
  • the modified loss function is as the above formula (7) Shown:
  • ⁇ t is the balance factor constant used to balance positive and negative samples
  • ⁇ is the adjustment constant used to adjust the weight reduction rate of simple samples
  • P t is the probability value of the corresponding score
  • ⁇ is a positive number, preferably 1.
  • the activation function has the characteristics of no upper bound and lower bound, smooth and non-monotonic, which improves the phenomenon of gradient disappearance during training.
  • the offline model training part also includes a solidification graph unit, which converts the parameters of the model into constants by fusing the model training parameter data with the network structure, so as to solidify the parameters in the model network structure to ensure offline training models and online predictions
  • the consistency of the model network structure, the online recommendation system can obtain the network structure and model training parameters at the same time by loading the solidified model file, thereby ensuring consistency.
  • the offline model training part also includes a model push unit to obtain the final binary model The files are pushed to the online recommendation system on a daily or hourly basis through a timed task according to a preset frequency. At the same time, the file content generation information summary algorithm is encoded for subsequent model verification, and the entire offline process is completed.
  • the online sorting part (4) is described below.
  • the online sorting part (4) also includes a feature generation unit (feature generator) and an online model prediction part, etc.
  • the feature generation unit in the online process and the feature generation in the offline process The design of the unit is the same.
  • the server After the server receives the user’s request, it obtains the user’s relevant characteristics, and at the same time obtains the candidate recommendation information matched by each user’s request through the recall part, pulls the candidate recommendation information characteristics of each candidate recommendation information, and compares them with the user.
  • the feature combination obtains the sample format consistent with the feature generation unit in the offline process.
  • the online model prediction part based on the improved multi-task model network structure of the offline training model, predicts multiple indicators during online prediction (such as click rate, duration, collection rate) , Like rate, etc.), and at the same time design a variety of target aggregation and sorting methods, such as click-through rate * duration, as the sorting target value, select the top N items as the final return recommendation result.
  • the above improvements optimize the candidate recommendation information Predict the details of the model and strengthen the anti-noise ability of the model, thereby significantly improving the key indicators of the online recommendation system.
  • the test results of the model are shown in Tables 1 and 2. The above experiments show that the joint learning click rate and duration can obtain better generalization
  • the ability to transform online recommendations has been significantly positively driving the core indicators of online recommendations, and the proportion of user interactions and explicit behaviors has also been significantly improved:
  • the embodiment of this application provides an artificial intelligence-based recommendation method, which is a multi-task learning recommendation method based on multi-level improvement. It is designed to design a multi-layer expert network, introduce a multi-index uncertainty factor, and solve the problem of classification imbalance. From a similar perspective, the ranking model of the personalized recommendation system has been deeply optimized and reconstructed. The offline evaluation of the model AUC, mean square error and other indicators have been significantly improved. At the same time, the ranking target of the online recommendation system is configurable. Time length, like rate, sharing rate, etc.), more flexible and controllable. This solution has been successfully applied to the recommendation system of news applications, and finally the core indicators of online recommendation are obviously positively driven, and the proportion of user interaction and explicit behavior is also Significant improvement.
  • the artificial intelligence-based recommendation device 255 stored in the memory 250
  • the software module in -B may include: an information acquisition module 2551, configured to acquire multiple candidate recommendation information corresponding to the user to be recommended; feature formation module 2552, configured to acquire the user characteristics of the user to be recommended and the information of each candidate recommendation information Candidate recommendation information features, and combine the user features with the candidate recommendation information features of each candidate recommendation information to form a fusion feature corresponding to each candidate recommendation information; feature processing module 2553, configured to match each candidate recommendation information Multi-level mapping processing is performed on the fusion features of, and each candidate recommendation information is scored corresponding to multiple indicators; the information sorting module 2554 is configured to perform multi-index aggregation processing on the scores corresponding to multiple indicators to obtain each candidate recommendation information According to the comprehensive score of each candidate recommendation information, the multiple candidate recommendation information is sorted in descending order
  • the information obtaining module 2551 is further configured to obtain at least one of the following types of candidate recommendation information: similar to the content of the historical browsing information corresponding to the user to be recommended, and the content similarity is not less than the content similarity threshold. Pieces of candidate recommendation information; multiple pieces of candidate recommendation information that are similar in behavior to the historical behavior information of the user to be recommended, and whose behavior similarity is not less than the behavior similarity threshold.
  • the feature forming module 2552 is further configured to: obtain at least one of the following user features corresponding to the user to be recommended: basic attribute features used to characterize the basic information of the user to be recommended; Social relationship characteristics; interactive behavior characteristics used to characterize user interaction behavior; reading psychological characteristics used to characterize user reading preferences; obtain at least one of the following candidate recommendation information characteristics corresponding to candidate recommendation information: used to represent the candidate recommendation information category Category features; label features used to characterize candidate recommendation information content; time features used to characterize the release time of candidate recommendation information; release features used to characterize candidate recommendation information sources; length features used to characterize the length of candidate recommendation information.
  • the feature forming module 2552 is further configured to: obtain at least one of the following environmental features corresponding to the user to be recommended: the time feature pushed to the user to be recommended; the user location feature of the user to be recommended; the user to be recommended The device characteristics of the device to be recommended; the network characteristics of the device used by the user to be recommended; the environmental characteristics, user characteristics, and candidate recommendation information characteristics corresponding to the user to be recommended are combined into a fusion feature corresponding to each candidate recommendation information.
  • the feature processing module 2553 is further configured to perform multi-level mapping processing on the fusion feature corresponding to each candidate recommendation information through multiple expert networks to obtain multiple sub-features corresponding to the multiple expert networks one-to-one , And weighting multiple sub-features to obtain the indicator features corresponding to each indicator; using the predictor corresponding to multiple indicators in the candidate recommendation information prediction model, and combining the indicator features corresponding to each indicator to score and predict the fusion feature After processing, the score of the fusion feature based on each index is obtained.
  • the feature processing module 2553 is further configured to map the fusion features to the feature space corresponding to each expert network through multiple expert networks in the candidate recommendation information prediction model, so as to obtain the information associated with each expert network. Sub-features of feature dimensions;
  • the fusion feature is processed with maximum likelihood to obtain the sub-feature weight distribution corresponding to each indicator; based on the sub-feature weight distribution corresponding to each expert network The weights of the obtained sub-features are weighted to the associated sub-features obtained by each expert network, and the index features corresponding to each index are obtained respectively.
  • the information ranking module 2554 is further configured to: obtain aggregation rules corresponding to multiple indicators; based on the operators included in the aggregation rules, perform calculation processing on the scores corresponding to the multiple indicators to obtain each candidate recommendation The information corresponds to a comprehensive score of multiple indicators.
  • the software modules stored in the artificial intelligence-based recommendation model training device 255-A of the memory 250 may include: a training sample set construction module 2556, configured to perform a log of the recommendation system Preprocessing to build a training sample set; model building module 2557, configured to build candidate recommendation information based on a weight generator corresponding to multiple indicators one-to-one, predictors corresponding to multiple indicators one-to-one, and multiple expert networks Prediction model; training module 2558, configured to train candidate recommendation information prediction models for multiple indicators through a collection of training samples; among them, the candidate recommendation information prediction model obtained by training is used for the recommendation system to aggregate and sort multiple indicators to The candidate recommendation information to be recommended is determined according to the ranking result.
  • a training sample set construction module 2556 configured to perform a log of the recommendation system Preprocessing to build a training sample set
  • model building module 2557 configured to build candidate recommendation information based on a weight generator corresponding to multiple indicators one-to-one, predictors corresponding to multiple indicators one-to-one, and multiple expert networks Prediction
  • the device 255-A further includes: a model online module 2559, configured to: convert the parameters of the candidate recommendation information prediction model into constants, and solidify them in the candidate recommendation information prediction model to generate a solidified binary model File: Push the binary model file to the recommendation system so that the candidate recommendation information prediction model used in the recommendation system is consistent with the candidate recommendation information prediction model obtained after training.
  • a model online module 2559 configured to: convert the parameters of the candidate recommendation information prediction model into constants, and solidify them in the candidate recommendation information prediction model to generate a solidified binary model File: Push the binary model file to the recommendation system so that the candidate recommendation information prediction model used in the recommendation system is consistent with the candidate recommendation information prediction model obtained after training.
  • the training sample set construction module 2556 is further configured to perform fusion processing on at least one of the exposure log, the playback log, the information frontal index, and the portrait feature log in the recommendation system according to the user equipment identification, Obtain sample data; perform positive sample sampling and negative sample sampling on the obtained sample data, so that the positive sample and negative sample obtained by sampling are kept in standard proportions; perform feature extraction processing on the positive and negative samples obtained after sampling, and obtain The features matched with each indicator, and the set of real results including the features and the corresponding indicators are determined as the training sample set required for training the candidate recommendation information prediction model.
  • the training module 2558 is further configured to initialize the candidate recommendation information prediction model, and initialize the loss function corresponding to multiple indicators.
  • the loss function includes feature samples and scores of corresponding feature samples; in the candidate recommendation information prediction model The following processing is performed during each iteration of the training process: through the candidate recommendation information prediction model, the feature samples included in the training sample set are scored, and the scores of the corresponding feature samples for each indicator are obtained; the true results and scores of the corresponding feature samples are substituted into the loss Function to determine the corresponding candidate recommendation information prediction model parameters when the loss function obtains the minimum value; update the candidate recommendation information prediction model according to the determined candidate recommendation information prediction model parameters.
  • the training module 2558 is further configured to map the feature samples to the feature space corresponding to each expert network through multiple expert networks in the candidate recommendation information prediction model to obtain the features associated with each expert network Dimensional sub-features; based on the weights of the sub-features corresponding to the sub-features obtained by each expert network in the sub-feature weight distribution, weighting the associated sub-features obtained by each expert network, respectively, to obtain the corresponding feature samples' index features; through candidate recommendation information
  • the predictor corresponding to each index included in the prediction model performs scoring prediction processing on the feature sample in combination with the index feature, and obtains the score of the feature sample based on each index.
  • the training module 2558 is further configured to: substitute the actual results and scores of the corresponding feature samples for each indicator into the loss function corresponding to each indicator; combine the loss weights corresponding to each loss function to perform the calculation on each loss function.
  • the weighted summation process is used to obtain the aggregate loss function of the corresponding candidate recommendation information prediction model; the aggregate loss function is minimized to obtain the candidate recommendation information prediction model parameters corresponding to the minimum value of the aggregate loss function; among them, the candidate recommendation information prediction model
  • the parameters include the structural parameters of the corresponding candidate recommendation information model and the loss weight of the corresponding aggregate loss function.
  • the embodiment of the present application provides a storage medium storing executable instructions, and the executable instructions are stored therein.
  • the processor will cause the processor to execute the artificial intelligence-based recommendation method provided by the embodiments of the present application.
  • the recommendation method based on artificial intelligence as shown in FIGS. 5A-5C and the training method of recommendation model based on artificial intelligence for example, the recommendation method based on artificial intelligence as shown in FIGS. 6A-6C.
  • the storage medium may be FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM, etc.; it may also be various devices including one or any combination of the foregoing memories. .
  • the executable instructions may be in the form of programs, software, software modules, scripts or codes, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and their It can be deployed in any form, including being deployed as an independent program or as a module, component, subroutine or other unit suitable for use in a computing environment.
  • executable instructions may but do not necessarily correspond to files in the file system, and may be stored as part of a file that saves other programs or data, for example, in a HyperText Markup Language (HTML, HyperText Markup Language) document
  • HTML HyperText Markup Language
  • One or more scripts in are stored in a single file dedicated to the program in question, or in multiple coordinated files (for example, a file storing one or more modules, subroutines, or code parts).
  • executable instructions can be deployed to be executed on one computing device, or on multiple computing devices located in one location, or on multiple computing devices that are distributed in multiple locations and interconnected by a communication network Executed on.
  • user characteristics and candidate recommendation information features of candidate recommendation information are combined to form a fusion feature, and through multi-level mapping processing on the fusion feature, scores corresponding to multiple indicators are obtained, and based on Multi-index aggregation processing is used to sort, to positively drive the indicators in the recommendation system, and to increase the proportion of user interaction display behaviors.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Optimization (AREA)
  • Mathematical Analysis (AREA)
  • Algebra (AREA)
  • Pure & Applied Mathematics (AREA)
  • Computational Mathematics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本申请提供了一种基于人工智能的推荐方法、装置、电子设备及计算机可读存储介质；方法包括：获取待推荐对象的对象特征、每个候选推荐信息的候选推荐信息特征，将对象特征分别与每个候选推荐信息的候选推荐信息特征组合成对应每个候选推荐信息的融合特征；对对应每个候选推荐信息的融合特征进行多层次映射处理，得到每个候选推荐信息分别对应多个指标的评分；对分别对应多个指标的评分进行多指标聚合处理，得到每个候选推荐信息的综合评分，以对多个候选推荐信息进行降序排序；在降序排序的结果中选择排序靠前的至少一个候选推荐信息，以执行对应待推荐对象的推荐操作。

Description

基于人工智能的推荐方法、装置、电子设备及存储介质
相关申请的交叉引用
本申请基于申请号为202010095076.2、申请日为2020年02月13日的中国专利申请提出，并要求该中国专利申请的优先权，该中国专利申请的全部内容在此引入本申请作为参考。
技术领域
本申请涉及人工智能技术，尤其涉及一种基于人工智能的推荐方法、装置、电子设备及计算机可读存储介质。
背景技术
人工智能(AI，Artificial Intelligence)是利用数字计算机或者数字计算机控制的机器模拟、延伸和扩展人的智能，感知环境、获取知识并使用知识获得最佳结果的理论、方法和技术及应用系统。
随着信息技术和互联网行业的发展，信息过载已经成为人们处理信息的挑战，个性化推荐系统通过更加精准的理解用户意图有效缓解了这类问题，但是，相关技术中的通过线性排序模型或者深度排序模型进行点击率预估的方式，仍然会推荐给用户排序靠前但是并不是用户真正喜欢的信息，这导致了信息的无效推荐，对于推荐系统的计算资源和通信资源造成了不必要的消耗。
发明内容
本申请实施例提供一种基于人工智能的推荐方法、装置、电子设备及计算机可读存储介质，能够实现信息的精准推荐。
本申请实施例的技术方案是这样实现的：
本申请实施例提供一种基于人工智能的推荐方法，包括：
获取对应待推荐对象的多个候选推荐信息；
获取所述待推荐对象的对象特征、以及每个候选推荐信息的候选推荐信息特征，并将所述对象特征分别与每个候选推荐信息的候选推荐信息特征进行组合，以形成对应每个候选推荐信息的融合特征；
对对应每个候选推荐信息的融合特征进行多层次映射处理，得到每个候选推荐信息分别对应多个指标的评分；
对所述分别对应多个指标的评分进行多指标聚合处理，得到每个候选推荐信息的综合评分，并根据每个候选推荐信息的综合评分对所述多个候选推荐信息进行降序排序；
在所述降序排序的结果中选择排序靠前的至少一个候选推荐信息，并基于 所选择的候选推荐信息执行对应所述待推荐对象的推荐操作。
本申请实施例提供一种基于人工智能的推荐装置，包括：
信息获取模块，用于获取对应待推荐对象的多个候选推荐信息；
特征形成模块，用于获取所述待推荐对象的对象特征、以及每个候选推荐信息的候选推荐信息特征，并将所述对象特征分别与每个候选推荐信息的候选推荐信息特征进行组合，以形成对应每个候选推荐信息的融合特征；
特征处理模块，用于对对应每个候选推荐信息的融合特征进行多层次映射处理，得到每个候选推荐信息分别对应多个指标的评分；
信息排序模块，用于对所述分别对应多个指标的评分进行多指标聚合处理，得到每个候选推荐信息的综合评分，并根据每个候选推荐信息的综合评分对所述多个候选推荐信息进行降序排序；
推荐模块，用于在所述降序排序的结果中选择排序靠前的至少一个候选推荐信息，并基于所选择的候选推荐信息执行对应所述待推荐对象的推荐操作。
本申请实施例提供一种基于人工智能的推荐模型训练方法，包括：
对推荐系统的日志进行预处理，以构建训练样本集合；
基于与多个指标一一对应的权重生成器、与所述多个指标一一对应的预测器、以及多个专家网络，构建候选推荐信息预测模型；
通过所述训练样本集合，对所述候选推荐信息预测模型进行多指标训练；
其中，所述训练得到的候选推荐信息预测模型用于供所述推荐系统进行多指标的聚合处理和排序，以根据排序结果确定待推荐的候选推荐信息。
本发明实施例提供一种基于人工智能的推荐模型训练装置，包括：
训练样本集合构建模块，用于对推荐系统的日志进行预处理，以构建训练样本集合；
模型构建模块，用于基于与多个指标一一对应的权重生成器、与所述多个指标一一对应的预测器、以及多个专家网络，构建候选推荐信息预测模型；
训练模块，用于通过所述训练样本集合，对所述候选推荐信息预测模型进行多指标训练；
其中，所述训练得到的候选推荐信息预测模型用于供所述推荐系统进行多指标的聚合处理和排序，以根据排序结果确定待推荐的候选推荐信息。
本申请实施例提供一种电子设备，包括：
存储器，用于存储可执行指令；
处理器，用于执行所述存储器中存储的可执行指令时，实现本申请实施例提供的基于人工智能的推荐方法、以及基于人工智能的推荐模型训练方法。
本申请实施例提供一种计算机可读存储介质，存储有可执行指令，用于引起处理器执行时，实现本申请实施例提供的基于人工智能的推荐方法、以及基于人工智能的推荐模型训练方法。
本申请实施例具有以下有益效果：
通过对融合特征的多层次映射处理、多指标聚合处理的综合评分进行排序以选取待推荐的信息，相对于单目标预测的更加全面和客观，从而实现了精准的个性化推荐，提高推荐系统的计算资源和通信资源的利用效率。
附图说明
图1A是本申请实施例提供的基于人工智能的推荐系统100-A的一个可选的架构示意图；
图1B是本申请实施例提供的基于人工智能的推荐系统100-B的一个可选的架构示意图；
图2A是本申请实施例提供的应用基于人工智能的推荐模型训练方法的训练服务器200-A的结构示意图；
图2B是本申请实施例提供的应用基于人工智能的推荐方法的应用服务器200-B的结构示意图；
图3是本申请实施例提供的基于人工智能的推荐模型训练方法中的模型架构图；
图4是本申请实施例提供的新型激活函数示意图
图5A-5C是本申请实施例提供的基于人工智能的推荐模型训练方法的一个可选的流程示意图；
图6A-6C是本申请实施例提供的基于人工智能的推荐方法的一个可选的流程示意图；
图7是本申请实施例提供的基于人工智能的推荐方法的应用场景产品示意图；
图8是本申请实施例提供的候选推荐信息预测模型训练及应用架构图。
具体实施方式
为了使本申请的目的、技术方案和优点更加清楚，下面将结合附图对本申请作进一步地详细描述，所描述的实施例不应视为对本申请的限制，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其它实施例，都属于本申请保护的范围。
在以下的描述中，涉及到“一些实施例”，其描述了所有可能实施例的子集，但是可以理解,“一些实施例”可以是所有可能实施例的相同子集或不同子集，并且可以在不冲突的情况下相互结合。
在以下的描述中，所涉及的术语“第一\第二”仅仅是是区别类似的对象，不代表针对对象的特定排序，可以理解地，“第一\第二”在允许的情况下可以互换特定的顺序或先后次序，以使这里描述的本申请实施例能够以除了在这里图示或描述的以外的顺序实施。
除非另有定义，本文所使用的所有的技术和科学术语与属于本申请的技术领域的技术人员通常理解的含义相同。本文中所使用的术语只是为了描述本申请实施例的目的，不是旨在限制本申请。
对本申请实施例进行进一步详细说明之前，对本申请实施例中涉及的名词和术语进行说明，本申请实施例中涉及的名词和术语适用于如下的解释。
1)多任务学习(MTL，Multi-Task Learning)，利用多个学习任务中所包含 的有用信息，为每个学习任务得到更为准确的学习器。
2)受试者工作特征曲线下的面积(AUC，Area Under Curve)，作为模型的评价标准，通过对受试者工作特征曲线下各部分的面积求和而得。
3)待推荐对象，即进行信息推荐的目标，由于信息呈现的媒介是终端，信息推荐的目标是操作相应终端的用户，因此下文中将“对象”与“用户”进行了等同的描述。可以理解地，这里的用户可以是能够操作终端的自然人，也可以是终端中运行的能够模拟人类行为的机器人程序。
4)候选推荐信息，即能够发送到终端中进行呈现以向相应终端的用户进行推荐的信息。
5)专家网络，每个专家网络都是一个前向传播网络，其输出为特征，分别对应于不同任务，为每个任务引入一个权重生成结构，输出专家网络的权重值，使得不同的综合任务以不同的方式利用专家网络的输出。
申请人发现对于真实场景中的推荐系统的个性化排序阶段，面临一些问题，主要是有很多不同甚至是冲突的优化目标，比如不仅希望用户观看，还希望用户能给出高评价并分享，推荐系统中也会经常有一些隐性偏见，比如用户是因为资讯内容排得靠前而点击，而非用户真正喜欢，因此模型产生的数据会引发模型训练偏置，从而形成一个反馈循环，越来越偏，为了有效的解决上述问题，本申请实施例提出了一种基于人工智能的推荐方法，分别通过设计多层专家网络、引入多指标不确定性因子、解决分类不平衡问题、改进模型激活函数等角度，对个性化推荐系统的候选推荐信息预测模型进行了深度优化改造，模型离线评测的AUC、均方根误差等指标有了明显提升，同时对在线推荐系统的排序目标做可配置性的设计，使得排序结果更加灵活可控，最终能够正向带动线上推荐的各项核心指标，用户互动显式行为比例也有了明显提升。
本申请实施例提供一种基于人工智能的推荐方法、装置、电子设备和计算机可读存储介质，能够正向带动线上推荐的各项核心指标，用户互动显式行为比例也有了明显提升，下面说明本申请实施例提供的电子设备的示例性应用，本申请实施例提供的电子设备可以实施为笔记本电脑，平板电脑，台式计算机，机顶盒，移动设备(例如，移动电话，便携式音乐播放器，个人数字助理，专用消息设备，便携式游戏设备)，车载计算机等各种类型的用户终端，也可以实施为服务器。下面，将说明设备实施为服务器时示例性应用。
参见图1A，图1A是本申请实施例提供的基于人工智能的推荐系统100-A的一个可选的架构示意图，终端400通过网络300连接应用服务器200-B以及训练服务器200-A，网络300可以是广域网或者局域网，又或者是二者的组合，训练服务器200-A负责离线训练候选推荐信息预测模型，训练服务器200-A中包括训练样本集合构建模块2556、模型构建模块2557、训练模块2558、以及模型上线模块2559，训练样本集合构建模块2556中包括训练库生成单元25561、采样单元25562、特征生成单元25563、训练模块2558中包括模型训练单元25581、模型上线模块2559中包括：固化图单元25591以及模型推送单元25592，通过训练库生成单元25561对原始数据进行有效预处理，通过采样单元25562 对经过有效预处理的原始数据进行正负样本采样，通过特征生成单元25563将样本信息做进一步特征提取处理，得到训练所需的样本数据，训练模块2558中的模型训练单元25581基于得到的样本数据进行训练由模型构建模块2557生成的候选推荐信息预测模型，并通过模型上线模块2559中的固化图单元25591以及模型推送单元25592将训练好的候选推荐信息预测模型上线至应用服务器200-B中的推荐系统。响应于应用服务器200-B接收到的来自于终端400的用户请求，通过训练好的候选推荐信息预测模型输出排序靠前的候选推荐信息，并将排序靠前的候选推荐信息返回至终端400。
在一些实施例中，训练服务器200-A可以是独立的物理服务器，也可以是多个物理服务器构成的服务器集群或者分布式系统，还可以是提供云服务、云数据库、云计算、云函数、云存储、网络服务、云通信、中间件服务、域名服务、安全服务、CDN、以及大数据和人工智能平台等基础云计算服务的云服务器。终端400可以是智能手机、平板电脑、笔记本电脑、台式计算机、智能音箱、智能手表等，但并不局限于此。终端以及服务器可以通过有线或无线通信方式进行直接或间接地连接，本申请实施例中不做限制。
参见图2A，图2A是本申请实施例提供的应用基于人工智能的推荐模型训练方法的训练服务器200-A的结构示意图，图2A所示的训练服务器200-A包括：至少一个处理器210、存储器250、至少一个网络接口220。终端200中的各个组件通过总线系统240耦合在一起。可理解，总线系统240用于实现这些组件之间的连接通信。总线系统240除包括数据总线之外，还包括电源总线、控制总线和状态信号总线。但是为了清楚说明起见，在图2A中将各种总线都标为总线系统240。
处理器210可以是一种集成电路芯片，具有信号的处理能力，例如通用处理器、数字信号处理器(DSP，Digital Signal Processor)，或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等，其中，通用处理器可以是微处理器或者任何常规的处理器等。
存储器250可以是可移除的，不可移除的或其组合。示例性的硬件设备包括固态存储器，硬盘驱动器，光盘驱动器等。存储器250可选地包括在物理位置上远离处理器210的一个或多个存储设备。
存储器250包括易失性存储器或非易失性存储器，也可包括易失性和非易失性存储器两者。非易失性存储器可以是只读存储器(ROM，Read Only Memory)，易失性存储器可以是随机存取存储器(RAM，Random Access Memory)。本申请实施例描述的存储器250旨在包括任意适合类型的存储器。
在一些实施例中，存储器250能够存储数据以支持各种操作，这些数据的示例包括程序、模块和数据结构或者其子集或超集，下面示例性说明。
操作系统251，包括用于处理各种基本系统服务和执行硬件相关任务的系统程序，例如框架层、核心库层、驱动层等，用于实现各种基础业务以及处理基于硬件的任务；
网络通信模块252，用于经由一个或多个(有线或无线)网络接口220到达其他计算设备，示例性的网络接口220包括：蓝牙、无线相容性认证(WiFi)、 和通用串行总线(USB，Universal Serial Bus)等；
在一些实施例中，本申请实施例提供的基于人工智能的推荐模型训练装置可以采用软件方式实现，图2A示出了存储在存储器250中的基于人工智能的推荐模型训练装置255-A，其可以是程序和插件等形式的软件，包括以下软件模块：训练样本集合构建模块2556、模型构建模块2557、训练模块2558、以及模型上线模块2559，这些模块是逻辑上的，因此根据所实现的功能可以进行任意的组合或进一步拆分，将在下文中说明各个模块的功能。
在另一些实施例中，本申请实施例提供的基于人工智能的推荐模型训练装置可以采用硬件方式实现，作为示例，本申请实施例提供的基于人工智能的推荐模型训练装置可以是采用硬件译码处理器形式的处理器，其被编程以执行本申请实施例提供的基于人工智能的推荐模型训练方法，例如，硬件译码处理器形式的处理器可以采用一个或多个应用专用集成电路(ASIC，Application Specific Integrated Circuit)、DSP、可编程逻辑器件(PLD，Programmable Logic Device)、复杂可编程逻辑器件(CPLD，Complex Programmable Logic Device)、现场可编程门阵列(FPGA，Field-Programmable Gate Array)或其他电子元件。
将结合本申请实施例提供的服务器的示例性应用和实施，说明本申请实施例提供的基于人工智能的推荐模型训练方法。
下面分两个阶段来分别说明本申请实施例提供的基于人工智能的推荐方法以及基于人工智能的推荐模型训练方法，第一阶段是候选推荐信息预测模型的离线训练阶段。
参见图3，图3是本申请实施例提供的基于人工智能的推荐模型训练方法中的模型架构图。
这里的模型为候选推荐信息预测模型，离散向量形式的用户特征(在图3中以Field0-Field5表示)、离散向量形式的候选推荐信息特征(在图3中以Field5-Field12表示)以及离散向量形式的环境特征(在图3中以Field13-Field15表示)被转化为稠密嵌入向量的形式，将稠密嵌入向量的形式的用户特征、候选推荐信息特征以及环境特征输入到和池化层，将各个通道的特征图的所有像素值求和，这样每个通道得到一个实数值，N个通道最终会得到一个长度为N的向量，该向量即为和池化的结果，将池化结果输入到图3中示出的多层专家网络(包括专家网络1-8)和对应各个指标的权重生成器，权重生成器基于下述公式(1)对和池化的结果进行处理，得到针对对应指标的专家网络的权重分布：
g k(x)＝softmax(W gkx)                                          (1)
其中，g k(x)是对应第k个指标的各层专家网络输出的特征的权重，W gk是对应第k个指标的权重生成器的权重参数，x是对所有输入特征进行和池化的结果，图3中左侧的权重生成器对应的指标为指标1，这里的指标1可以为点击率，图3中右侧的权重生成器对应的指标为指标2，这里的指标2可以为时长，以左侧的权重生成器为例，其可以生成分别分配给专家网络1-8的8个权重，每个专家网络输出一个子特征，分别与对应的权重相乘再进行求和，得到 用于预测指标1的特征，详细过程可以参见公式(2)：
Figure PCTCN2020126549-appb-000001
其中，f k是对应第k个指标的多层专家网络的输出，g k(x) i是对应第k个指标的第i个专家网络输出的特征的权重，f i(x)是第i个专家网络输出的特征。
将对各层专家网络输出的特征加权求和得到的特征输入全连接层，全连接层由新型激活函数(Swish(256))构成，其原始公式参见公式(3)，变形公式参见公式(4)：
f(x)＝x*sigmoid(x)                                            (3)
f(x)＝x*sigmoid(β*x)                                         (4)
下面说明本申请实施例提供的Swish激活函数。参见图4，图4是本申请实施例提供的新型激活函数示意图，这里的Swish(256)激活函数拥有不饱和，光滑，非单调性的特征，当β(beta)为1时，激活函数具备无上界有下界、平滑且非单调的特性，改善了训练过程中发生梯度消失的现象，当β为0.1时，激活函数具备无上界有下界、平滑的特征，当β为10时，激活函数具备无上界有下界且非单调的特性。
继续说明图3，将全连接层输出的特征通过Sigmoid激活函数输出得到对应指标1的预测评分(图3中的点击率输出Pctr)，对于指标2而言，将全连接层输出的特征通过“lambda x：x”表达式，将输入值返回成输出值，得到对应指标2的预测评分(图3中的时长输出Pdur)，基于输入到全连接层的特征，得到最后对应各个指标的评分的过程可以通过公式(5)实现：
y k＝h k(f k(x))                                                 (5)
其中，f k是对应第k个指标的多层专家网络的输出，h k是对应第k个指标的塔(Tower)网络，塔网络即为图3中示出的全连接层。
参见图5A，图5A是本申请实施例提供的基于人工智能的推荐模型训练方法的一个可选的流程示意图，将结合图5A示出的步骤101-103进行说明。
在步骤101中，训练服务器对推荐系统的日志进行预处理，以构建训练样本集合。
这里的训练服务器是用于训练上述候选推荐信息预测模型的服务器，这里的日志包括推荐系统中用户行为数据，用户行为数据是与用户行为事件相关的数据，用户行为事件三要素包括：操作，定义一个操作动作(如点击、拖拽)；参数/属性，参数可以是任何和这个事件相关的属性，包括触发这个事件的(人、时间、地点、设备、操作的业务信息)；属性值，参数/属性的值，训练样本集合是对日志中的原始数据进行预处理得到的
参见图5C，基于图5A，图5C是本申请实施例提供的基于人工智能的推荐模型训练方法的一个可选的流程示意图，步骤101中通过对推荐系统的日志进行预处理，以构建训练样本集合可以通过步骤1011-步骤1013实现。
在步骤1011中，将推荐系统中的曝光日志、播放日志、资讯正排索引及画 像特征日志中的至少一种，按照用户设备标识进行融合处理，得到样本数据。
在一些实施例中，首先对埋点日志中的曝光、点击、转化和停留时长等数据做抽取解析，如基于曝光序列号关联各类操作、解析埋点参数(例如日志中记录的实时特征)、解析上下文特征等，然后对所得到的样本数据进行过滤，例如过滤恶意用户样本、过滤无效曝光样本等，例如，针对同一个候选推荐信息在不同时间对同一个用户曝光多次的情况，训练样本集合中会出现同一个用户对同一个候选推荐信息点击与不点击并存的情况，如果多次曝光的间隙非常短，考虑只使用其中的一次曝光数据，或者，为了避免高度活跃用户对损失函数的影响，对日志中每个用户提取相同数量的原始数据。
在步骤1012中，对所得到的样本数据进行正样本采样以及负样本采样，使得采样所得到的正样本以及负样本保持为标准比例。
在一些实施例中，对过滤后的样本数据做特征抽取，生成带特征的样本，主要从用户和候选推荐信息两个维度做特征工程，还可以从用户、候选推荐信息以及环境信息三个维度做特征工程，再按照一定正负样本比例进行正负样本采样，使得采样所得到的正样本和负样本保持为标准比例，标准比例是一个设定比例，可以根据推荐系统的先验数据确定。
这里的正样本和负样本是针对各个指标而言的，指标可以是点击率和时长，对于点击率任务而言，正样本可以是对应预测点击率的各种样本数据，且样本数据中的样本标签是高点击率(这里的高点击率是一个相对概念，可以预先设定高于点击率阈值即为高点击率)，负样本可以是对应预测点击率的各种样本数据，且样本数据中的样本标签是低点击率(这里的低点击率是一个相对概念，可以预先设定低于点击率阈值即为低点击率)，那么负样本的数量可能会很多，如果直接拿去训练是不合理的，所以需要保证正负样本的比例再进行训练。
申请人发现，原始数据可能会存在如下问题：抽样数量太少，数据量太小不能代表全体样本，抽样的数量一定要尽量的去代表全体数据；抽样偏差，抽样的数量要保证，同时质量也要保证；不满足模型需求的数据，有些算法对数据的平衡性非常的敏感，以及对特征分布或者特征分段要求很高，对于这样的模型，如果忽略注意正负样本之间的比例，就会导致性能显著下降；不满足业务需求的数据，在特定的业务场景下，正样本本来就很少，此时训练出来的参数表达能力有限，模型是否会不符合业务的召回率需求，解决样本不均衡问题的方法是进行采样处理，包括采样得到正样本和采样得到负样本两种，最终使得采样所得到的正样本以及负样本保持为设定比例。通过上述实施方式，解决了样本不均很的问题，使得模型符合业务的召回需求，推荐系统的推荐性能显著提高，从而有效提高推荐系统的通信资源以及计算资源的利用效率。
在步骤1013中，对经过采样得到的正样本和负样本进行特征提取处理，得到与各个指标匹配的特征，并将包括特征以及对应的指标的真实结果的集合，确定为训练候选推荐信息预测模型所需的训练样本集合。
在一些实施例中，以正样本为例，对正样本进行特征提取，得到对应正样本的至少一个特征，以及对应正样本的指标的真实结果(标签)，例如当指标是点击率时，这里的真实结果可以是高点击率，将特征以及对应指标的真实结果 作为一个训练样本。
在步骤102中，训练服务器基于与多个指标一一对应的权重生成器、与多个指标一一对应的预测器、以及多个专家网络，构建候选推荐信息预测模型。
参见图3，候选推荐信息预测模型中包括特征输入部分，特征按照区(field)进行类别区分，首先分为了用户特征、候选推荐信息特征以及环境特征，用户特征中又按照不同类型的特征通过区(field)进行区分，特征输入至和池化层，经过和池化层输入到多个专家网络(专家网络1-8)，候选推荐信息预测模型中还包括两个权重生成器，图3中左侧的权重生成器对应于指标1，图3中右侧的权重生成器对应指标2，候选推荐信息预测模型中还包括两个塔网络，每个塔网络包括由全连接层以及激活函数组成的预测器，每个预测器分别对应不用的指标(指标1和指标2)。
在步骤103中，训练服务器通过训练样本集合，对候选推荐信息预测模型进行多指标训练，其中，训练得到的候选推荐信息预测模型用于供推荐系统进行多指标的聚合处理和排序，以根据排序结果确定待推荐的候选推荐信息。
在一些实施例中，步骤103中通过训练样本集合，对候选推荐信息预测模型进行多指标训练，可以通过以下技术方案实现，初始化候选推荐信息预测模型，并初始化对应多个指标的损失函数，损失函数包括特征样本以及对应特征样本的评分；在候选推荐信息预测模型的每次迭代训练过程中执行以下处理：通过候选推荐信息预测模型，对训练样本集合包括的特征样本进行评分，得到对应特征样本的针对各个指标的评分；将对应特征样本的真实结果和评分代入损失函数，以确定损失函数取得最小值时对应的候选推荐信息预测模型参数；根据所确定的候选推荐信息预测模型参数更新候选推荐信息预测模型。
在一些实施例中，上述通过候选推荐信息预测模型，对训练样本集合包括的特征样本进行评分，得到对应特征样本的针对各个指标的评分的步骤，可以通过以下技术方案实现，通过候选推荐信息预测模型中的多个专家网络，将特征样本分别映射到对应各个专家网络的特征空间，以获得各个专家网络所关联的特征维度的子特征；基于子特征权重分布中对应各个专家网络所获得的子特征的权重，对各个专家网络所获得关联的子特征进行加权处理，分别得到对应特征样本的指标特征；通过候选推荐信息预测模型中包括的对应各个指标的预测器，结合指标特征对特征样本进行评分预测处理，得到特征样本基于各个指标的评分。
在一些实施例中，特征样本包括：用户特征、候选推荐信息特征以及环境特征，将特征样本输入到多层专家网络和对应各个指标的权重生成器，权重生成器基于上述公式(1)对和池化的结果进行处理，得到针对对应指标的专家网络的权重分布。
图3中左侧的权重生成器对应的指标为指标1，指标1可以为点击率，图3中右侧的权重生成器对应的指标为指标2，指标2可以为时长，以图3中左侧的权重生成器为例，其可以生成分别分配给专家网络1-8的8个权重，每个专家网络输出一个子特征，分别与对应的权重相乘再进行求和，得到用于预测指标1的指标特征，详细过程可以参见公式(2)。将对各层专家网络输出的特征 加权求和得到的指标特征输入到预测器(塔网络)中的全连接层，全连接层由新型激活函数(Swish(256))构成，其原始公式参见公式(3)，变形公式参见公式(4)。
在一些实施例中，上述将对应特征样本的真实结果和评分代入损失函数，以确定损失函数取得最小值时对应的候选推荐信息预测模型参数的步骤，可以通过以下技术方案实现，将对应特征样本的针对各个指标的真实结果和评分分别代入对应各个指标的损失函数；结合对应各个损失函数的损失权重，对各个损失函数进行加权求和处理，得到对应候选推荐信息预测模型的聚合损失函数；对聚合损失函数进行最小化处理，得到聚合损失函数取得最小值时对应的候选推荐信息预测模型参数；其中，候选推荐信息预测模型参数包括对应候选推荐信息模型的结构参数、以及对应聚合损失函数的损失权重。
在一些实施例中，聚合损失函数为：
Figure PCTCN2020126549-appb-000002
其中，
Figure PCTCN2020126549-appb-000003
是对应连续型指标中的第i指标的第i损失函数，
Figure PCTCN2020126549-appb-000004
是对应离散型指标中的第i指标的第i损失函数，
Figure PCTCN2020126549-appb-000005
是对应连续型指标中的第i损失函数的权重，
Figure PCTCN2020126549-appb-000006
用于表征对连续型指标中的第i指标进行预测评分的不确定性，
Figure PCTCN2020126549-appb-000007
是对应离散型指标中的第i损失函数的权重，
Figure PCTCN2020126549-appb-000008
用于表征对离散型指标中的第i指标进行预测评分的不确定性，W是候选推荐信息预测模型中对应各个指标的结构权重；对应离散型指标的第i损失函数的表达式为：
Figure PCTCN2020126549-appb-000009
其中，α t为用于平衡正负样本的平衡因子常数，γ为用于调节简单样本权重降低速率的调节常数，P t为对应评分的概率值。
在一些实施例中，由于输出的不同，对应各指标的损失函数的权重也有所不同，当第一指标的输出为离散型输出，且第二指标的输出也为离散型输出时，聚合损失函数为：
Figure PCTCN2020126549-appb-000010
其中，L 1(W)是对应第一指标的第一损失函数，L 2(W)是对应第二指标的第二损失函数，
Figure PCTCN2020126549-appb-000011
是对应第一损失函数的权重，σ 1用于表征对第一指标进行预测评分的不确定性，
Figure PCTCN2020126549-appb-000012
是对应第二损失函数的权重，σ 2用于表征对第二指标进行预测评分的不确定性。
在一些实施例中，当第一指标的输出为连续型输出，且第二指标的输出为离散型输出时，聚合损失函数为：
Figure PCTCN2020126549-appb-000013
其中，L 1(W)是对应第一指标的第一损失函数，L 2(W)是对应第二指标的第二损失函数，
Figure PCTCN2020126549-appb-000014
是对应第一损失函数的权重，σ 1用于表征对第一指标进行预测评分的不确定性，
Figure PCTCN2020126549-appb-000015
是对应第二损失函数的权重，σ 2用于表征对第二指标进行预测评分的不确定性。
在一些实施例中，在执行完步骤103之后，还可以执行步骤104-105，参见图5B，图5B是本申请实施例提供的基于人工智能的推荐模型训练方法的一个可选的流程示意图。
在步骤104中，训练服务器将候选推荐信息预测模型的参数转换为常量，并固化在候选推荐信息预测模型中，以生成固化后的二进制模型文件。
在步骤105中，训练服务器将二进制模型文件推送至推荐系统，以使推荐系统中所使用的候选推荐信息预测模型与经过训练得到的候选推荐信息预测模型结构一致。
在一些实施例中，通过将模型训练参数数据与网络结构相融合的方式，通过将模型参数转换为常量的方式将参数固化在模型网络结构当中，以保证离线训练模型与在线预测模型网络结构的一致性，在线推荐系统通过加载固化后的模型文件，即可同时获取网络结构以及模型训练参数，进而保证一致性，将最终得到的二进制模型文件通过定时任务按照预设的频次，按天或者按小时推送至在线推荐系统，同时将文件内容生成信息摘要算法进行编码，以用于后续模型校验，至此完成整个离线训练过程。
第二阶段是候选推荐信息预测模型的应用阶段，应用于应用服务器200-B。
参见图1B，图1B是本申请实施例提供的基于人工智能的推荐系统100-B的一个可选的架构示意图，终端400通过网络300连接应用服务器200-B以及训练服务器200-A，网络300可以是广域网或者局域网，又或者是二者的组合，应用服务器200-B负责通过候选推荐信息预测模型对候选推荐信息进行排序，应用服务器200-B中包括信息获取模块2551、特征形成模块2552、特征处理模块2553、信息排序模块2554以及推荐模块2555，应用服务器200-B响应于接收到来自于终端400的用户请求，信息获取模块2551从召回系统500中获取对应用户请求的候选推荐信息，应用服务器200-B通过从训练服务器200-A获取到的训练好的候选推荐信息预测模型对候选推荐信息进行排序，并根据排序结果将对应排序结果的候选推荐信息推送至用户所使用的终端400。
在一些实施例中，应用服务器200-B可以是独立的物理服务器，也可以是多个物理服务器构成的服务器集群或者分布式系统，还可以是提供云服务、云数据库、云计算、云函数、云存储、网络服务、云通信、中间件服务、域名服务、安全服务、CDN、以及大数据和人工智能平台等基础云计算服务的云服务器。终端400可以是智能手机、平板电脑、笔记本电脑、台式计算机、智能音箱、智能手表等，但并不局限于此。终端以及服务器可以通过有线或无线通信 方式进行直接或间接地连接，本发明实施例中不做限制。
参见图2B，图2B是本申请实施例提供的应用基于人工智能的推荐方法的应用服务器200-B的结构示意图，图2B所示的应用服务器200-B的结构与训练服务器200-A中的结构相同，除了应用服务器200-B中包括的是基于人工智能的推荐装置255-B而不是基于人工智能的推荐模型训练装置255-A。
在一些实施例中，本申请实施例提供的基于人工智能的推荐装置255-B可以采用软件方式实现，图2B示出了存储在存储器250中的基于人工智能的推荐装置255-B，其可以是程序和插件等形式的软件，包括以下软件模块：信息获取模块2551、特征形成模块2552、特征处理模块2553、信息排序模块2554以及推荐模块2555，这些模块是逻辑上的，因此根据所实现的功能可以进行任意的组合或进一步拆分，将在下文中说明各个模块的功能。
在另一些实施例中，本申请实施例提供的基于人工智能的推荐装置可以采用硬件方式实现，作为示例，本申请实施例提供的基于人工智能的推荐装置可以是采用硬件译码处理器形式的处理器，其被编程以执行本申请实施例提供的基于人工智能的推荐方法，例如，硬件译码处理器形式的处理器可以采用一个或多个应用专用集成电路(ASIC，Application Specific Integrated Circuit)、DSP、可编程逻辑器件(PLD，Programmable Logic Device)、复杂可编程逻辑器件(CPLD，Complex Programmable Logic Device)、现场可编程门阵列(FPGA，Field-Programmable Gate Array)或其他电子元件。
参见图6A，图6A是本申请实施例提供的基于人工智能的推荐方法的一个可选的流程示意图，将结合图6A示出的步骤201-203进行说明。
在步骤201中，应用服务器获取对应待推荐用户的多个候选推荐信息。
在一些实施例中，步骤201中获取对应待推荐用户的多个候选推荐信息，可以通过以下技术方案实现，获取以下至少一个类型的候选推荐信息：与对应待推荐用户的历史浏览信息的内容相似，且内容相似度不小于内容相似度阈值的多个候选推荐信息；与对应待推荐用户的历史行为信息的行为相似，且行为相似度不小于行为相似度阈值的多个候选推荐信息。
在一些实施例中，可以通过推荐系统中的召回模块获取多个候选推荐信息，这里获取多个候选推荐信息是响应于待推荐用户的用户请求，这里的用户请求可以是携带有特定目标的查询请求，还可以是初始化应用的请求，这里的行为相似度指的是用户历史行为信息与候选推荐信息之间的相似度，内容相似度指的是用户历史浏览信息与候选推荐信息之间的相似度。
在步骤202中，应用服务器获取待推荐用户的用户特征、以及每个候选推荐信息的候选推荐信息特征，并将用户特征分别与每个候选推荐信息的候选推荐信息特征进行组合，以形成对应每个候选推荐信息的融合特征。
在一些实施例中，步骤202中获取待推荐用户的用户特征、以及每个候选推荐信息的候选推荐信息特征，可以通过以下技术方案实现，获取对应待推荐用户的以下至少之一的用户特征：用于表征待推荐用户的基本信息的基础属性特征；用于表征用户社会关系的社会关系特征；用于表征用户互动行为的互动行为特征；用于表征用户阅读偏好的阅读心理特征；获取对应候选推荐信息的 以下至少之一的候选推荐信息特征：用于表征候选推荐信息类别的类别特征；用于表征候选推荐信息内容的标签特征；用于表征候选推荐信息发布时间的时间特征；用于表征候选推荐信息来源的发布特征；用于表征候选推荐信息长度的长度特征。
在一些实施例中，基本信息可以是用户的性别、年龄、长期居住地等基本属性，社会关系可以是是否婚配或者工作岗位等等具有社会属性，用户互动行为可以是点赞、转发或者收藏等等行为，阅读偏好可以是阅读兴趣，兴趣点可以是娱乐八卦或是国际新闻等等，候选推荐信息类别可以是信息展示载体类别，例如，视频信息、图像信息或者文本信息，候选推荐信息内容可以是内容主题，例如教育话题或者娱乐话题等等。
在一些实施例中，步骤202中将用户特征分别与每个候选推荐信息的候选推荐信息特征进行组合，以形成对应每个候选推荐信息的融合特征，可以通过以下技术方案实现，获取对应待推荐用户的以下至少之一的环境特征：向待推荐用户进行推送的时间特征；待推荐用户的用户位置特征；待推荐用户的设备特征；待推荐用户所使用的设备所处的网络特征；将对应待推荐用户的环境特征、用户特征以及候选推荐信息特征，组合成对应每个候选推荐信息的融合特征。
在一些实施例中，环境特征对于指标预测也有影响，向待推荐用户进行推送的时间会影响到待推荐用户是否有空闲观看候选推荐信息，待推荐用户的位置表征待推荐用户当前所处的生活场景，不同的生活场景对于指标的预测是有影响的，例如，当位置特征表征待推荐用户处在电影院，对于时长和点击率的预测结果和位置特征表征待推荐用户处在自习室的预测结果会有较大差异，待推荐用户所使用的设备所处的网络会影响待推荐用户是否希望接收到视频等需要耗费较大网络资源的候选推荐信息。
在步骤203中，应用服务器对对应每个候选推荐信息的融合特征进行多层次映射处理，得到每个候选推荐信息分别对应多个指标的评分。
参见图6B，基于图6A，图6B是本申请实施例提供的基于人工智能的推荐方法的一个可选的流程示意图，步骤203中对对应每个候选推荐信息的融合特征进行多层次映射处理，得到每个候选推荐信息分别对应多个指标的评分可以通过步骤2031-步骤2032实现。
在步骤2031中，通过多个专家网络，对对应每个候选推荐信息的融合特征进行多层次映射处理，得到与多个专家网络一一对应的多个子特征，并对多个子特征进行加权处理，得到对应各个指标的指标特征。
参见图6C，基于图6A，图6C是本申请实施例提供的基于人工智能的推荐方法的一个可选的流程示意图，步骤2031中通过多个专家网络，对对应每个候选推荐信息的融合特征进行多层次映射处理，得到与多个专家网络一一对应的多个子特征，并对多个子特征进行加权处理，得到对应各个指标的指标特征可以通过步骤20311-步骤20313实现。
在步骤20311中，通过候选推荐信息预测模型中的多个专家网络，将融合特征分别映射到对应各个专家网络的特征空间，以获得各个专家网络所关联的 特征维度的子特征。
作为示例，各层专家网络可以看成是多个全连接层，每个专家网络通常是数层规模比较小的全连接层。
在步骤20312中，通过候选推荐信息预测模型中与多个指标一一的对应的权重生成器，对融合特征进行最大似然处理，以得到对应各个指标的子特征权重分布。
权重生成器用来选择每个专家网络的信号占比，每个专家网络都有其擅长的预测方向，最后共同作用于分别对应各个权重生成器的指标，这里的各个指的是多个权重生成器中的每一个，关于各个的含义在全文中均表示多个对象中的每一个。
在步骤20313中，基于子特征权重分布中对应各个专家网络所获得的子特征的权重，对各个专家网络所获得关联的子特征进行加权处理，分别得到对应各个指标的指标特征。
在步骤2032中，通过候选推荐信息预测模型中与多个指标一一对应的预测器，并结合对应各个指标的指标特征对融合特征进行评分预测处理，得到融合特征基于各个指标的评分。
作为示例，将每一个权重生成器认为是加权求和池化操作，如果将权重生成器换成选择最大的操作，x为输入，则各层专家网络输出子特征分量最大对应的专家网络被唯一选中，向上传递信号。
在步骤204中，应用服务器对分别对应多个指标的评分进行多指标聚合处理，得到每个候选推荐信息的综合评分，并根据每个候选推荐信息的综合评分对多个候选推荐信息进行降序排序。
在一些实施例中，步骤204中对分别对应多个指标的评分进行多指标聚合处理，得到每个候选推荐信息的综合评分，可以通过以下技术方案实现，获取对应多个指标的聚合规则；基于聚合规则所包括的算子，对分别对应多个指标的评分进行计算处理，得到每个候选推荐信息对应多个指标的综合评分。
在一些实施例中，多个指标的聚合规则可以对应不同多指标聚合处理的方式，这里的多指标可以是时长、点击率、收藏数、转发数等等指标中的至少两个，聚合规则可以是带参数的相加、相乘或者其他运算规则，这里可以将时长以及点击率这两个指标进行相乘作为多指标聚合处理方式。
在步骤205中，应用服务器在降序排序的结果中选择排序靠前的至少一个候选推荐信息，并基于所选择的候选推荐信息执行对应待推荐用户的推荐操作。
在一些实施例中，在降序排序的结果中选择排序靠前的N个候选推荐信息作为即将要推荐给待推荐用户的候选推荐信息，其中，N为正整数，并基于所选择的候选推荐信息执行对应待推荐用户的推荐操作，推荐操作可以是直接推送至待推荐用户，还可以基于所选择的候选推荐信息进行重排序，这里的重排序指的是按照不同的多指标聚合处理的方式对所选择的候选推荐信息进行重新排序，以从更加全面的角度获取受待推荐用户喜欢的推荐信息。
在一些实施例中，基于所选择的候选推荐信息执行对应待推荐用户的推荐操作，这里所产生的作为推荐结果的候选推荐信息，是基于多指标预测所得到 的结果，而不是单独某一项指标预测所产生的结果，因此推荐准确率较高，容易引起待推荐用户的兴趣，待推荐用户会阅读、点击、收藏以及转发作为推荐结果的候选推荐信息，从而提升了互动行为比例，待推荐用户会阅读、点击、收藏以及转发的操作会作为用户画像的原始信息进行存档保留，并用于按时间间隔对模型进行训练，以持续性提升模型预测性能。
下面，将说明本申请实施例在一个实际的应用场景中的示例性应用，参见图7，图7是本申请实施例提供的基于人工智能的推荐方法的应用场景产品示意图，该界面中展示了三个经过排序处理的候选推荐信息，这些经过排序处理的候选推荐信息是通过候选推荐信息预测模型进行排序得到的，应用产品可以是新闻客户端，通过有效利用用户历史的隐式及显式行为，基于多层次改进的多任务学习训练方法训练推荐系统中的候选推荐信息预测模型，同时对在线推荐系统的排序目标做可配置性的设计，为用户提供精准的个性化新闻资讯推荐。
参加图8，图8是本申请实施例提供的候选推荐信息预测模型训练及应用架构图，如图8所示，图8所示的架构图包括离线模型训练部分和在线推荐系统部分，在接收到用户请求后，可以返回个性化的推荐信息列表，与推荐服务中控具有直接交互的单元(1)-(6)均为在线服务，(7)为离线部分，下面分别进行说明。
通过业务接入部分(1)接收用户请求，获取待推荐用户的相关特征，这里的相关特征可以是用户曝光、用户反馈以及浏览时长，通过召回部分(3)得到响应于用户请求的个性化的候选推荐信息，通过在线排序部分(4)对获取的候选推荐信息进行排序，通过重排序部分(5)对经过排序得到的候选推荐信息进行基于不同策略的重排序，通过用户画像部分(2)获取用户画像，通过特征获取部分(6)进行特征提取，通过离线模型训练部分(7)对模型进行离线训练。
业务接入部分(1)：接收用户的客户端发送的用户请求，基于用户请求可以获取到用户的用户特征，这里的用户特征可以包括偏好特征、行为特征等等，推荐服务中控响应于用户请求，从召回部分获取与用户请求匹配的候选推荐信息，并通过本申请实施例提供的候选推荐信息预测模型对召回部分获取到的候选推荐信息进行基于点击率的多指标聚合排序，并将推荐结果通过业务接入部分返回给客户端，例如，在图7中展示了返回给客户端的推荐结果，这里所展示的三条候选推荐信息均是符合用户特征的推荐结果，通过业务接入部分可以获取用户曝光数据、用户反馈数据以及浏览时长数据。
召回部分(3)：从全量信息集合中触发尽可能多的正确结果，并将结果返给在线排序部分(4)，召回的方式有多种，有协同过滤，主题模型、内容召回以及热点召回，在推荐系统中，用户不会提供明确的检索词输入，因此推荐系统需要根据用户画像、内容画像等各种信息为用户推荐可能感兴趣的内容。
在在线排序部分(4)中，针对任意一个推荐任务，通过特征生成器获取推荐任务的特征，并基于得到的特征以及候选推荐信息预测模型预测每个候选推荐信息的点击率(点击率结果)。
在特征获取部分(6)中，特征获取的来源可以是用户日志、新闻细节、用 户会话以及实时特征。
重排序部分(5)：排序和重排序之间的区别在于，排序所得到的是某种聚合规则下的结果，但是单一一种聚合规则存在局限性，因此需要进行重排序，在重排序的过程中，采用与排序所基于的聚合规则不同的聚合规则进行排序，例如，在排序阶段是通过点击率和时长的乘积作为聚合方式，而重排序阶段采用转发率与时长的乘积作为聚合方式，来对经过排序的候选推荐信息进行重排序。
接下来介绍离线模型训练部分(7)，离线模型训练部分可包括：训练样本集合构建模块、模型构建模块、训练模块、以及模型上线模块，训练样本集合构建模块中包括训练库生成单元、采样单元、特征生成单元、训练模块中包括模型训练单元、模型上线模块中包括：固化图单元以及模型推送单元，训练库生成单元用于对原始数据进行有效预处理，是提高训练模型准确度的前提，将用户的曝光日志、播放日志、资讯正排索引及画像特征日志，按照用户设备号进行融合，同时删除缺失有效单特征的数据，保证每条样本的有效单特性信息无缺失，采样单元用于基于训练库生成单元中得到的原始训练库对样本进行正负样本采样，使得正负样本保持在合理的比例(设定比例)，同时过滤播放时长在一定阈值之下的样本，并针对不同信息时长和观看时长采取分段门限正负样本设定方案，特征生成单元(特征生成器)用于将采样单元得到的样本信息做进一步提取处理，结合具体的预测目标选用适当的特征，从而生成离线模型训练所需的样本数据(特征与标签)，模型训练单元用于利用特征生成单元生成的样本数据，基于多层次改进的多任务学习模型进行离线模型训练，候选推荐信息预测模型中采用多层专家网络，多层专家网络促使多任务更好的进行底层特征共享，每一个任务独立的门限网络用来决定不同专家网络结果的使用程度，专家网络和门限网络均为三维张量，专家网络可以是简单的全连接层结构，候选推荐信息预测模型是基于以下公式输出各个指标的评分：
y k＝h k(f k(x))                                                 (5)
Figure PCTCN2020126549-appb-000016
g k(x)＝softmax(W gkx)                                           (1)
其中，y k是对应第k个指标的输出，h k是对应第k个指标的预测器的网络，f k是对应第k个指标的多层专家网络的输出，g k(x) i是对应第k个指标的第i个专家网络输出的特征的权重，f i(x)是第i个专家网络输出的特征，W gk是对应第k个指标的门限网络的权重参数，x是输入的特征，训练上述候选推荐信息预测模型时引入多指标不确定性因子，相关技术中聚合多任务的损失函数，采用网格搜索的方式，参见公式(10)：
Figure PCTCN2020126549-appb-000017
其中，w i是对应第i个指标的损失函数的权重，L i是对应第i个指标的损失函数，L total是聚合多任务的损失函数，虽然网络搜索的方式简单有效，但权值调节较耗时，因此，引入不确定性来衡量多任务间损失函数的权重。
在很多情况下，深度学习技术在各个领域都有极佳的表现，这种表现依赖于强大的算力和深厚的网络结构，会对问题给出一个特定的结果，大多数情况下，深度学习会对问题给出一个答案，且模型不会给出最终输出的结果的置信度，这个答案是模型在众多候选中找到概率最大的，但是在极端情况下，如分类标签是A和B，但是在测试阶段输入C类的图像，那么分类器大概率会带来无法预知的结果，通过贝叶斯建模实现置信度的输出，在贝叶斯模型中，有两种主要类型的不确定性可以建模：认知不确定性，是模型中的固有不确定性，由于训练数据量的不足，导致对于模型没见过的数据会有很低的置信度，认知不确定性解释了模型参数的不确定性，认知不确定性可以通过增加训练数据消除；偶然不确定性，在数据标注时如果出现比较大的标注误差，这个误差不是模型带入的，而是数据本身就存在的，数据集里的偏置越大，偶然不确定性就越大，其中，偶然不确定性可以细分为两类：(1)数据依赖型或异方差不确定性，这种不确定性取决于输入的数据，并且将预测结果作为模型的输出；(2)任务依赖型或同方差不确定性，不依赖于输入数据，也不会是模型输出结果，而是对所有输入数据相同的常量，对不同任务不同的变量，基于这个特性，叫做任务依赖型不确定性，因为在多任务学习中，任务的不确定性表明了相对置信度，反映了回归和分类问题中固有的不确定性，因此把同方差不确定性作为噪声来对多任务学习中的权重进行优化，多任务损失函数，这个损失函数利用同方差不确定性来最大化高斯似然估计，首先提出一个对于回归问题的概率模型定义，参见以下公式(11)：
p(y|f W(x))＝N(f W(x),σ 2)                                        (11)
其中，f W是神经网络的输出，x是输入数据，W是权重，对于分类问题，通常会将输出压入激活函数中，参见以下公式：
Figure PCTCN2020126549-appb-000018
接下来定义多任务的似然函数，参见以下公式：
p(y 1,...,y k|f W(x))＝p(y 1|f W(x))...p(y k|f W(x))                          (13)
其中，y 1是多任务中每个子任务的输出，因而，极大似然估计就可以通过以下公式(14)表明，该极大似然估计与公式(14)中的范数成正比，其中，σ是高斯分布的标准差，也是模型的噪声，接下来的任务就是根据W和σ最大化似然分布：
Figure PCTCN2020126549-appb-000019
以两个任务为例，即存在两个输出，当两个输出分别为连续型输出和独立型输出时，分别利用高斯分布和最大似然分布进行建模，可以得到连续型输出和独立型输出的多任务聚合损失函数，参见以下公式：
Figure PCTCN2020126549-appb-000020
其中，L 1(W)是对应第一指标的第一损失函数，L 2(W)是对应第二指标的第二损失函数，
Figure PCTCN2020126549-appb-000021
是对应第一损失函数的权重，σ 1用于表征对第一指标进行预测评分的不确定性，
Figure PCTCN2020126549-appb-000022
是对应第二损失函数的权重，σ 2用于表征对第二指标进行预测评分的不确定性，W是候选推荐信息预测模型中对应各个指标的结构权重。
训练的任务是最小化这个极大似然估计，所以，当σ(噪声)增大时，相对应的权重就会降低，另一方面，随着噪声σ减小，相对应的权重就要增加。
同时，针对于独立型输出的任务，还需要解决分类不平衡的问题，因此针对独立型输出的损失函数还需要在二分类交叉熵损失函数的基础上进行修改，修改后的损失函数如上述公式(7)所示：
L 2(W)＝-α t(1-P t) γlog(P t)                                         (7)
其中，α t为用于平衡正负样本的平衡因子常数，γ为用于调节简单样本权重降低速率的调节常数，P t为对应评分的概率值，让模型自适应的去学习艰难样本，同时平衡正负样本比例，在原有交叉熵损失函数的基础上加了一个γ因子，其中，γ>0能够减少易分类样本的损失，让模型自适应的去学习艰难样本，更加关注于这种难以区分的样本，减少了简单样本的影响，还加入平衡因子α t，用来平衡正负样本本身的比例不均，除了对交叉熵损失函数进行改进之外，还对网络激活函数进行调优，在常见的激活函数进行了改进，使该激活函数具备无上界有下界、平滑且非单调的特性，参见公式：
f(x)＝x*sigmoid(β*x)                                          (4)
其中，β为正数，优选为1，当β为1时，激活函数具备无上界有下界、平滑且非单调的特性，改善了训练过程中发生梯度消失的现象。
离线模型训练部分还包括固化图单元，通过将模型训练参数数据与网络结构相融合的方式，将模型的参数转换为常量，从而将参数固化在模型网络结构当中，以保证离线训练模型与在线预测模型网络结构的一致性，在线推荐系统通过加载固化后的模型文件，即可同时获取网络结构以及模型训练参数，进而保证一致性，离线模型训练部分还包括模型推送单元，将最终得到的二进制模型文件通过定时任务按照预设的频次，按天或者按小时推送至在线推荐系统，同时将文件内容生成信息摘要算法进行编码，以用于后续模型校验，至此完成整个离线流程。
下面对在线排序部分(4)进行说明，在线排序部分(4)中还包括特征生 成单元(特征生成器)及在线模型预测部分等等，在线流程中特征生成单元与离线流程中的特征生成单元的设计方式一致，服务器接收用户请求后，获取用户的相关特征，同时通过召回部分得到每个用户请求与之匹配的候选推荐信息，拉取每个候选推荐信息的候选推荐信息特征，与用户特征结合得到与离线流程中特征生成单元中一致的样本格式，在线模型预测部分，根据离线训练模型的改进多任务模型网络结构，在线预测时对多指标进行预测(例如点击率、时长、收藏率、点赞率等)，同时设计多种目标间的聚合排序方式，例如点击率*时长，作为排序目标值，选取排序靠前的N项做为最终返回推荐结果，上述改进优化了候选推荐信息预测模型的细节，且加强了模型抗噪声能力，进而显著提升线上推荐系统的关键指标，模型的测试结果参见表1和表2，上述实验表明联合学习点击率和时长能够获得更好的泛化能力，对线上推荐的各项核心指标正向带动明显，用户互动显式行为比例也有了明显提升：
模型 AUC Log(损失函数) 均方误差
Pctr(单点击率) 0.7045 0.3565 NA
Pctr(多指标) 0.7114 0.3302 NA
Pdur(单时长) 0.6952 NA 2623.2
Pdur(多指标) 0.7030 NA 2607.8
Pctr*Pdur(多指标) 0.7121 0.3302 2607.8
表1：对模型进行联合学习的指标测试结果
模型 人均时长 人均点击 点击率
深度因子分解机 +1.5％ +2.4％ 16.9％
多任务神经网络 +3.1％ +2.7％ 17.7％
具有多专家模型的多任务神经网络 +3.5％ +2.9％ 19.3％
改进的多任务神经网络 +3.8％ +3.4％ 20.1％
表2：各个模型的指标测试结果
本申请实施例提供了一种基于人工智能的推荐方法，是一种基于多层次改进的多任务学习推荐方法，分别通过设计多层专家网络、引入多指标不确定性因子、解决分类不平衡问题等角度，对个性化推荐系统的排序模型进行了深度优化改造，模型离线评测AUC、均方误差等指标有了明显提升，同时对在线推荐系统的排序目标做可配置性的设计(点击率、时长、点赞率、分享率等)，更加灵活可控，本方案已经成功应用于新闻应用的推荐系统，最终对线上推荐的各项核心指标正向带动明显，用户互动显式行为比例也有了明显提升。
下面继续说明本申请实施例提供的基于人工智能的推荐装置255的实施为软件模块的示例性结构，在一些实施例中，如图2B所示，存储在存储器250的基于人工智能的推荐装置255-B中的软件模块可以包括：信息获取模块2551，配置为获取对应待推荐用户的多个候选推荐信息；特征形成模块2552，配置为获取待推荐用户的用户特征、以及每个候选推荐信息的候选推荐信息特征，并将用户特征分别与每个候选推荐信息的候选推荐信息特征进行组合，以形成对应每个候选推荐信息的融合特征；特征处理模块2553，配置为对对应每个候选推荐信息的融合特征进行多层次映射处理，得到每个候选推荐信息分别对应多个指标的评分；信息排序模块2554，配置为对分别对应多个指标的评分进行多指标聚合处理，得到每个候选推荐信息的综合评分，并根据每个候选推荐信息的综合评分对多个候选推荐信息进行降序排序；推荐模块2555，配置为在降序排序的结果中选择排序靠前的至少一个候选推荐信息，并基于所选择的候选推荐信息执行对应待推荐用户的推荐操作。
在一些实施例中，信息获取模块2551，还配置为：获取以下至少一个类型的候选推荐信息：与对应待推荐用户的历史浏览信息的内容相似，且内容相似度不小于内容相似度阈值的多个候选推荐信息；与对应待推荐用户的历史行为信息的行为相似，且行为相似度不小于行为相似度阈值的多个候选推荐信息。
在一些实施例中，特征形成模块2552，还配置为：获取对应待推荐用户的以下至少之一的用户特征：用于表征待推荐用户的基本信息的基础属性特征；用于表征用户社会关系的社会关系特征；用于表征用户互动行为的互动行为特征；用于表征用户阅读偏好的阅读心理特征；获取对应候选推荐信息的以下至少之一的候选推荐信息特征：用于表征候选推荐信息类别的类别特征；用于表征候选推荐信息内容的标签特征；用于表征候选推荐信息发布时间的时间特征；用于表征候选推荐信息来源的发布特征；用于表征候选推荐信息长度的长度特征。
在一些实施例中，特征形成模块2552，还配置为：获取对应待推荐用户的以下至少之一的环境特征：向待推荐用户进行推送的时间特征；待推荐用户的用户位置特征；待推荐用户的设备特征；待推荐用户所使用的设备所处的网络特征；将对应待推荐用户的环境特征、用户特征以及候选推荐信息特征，组合成对应每个候选推荐信息的融合特征。
在一些实施例中，特征处理模块2553，还配置为：通过多个专家网络，对对应每个候选推荐信息的融合特征进行多层次映射处理，得到与多个专家网络一一对应的多个子特征，并对多个子特征进行加权处理，得到对应各个指标的指标特征；通过候选推荐信息预测模型中与多个指标一一对应的预测器，并结合对应各个指标的指标特征对融合特征进行评分预测处理，得到融合特征基于各个指标的评分。
在一些实施例中，特征处理模块2553，还配置为：通过候选推荐信息预测模型中的多个专家网络，将融合特征分别映射到对应各个专家网络的特征空间，以获得各个专家网络所关联的特征维度的子特征；
通过候选推荐信息预测模型中与多个指标一一的对应的权重生成器，对融合特征进行最大似然处理，以得到对应各个指标的子特征权重分布；基于子特征权重分布中对应各个专家网络所获得的子特征的权重，对各个专家网络所获得关联的子特征进行加权处理，分别得到对应各个指标的指标特征。
在一些实施例中，信息排序模块2554，还配置为：获取对应多个指标的聚合规则；基于聚合规则所包括的算子，对分别对应多个指标的评分进行计算处理，得到每个候选推荐信息对应多个指标的综合评分。
在一些实施例中，如图2A所示，存储在存储器250的基于人工智能的推荐模型训练装置255-A中的软件模块可以包括：训练样本集合构建模块2556，配置为对推荐系统的日志进行预处理，以构建训练样本集合；模型构建模块2557，配置为基于与多个指标一一对应的权重生成器、与多个指标一一对应的预测器、以及多个专家网络，构建候选推荐信息预测模型；训练模块2558，配置为通过训练样本集合，对候选推荐信息预测模型进行多指标训练；其中，训练得到的候选推荐信息预测模型用于供推荐系统进行多指标的聚合处理和排序，以根据排序结果确定待推荐的候选推荐信息。
在一些实施例中，装置255-A还包括：模型上线模块2559，配置为：将候选推荐信息预测模型的参数转换为常量，并固化在候选推荐信息预测模型中，以生成固化后的二进制模型文件；将二进制模型文件推送至推荐系统，以使推荐系统中所使用的候选推荐信息预测模型与经过训练得到的候选推荐信息预测模型结构一致。
在一些实施例中，训练样本集合构建模块2556，还配置为：将推荐系统中的曝光日志、播放日志、资讯正排索引及画像特征日志中的至少一种，按照用户设备标识进行融合处理，得到样本数据；对所得到的样本数据进行正样本采样以及负样本采样，使得采样所得到的正样本以及负样本保持为标准比例；对经过采样得到的正样本和负样本进行特征提取处理，得到与各个指标匹配的特征，并将包括特征以及对应的指标的真实结果的集合，确定为训练候选推荐信息预测模型所需的训练样本集合。
在一些实施例中，训练模块2558，还配置为：初始化候选推荐信息预测模型，并初始化对应多个指标的损失函数，损失函数包括特征样本以及对应特征样本的评分；在候选推荐信息预测模型的每次迭代训练过程中执行以下处理：通过候选推荐信息预测模型，对训练样本集合包括的特征样本进行评分，得到对应特征样本的针对各个指标的评分；将对应特征样本的真实结果和评分代入损失函数，以确定损失函数取得最小值时对应的候选推荐信息预测模型参数；根据所确定的候选推荐信息预测模型参数更新候选推荐信息预测模型。
在一些实施例中，训练模块2558，还配置为：通过候选推荐信息预测模型中的多个专家网络，将特征样本分别映射到对应各个专家网络的特征空间，以获得各个专家网络所关联的特征维度的子特征；基于子特征权重分布中对应各个专家网络所获得的子特征的权重，对各个专家网络所获得关联的子特征进行加权处理，分别得到对应特征样本的指标特征；通过候选推荐信息预测模型中包括的对应各个指标的预测器，结合指标特征对特征样本进行评分预测处理， 得到特征样本基于各个指标的评分。
在一些实施例中，训练模块2558，还配置为：将对应特征样本的针对各个指标的真实结果和评分分别代入对应各个指标的损失函数；结合对应各个损失函数的损失权重，对各个损失函数进行加权求和处理，得到对应候选推荐信息预测模型的聚合损失函数；对聚合损失函数进行最小化处理，得到聚合损失函数取得最小值时对应的候选推荐信息预测模型参数；其中，候选推荐信息预测模型参数包括对应候选推荐信息模型的结构参数、以及对应聚合损失函数的损失权重。
本申请实施例提供一种存储有可执行指令的存储介质，其中存储有可执行指令，当可执行指令被处理器执行时，将引起处理器执行本申请实施例提供的基于人工智能的推荐方法，例如，如图5A-5C示出的基于人工智能的推荐方法、以及基于人工智能的推荐模型训练方法，例如，如图6A-6C示出的基于人工智能的推荐方法。
在一些实施例中，存储介质可以是FRAM、ROM、PROM、EPROM、EEPROM、闪存、磁表面存储器、光盘、或CD-ROM等存储器；也可以是包括上述存储器之一或任意组合的各种设备。
在一些实施例中，可执行指令可以采用程序、软件、软件模块、脚本或代码的形式，按任意形式的编程语言(包括编译或解释语言，或者声明性或过程性语言)来编写，并且其可按任意形式部署，包括被部署为独立的程序或者被部署为模块、组件、子例程或者适合在计算环境中使用的其它单元。
作为示例，可执行指令可以但不一定对应于文件系统中的文件，可以可被存储在保存其它程序或数据的文件的一部分，例如，存储在超文本标记语言(HTML，Hyper Text Markup Language)文档中的一个或多个脚本中，存储在专用于所讨论的程序的单个文件中，或者，存储在多个协同文件(例如，存储一个或多个模块、子程序或代码部分的文件)中。
作为示例，可执行指令可被部署为在一个计算设备上执行，或者在位于一个地点的多个计算设备上执行，又或者，在分布在多个地点且通过通信网络互连的多个计算设备上执行。
综上所述，通过本申请实施例将用户特征与候选推荐信息的候选推荐信息特征组合，以形成融合特征，并通过对融合特征的多层次映射处理，得到对应多个指标的评分，并基于多指标聚合处理进行排序，以正向带动推荐系统中指标，提高用户互动显示行为比例。
以上所述，仅为本申请的实施例而已，并非用于限定本申请的保护范围。凡在本申请的精神和范围之内所作的任何修改、等同替换和改进等，均包含在本申请的保护范围之内。

Claims (15)

  1. 一种基于人工智能的推荐方法，所述方法由电子设备执行，所述方法包括：
    获取对应待推荐对象的多个候选推荐信息；
    获取所述待推荐对象的对象特征、以及每个所述候选推荐信息的候选推荐信息特征，并将所述对象特征分别与每个所述候选推荐信息的候选推荐信息特征进行组合，以形成对应每个所述候选推荐信息的融合特征；
    对对应每个所述候选推荐信息的融合特征进行多层次映射处理，得到每个所述候选推荐信息分别对应多个指标的评分；
    对每个所述候选推荐信息的分别对应多个指标的评分进行多指标聚合处理，得到每个所述候选推荐信息的综合评分，并根据每个候选推荐信息的综合评分对所述多个候选推荐信息进行降序排序；
    在所述降序排序的结果中选择排序靠前的至少一个候选推荐信息，并基于所选择的候选推荐信息执行对应所述待推荐对象的推荐操作。
  2. 根据权利要求1所述的方法，其中，所述获取对应待推荐对象的多个候选推荐信息，包括：
    获取以下至少一个类型的候选推荐信息：
    与对应所述待推荐对象的历史浏览信息的内容相似，且内容相似度不小于内容相似度阈值的多个候选推荐信息；
    与对应所述待推荐对象的历史行为信息的行为相似，且行为相似度不小于行为相似度阈值的多个候选推荐信息。
  3. 根据权利要求1所述的方法，其中，所述获取所述待推荐对象的对象特征、以及每个所述候选推荐信息的候选推荐信息特征，包括：
    获取对应所述待推荐对象的以下至少之一的对象特征：
    用于表征所述待推荐对象的基本信息的基础属性特征；用于表征对象社会关系的社会关系特征；用于表征对象互动行为的互动行为特征；用于表征对象阅读偏好的阅读心理特征；
    获取对应所述候选推荐信息的以下至少之一的候选推荐信息特征：
    用于表征候选推荐信息类别的类别特征；用于表征候选推荐信息内容的标签特征；用于表征候选推荐信息发布时间的时间特征；用于表征候选推荐信息来源的发布特征；用于表征候选推荐信息长度的长度特征。
  4. 根据权利要求1所述的方法，其中，所述将所述对象特征分别与每个所述候选推荐信息的候选推荐信息特征进行组合，以形成对应每个所述候选推荐信息的融合特征，包括：
    获取对应所述待推荐对象的以下至少之一的环境特征：
    向所述待推荐对象进行推送的时间特征；所述待推荐对象的对象位置特征；所述待推荐对象的设备特征；所述待推荐对象所使用的设备所处的网络特征；
    将对应所述待推荐对象的环境特征、所述对象特征以及每个所述候选推荐信息特征，组合成对应每个所述候选推荐信息的融合特征。
  5. 根据权利要求1所述的方法，其中，所述对对应每个所述候选推荐信息的融合特征进行多层次映射处理，得到每个所述候选推荐信息分别对应多个指标的评分，包括：
    通过多个专家网络，对对应每个所述候选推荐信息的融合特征进行多层次映射处理，得到与所述多个专家网络一一对应的多个子特征，并对所述多个子特征进行加权处理，得到对应各个指标的指标特征；
    通过候选推荐信息预测模型中与多个指标一一对应的预测器，并结合对应各个所述指标的指标特征对所述融合特征进行评分预测处理，得到所述融合特征基于各个指标的评分。
  6. 根据权利要求5所述的方法，其中，所述通过多个专家网络，对对应每个所述候选推荐信息的融合特征进行多层次映射处理，得到与所述多个专家网络一一对应的多个子特征，包括：
    通过所述候选推荐信息预测模型中的多个专家网络，将所述融合特征分别映射到对应所述各个专家网络的特征空间，以获得所述各个专家网络所关联的特征维度的子特征；
    所述对所述多个子特征进行加权处理，得到对应各个指标的指标特征，包括：
    通过所述候选推荐信息预测模型中与所述多个指标一一的对应的权重生成器，对所述融合特征进行最大似然处理，以得到对应各个所述指标的子特征权重分布；
    基于所述子特征权重分布中对应所述各个专家网络所获得的子特征的权重，对所述各个专家网络所获得关联的子特征进行加权处理，分别得到对应各个所述指标的指标特征。
  7. 根据权利要求1所述的方法，其中，所述对所述分别对应多个指标的评分进行多指标聚合处理，得到每个候选推荐信息的综合评分，包括：
    获取对应所述多个指标的聚合规则；
    基于所述聚合规则所包括的算子，对分别对应所述多个指标的评分进行计算处理，得到每个候选推荐信息对应所述多个指标的综合评分。
  8. 一种基于人工智能的推荐模型训练方法，所述方法由电子设备执行，所述方法包括：
    对推荐系统的日志进行预处理，以构建训练样本集合；
    基于多个专家网络、与多个指标一一对应的权重生成器、与所述多个指标一一对应的预测器，构建候选推荐信息预测模型；
    通过所述训练样本集合，对所述候选推荐信息预测模型进行多指标训练；
    其中，所述训练得到的候选推荐信息预测模型用于供所述推荐系统进行多指标的聚合处理和排序，以根据排序结果确定待推荐的候选推荐信息。
  9. 根据权利要求8所述的方法，其中，所述对推荐系统的日志进行预处理，以构建训练样本集合，包括：
    将所述推荐系统中的曝光日志、播放日志、资讯正排索引及画像特征日志中的至少一种，按照对象设备标识进行融合处理，得到样本数据；
    对所得到的样本数据进行正样本采样以及负样本采样，使得采样所得到的正样本以及负样本保持为标准比例；
    对经过采样得到的正样本和负样本进行特征提取处理，得到与各个指标匹配的特征，并将包括所述特征以及对应的指标的真实结果的集合，确定为训练所述候选推荐信息预测模型所需的训练样本集合。
  10. 根据权利要求8所述的方法，其中，所述通过所述训练样本集合，对所述候选推荐信息预测模型进行多指标训练，包括：
    初始化所述候选推荐信息预测模型，并初始化对应多个指标的损失函数，所述损失函数包括特征样本以及对应所述特征样本的评分；
    在所述候选推荐信息预测模型的每次迭代训练过程中执行以下处理：
    通过所述候选推荐信息预测模型，对所述训练样本集合包括的特征样本进行评分，得到对应所述特征样本的针对各个所述指标的评分；
    将对应所述特征样本的真实结果和所述评分代入所述损失函数，以确定所述损失函数取得最小值时对应的候选推荐信息预测模型参数；
    根据所确定的候选推荐信息预测模型参数更新所述候选推荐信息预测模型。
  11. 根据权利要求10所述的方法，其中，所述通过所述候选推荐信息预测模型，对所述训练样本集合包括的特征样本进行评分，得到对应所述特征样本的针对各个所述指标的评分，包括：
    通过所述候选推荐信息预测模型中的多个专家网络，将所述特征样本分别映射到对应各个所述专家网络的特征空间，以获得各个所述专家网络所关联的特征维度的子特征；
    通过所述候选推荐信息预测模型中包括的对应各个所述指标的权重生成器，对所述融合特征进行最大似然处理，得到对应各个所述指标的子特征权重分布；
    基于所述子特征权重分布中对应所述各个专家网络所获得的子特征的权重，对各个所述专家网络所获得关联的子特征进行加权处理，分别得到对应所述特征样本的指标特征；
    通过候选推荐信息预测模型中包括的对应各个指标的预测器，结合所述指标特征对所述特征样本进行评分预测处理，得到所述特征样本基于各个指标的评分。
  12. 根据权利要求10所述的方法，其中，所述将对应所述特征样本的真实结果和所述评分代入所述损失函数，以确定所述损失函数取得最小值时对应的候选推荐信息预测模型参数，包括：
    将对应所述特征样本的针对各个指标的真实结果和所述评分分别代入对应各个所述指标的损失函数；
    结合对应各个损失函数的损失权重，对所述各个损失函数进行加权求和处理，得到对应所述候选推荐信息预测模型的聚合损失函数；
    对所述聚合损失函数进行最小化处理，得到所述聚合损失函数取得最小值时对应的候选推荐信息预测模型参数；
    其中，所述候选推荐信息预测模型参数包括对应所述候选推荐信息模型的结构参数、以及对应所述聚合损失函数的损失权重。
  13. 一种基于人工智能的推荐装置，所述装置包括：
    信息获取模块，配置为获取对应待推荐对象的多个候选推荐信息；
    特征形成模块，配置为获取所述待推荐对象的对象特征、以及每个所述候选推荐信息的候选推荐信息特征，并将所述对象特征分别与每个所述候选推荐信息的候选推荐信息特征进行组合，以形成对应每个所述候选推荐信息的融合特征；
    特征处理模块，配置为对对应每个所述候选推荐信息的融合特征进行多层次映射处理，得到每个所述候选推荐信息分别对应多个指标的评分；
    信息排序模块，配置为对每个所述候选推荐信息的分别对应多个指标的评分进行多指标聚合处理，得到每个所述候选推荐信息的综合评分，并根据每个候选推荐信息的综合评分对所述多个候选推荐信息进行降序排序；
    推荐模块，用于在所述降序排序的结果中选择排序靠前的至少一个候选推荐信息，并基于所选择的候选推荐信息执行对应所述待推荐对象的推荐操作。
  14. 一种电子设备，包括：
    存储器，用于存储可执行指令；
    处理器，用于执行所述存储器中存储的可执行指令时，实现权利要求1至7任一项所述的基于人工智能的推荐方法、或权利要求8至12任一项所述的基于人工智能的推荐模型训练方法。
  15. 一种计算机可读存储介质，存储有可执行指令，用于被处理器执行时，实现权利要求1至7任一项所述的基于人工智能的推荐方法、或权利要求8至12任一项所述的基于人工智能的推荐模型训练方法。
PCT/CN2020/126549 2020-02-13 2020-11-04 基于人工智能的推荐方法、装置、电子设备及存储介质 WO2021159776A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/699,421 US20220215032A1 (en) 2020-02-13 2022-03-21 Ai-based recommendation method and apparatus, electronic device, and storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010095076.2 2020-02-13
CN202010095076.2A CN111291266B (zh) 2020-02-13 2020-02-13 基于人工智能的推荐方法、装置、电子设备及存储介质

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/699,421 Continuation US20220215032A1 (en) 2020-02-13 2022-03-21 Ai-based recommendation method and apparatus, electronic device, and storage medium

Publications (1)

Publication Number Publication Date
WO2021159776A1 true WO2021159776A1 (zh) 2021-08-19

Family

ID=71024518

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/126549 WO2021159776A1 (zh) 2020-02-13 2020-11-04 基于人工智能的推荐方法、装置、电子设备及存储介质

Country Status (3)

Country Link
US (1) US20220215032A1 (zh)
CN (1) CN111291266B (zh)
WO (1) WO2021159776A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113743081A (zh) * 2021-09-03 2021-12-03 西安邮电大学 技术服务信息的推荐方法
CN114491319A (zh) * 2022-01-21 2022-05-13 北京百度网讯科技有限公司 信息排序方法及装置、电子设备和存储介质
CN115018009A (zh) * 2022-07-07 2022-09-06 北京百度网讯科技有限公司 对象描述方法、网络模型的训练方法及装置
CN115311215A (zh) * 2022-07-22 2022-11-08 锋睿领创(珠海)科技有限公司 一种高速高精度的六面体检测系统、方法和存储介质

Families Citing this family (59)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111291266B (zh) * 2020-02-13 2023-03-21 深圳市雅阅科技有限公司 基于人工智能的推荐方法、装置、电子设备及存储介质
CN111783980B (zh) * 2020-06-28 2023-04-07 大连理工大学 基于双重协作生成式对抗网络的排序学习方法
CN111767429B (zh) * 2020-06-29 2023-06-02 北京奇艺世纪科技有限公司 一种视频推荐方法、装置及电子设备
CN113902952B (zh) * 2020-07-06 2024-09-20 北京达佳互联信息技术有限公司 特征评估方法、装置、电子设备及存储介质
CN111768868A (zh) * 2020-07-06 2020-10-13 上海商汤智能科技有限公司 信息处理方法及装置、电子设备和存储介质
CN112016589A (zh) * 2020-07-24 2020-12-01 苏州智制云科技有限公司 一种多特征复杂零件模型到机床装备的端到端推荐方法
CN112948671A (zh) * 2020-08-02 2021-06-11 吕维东 基于大数据和人工智能的业务处理方法及业务处理系统
CN111898028B (zh) * 2020-08-07 2024-04-19 北京小米移动软件有限公司 一种实体对象推荐方法、装置及存储介质
CN112000888B (zh) * 2020-08-24 2024-02-02 北京达佳互联信息技术有限公司 信息推送方法、装置、服务器及存储介质
CN112035741B (zh) * 2020-08-28 2022-08-30 康键信息技术（深圳）有限公司 基于用户体检数据的预约方法、装置、设备及存储介质
CN112331340B (zh) * 2020-10-14 2021-11-23 国家卫生健康委科学技术研究所 育龄夫妇妊娠概率的智能预测方法及系统
CN112348532A (zh) * 2020-10-19 2021-02-09 前海飞算科技(深圳)有限公司 最佳推荐信息处理方法和装置、存储介质和电子设备
CN112258285A (zh) * 2020-10-26 2021-01-22 北京沃东天骏信息技术有限公司 一种内容推荐方法及装置、设备、存储介质
CN112464104B (zh) * 2020-11-13 2024-05-14 中国科学院深圳先进技术研究院 一种基于网络自协作的隐式推荐方法和系统
CN114528435A (zh) * 2020-11-23 2022-05-24 北京达佳互联信息技术有限公司 搜索场景下的视频排序方法、装置、电子设备及存储介质
CN112686390B (zh) * 2020-12-28 2023-11-17 科大讯飞股份有限公司 联合建模方法、系统、装置及具有存储功能的装置
CN112699305A (zh) * 2021-01-08 2021-04-23 网易传媒科技（北京）有限公司 多目标推荐方法、装置、计算设备以及介质
CN112948449A (zh) * 2021-02-23 2021-06-11 北京三快在线科技有限公司 一种信息推荐的方法及装置
CN113158241A (zh) * 2021-04-06 2021-07-23 深圳市洞见智慧科技有限公司 基于联邦学习的岗位推荐方法及装置
CN113112148B (zh) * 2021-04-09 2022-08-05 北京邮电大学 推荐系统模型评测结果的评测方法及电子设备
CN112989074A (zh) * 2021-04-23 2021-06-18 腾讯科技（深圳）有限公司 多媒体信息推荐方法、装置、电子设备及存储介质
CN113127750B (zh) * 2021-05-19 2024-04-02 同盾科技有限公司 信息列表生成方法及装置、存储介质、电子设备
CN113343006A (zh) * 2021-06-28 2021-09-03 北京三快在线科技有限公司 图片推荐方法、装置、设备及存储介质
CN113688335B (zh) * 2021-07-23 2023-09-01 北京三快在线科技有限公司 排序理由生成方法、装置、电子设备及存储介质
CN113326440B (zh) * 2021-08-03 2021-11-02 腾讯科技（深圳）有限公司 基于人工智能的推荐方法、装置及电子设备
CN113628716A (zh) * 2021-08-05 2021-11-09 翼健(上海)信息科技有限公司 一种处方推荐系统
CN113627559A (zh) * 2021-08-19 2021-11-09 腾讯音乐娱乐科技（深圳）有限公司 一种潜力爆款歌曲挖掘方法、装置、设备及介质
CN113689237A (zh) * 2021-08-20 2021-11-23 北京达佳互联信息技术有限公司 待投放媒体资源、媒体资源处理模型的确定方法及装置
CN113642670A (zh) * 2021-08-30 2021-11-12 平安医疗健康管理股份有限公司 基于机器学习的医疗评估数据的处理方法及装置
CN113794900B (zh) * 2021-08-31 2023-04-07 北京达佳互联信息技术有限公司 视频处理方法和装置
CN113836327B (zh) * 2021-09-26 2024-08-20 杭州网易云音乐科技有限公司 推荐方法、介质、装置和计算设备
CN113903420A (zh) * 2021-09-29 2022-01-07 清华大学 一种语义标签确定模型的构建方法、病历解析方法
CN113626719B (zh) * 2021-10-12 2022-02-08 腾讯科技（深圳）有限公司 信息推荐方法、装置、设备、存储介质及计算机程序产品
CN114139046B (zh) * 2021-10-29 2023-03-24 北京达佳互联信息技术有限公司 对象推荐方法、装置、电子设备及存储介质
CN114170425A (zh) * 2021-11-02 2022-03-11 阿里巴巴（中国）有限公司 模型训练、图像分类方法、服务器及存储介质
CN114036397B (zh) * 2021-11-26 2023-08-01 北京百度网讯科技有限公司 数据推荐方法、装置、电子设备和介质
CN114186126A (zh) * 2021-11-29 2022-03-15 北京达佳互联信息技术有限公司 一种对象推荐方法、装置、电子设备及存储介质
CN114265979B (zh) * 2021-12-17 2022-11-18 北京百度网讯科技有限公司 确定融合参数的方法、信息推荐方法和模型训练方法
CN114417194A (zh) * 2021-12-30 2022-04-29 北京百度网讯科技有限公司 推荐系统排序方法、参数预测模型训练方法及装置
CN114491283B (zh) * 2022-04-02 2022-07-22 浙江口碑网络技术有限公司 对象推荐方法、装置及电子设备
CN114897162A (zh) * 2022-05-18 2022-08-12 Oppo广东移动通信有限公司 对象选取模型的训练方法、选取方法、装置及电子设备
CN114996487B (zh) * 2022-05-24 2023-04-07 北京达佳互联信息技术有限公司 媒体资源推荐方法、装置、电子设备以及存储介质
CN114756762B (zh) * 2022-06-13 2022-09-02 腾讯科技（深圳）有限公司 数据处理方法、装置、设备及存储介质
CN114912030A (zh) * 2022-06-27 2022-08-16 平安银行股份有限公司 权益模型训练方法、推荐方法及电子终端和计算机介质
CN115203564A (zh) * 2022-07-20 2022-10-18 百度在线网络技术（北京）有限公司 信息流推荐方法、装置及计算机程序产品
CN114969301A (zh) * 2022-07-29 2022-08-30 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) 在线编程教学的自适应推荐方法、装置、设备及存储介质
CN115292612B (zh) * 2022-10-10 2023-01-24 北京锘崴信息科技有限公司 基于隐私保护的信息和金融信息推荐方法及装置
CN115544385B (zh) * 2022-11-22 2023-04-04 浙江大华技术股份有限公司 平台推荐方法、电子设备和计算机可读存储介质
CN117009575A (zh) * 2022-11-30 2023-11-07 腾讯科技（深圳）有限公司 一种视频推荐方法、装置及电子设备
CN116312741A (zh) * 2023-01-28 2023-06-23 深圳太力生物技术有限责任公司 培养基配方优化方法、装置、计算机设备和存储介质
CN118467812A (zh) * 2023-02-09 2024-08-09 腾讯科技（深圳）有限公司 推荐处理方法、装置、设备、存储介质及程序产品
CN116108282B (zh) * 2023-04-12 2023-08-29 荣耀终端有限公司 一种信息推荐模型的训练方法、信息推荐方法及设备
CN116610745B (zh) * 2023-06-08 2023-12-26 深圳市巨灵数字有限公司 应用数字孪生技术的ai场景信息推送处理方法及系统
CN116450465B (zh) * 2023-06-14 2023-09-15 建信金融科技有限责任公司 数据处理方法、装置、设备及介质
CN116578652B (zh) * 2023-07-13 2024-01-16 中国人民解放军国防科技大学 一种多表关联数据集回填系统和方法
CN116884759B (zh) * 2023-07-19 2024-03-22 重庆望变电气（集团）股份有限公司 一种铁芯叠放工艺方案生成系统和方法
CN116776160B (zh) * 2023-08-23 2023-11-10 腾讯科技（深圳）有限公司 一种数据处理方法和相关装置
CN117216392B (zh) * 2023-09-19 2024-07-12 北京大学 一种基于缺失数据填补的推荐方法和系统
CN117194794B (zh) * 2023-09-20 2024-03-26 江苏科技大学 一种信息推荐方法、装置、计算机设备及计算机存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110162700A (zh) * 2019-04-23 2019-08-23 腾讯科技（深圳）有限公司 信息推荐及模型的训练方法、装置、设备以及存储介质
CN110266745A (zh) * 2019-03-08 2019-09-20 腾讯科技（深圳）有限公司 基于深度网络的信息流推荐方法、装置、设备及存储介质
CN110263257A (zh) * 2019-06-24 2019-09-20 北京交通大学 基于深度学习的多源异构数据混合推荐模型
CN111291266A (zh) * 2020-02-13 2020-06-16 腾讯科技（北京）有限公司 基于人工智能的推荐方法、装置、电子设备及存储介质

Family Cites Families (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040210661A1 (en) * 2003-01-14 2004-10-21 Thompson Mark Gregory Systems and methods of profiling, matching and optimizing performance of large networks of individuals
US11093992B2 (en) * 2017-05-05 2021-08-17 Reai Inc. Smart matching for real estate transactions
US11238083B2 (en) * 2017-05-12 2022-02-01 Evolv Technology Solutions, Inc. Intelligently driven visual interface on mobile devices and tablets based on implicit and explicit user actions
CN110020094B (zh) * 2017-07-14 2023-06-13 阿里巴巴集团控股有限公司 一种搜索结果的展示方法和相关装置
CN107958030B (zh) * 2017-11-17 2021-08-24 北京奇虎科技有限公司 视频封面推荐模型优化方法及装置
CN108830199B (zh) * 2018-05-31 2021-04-16 京东方科技集团股份有限公司 识别交通灯信号的方法、装置、可读介质及电子设备
US11157765B2 (en) * 2018-11-21 2021-10-26 Samsung Electronics Co., Ltd. Method and system for determining physical characteristics of objects
CN110263243B (zh) * 2019-01-23 2024-05-10 深圳市雅阅科技有限公司 媒体信息推荐方法、装置、存储介质和计算机设备
CN110162703B (zh) * 2019-05-13 2024-08-20 深圳市雅阅科技有限公司 内容推荐方法、训练方法、装置、设备及存储介质
CN110413877B (zh) * 2019-07-02 2023-03-28 创新先进技术有限公司 一种资源推荐方法、装置及电子设备
US11232506B1 (en) * 2019-07-03 2022-01-25 Stitch Fix, Inc. Contextual set selection
US11574353B2 (en) * 2019-10-09 2023-02-07 Target Brands, Inc. Compatibility based furniture recommendations
CN110781391B (zh) * 2019-10-22 2023-12-12 深圳市雅阅科技有限公司 一种信息推荐方法、装置、设备及存储介质
CN113508378B (zh) * 2019-10-31 2024-07-05 华为技术有限公司 推荐模型的训练方法、推荐方法、装置及计算机可读介质
US12002569B2 (en) * 2021-01-06 2024-06-04 Optum Technology, Inc. Generating multi-dimensional recommendation data objects based on decentralized crowd sourcing

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110266745A (zh) * 2019-03-08 2019-09-20 腾讯科技（深圳）有限公司 基于深度网络的信息流推荐方法、装置、设备及存储介质
CN110162700A (zh) * 2019-04-23 2019-08-23 腾讯科技（深圳）有限公司 信息推荐及模型的训练方法、装置、设备以及存储介质
CN110263257A (zh) * 2019-06-24 2019-09-20 北京交通大学 基于深度学习的多源异构数据混合推荐模型
CN111291266A (zh) * 2020-02-13 2020-06-16 腾讯科技（北京）有限公司 基于人工智能的推荐方法、装置、电子设备及存储介质

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113743081A (zh) * 2021-09-03 2021-12-03 西安邮电大学 技术服务信息的推荐方法
CN113743081B (zh) * 2021-09-03 2023-08-01 西安邮电大学 技术服务信息的推荐方法
CN114491319A (zh) * 2022-01-21 2022-05-13 北京百度网讯科技有限公司 信息排序方法及装置、电子设备和存储介质
CN115018009A (zh) * 2022-07-07 2022-09-06 北京百度网讯科技有限公司 对象描述方法、网络模型的训练方法及装置
CN115018009B (zh) * 2022-07-07 2023-04-07 北京百度网讯科技有限公司 对象描述方法、网络模型的训练方法及装置
CN115311215A (zh) * 2022-07-22 2022-11-08 锋睿领创(珠海)科技有限公司 一种高速高精度的六面体检测系统、方法和存储介质

Also Published As

Publication number Publication date
CN111291266A (zh) 2020-06-16
US20220215032A1 (en) 2022-07-07
CN111291266B (zh) 2023-03-21

Similar Documents

Publication Publication Date Title
WO2021159776A1 (zh) 基于人工智能的推荐方法、装置、电子设备及存储介质
US11893071B2 (en) Content recommendation method and apparatus, electronic device, and storage medium
CN111444428B (zh) 基于人工智能的信息推荐方法、装置、电子设备及存储介质
CN111538912B (zh) 内容推荐方法、装置、设备及可读存储介质
WO2022016522A1 (zh) 推荐模型的训练方法、推荐方法、装置及计算机可读介质
US11381651B2 (en) Interpretable user modeling from unstructured user data
CN111382361B (zh) 信息推送方法、装置、存储介质和计算机设备
CN111966914B (zh) 基于人工智能的内容推荐方法、装置和计算机设备
Bhaskaran et al. An efficient personalized trust based hybrid recommendation (tbhr) strategy for e-learning system in cloud computing
CN112119388A (zh) 训练图像嵌入模型和文本嵌入模型
US12020267B2 (en) Method, apparatus, storage medium, and device for generating user profile
CN112074828A (zh) 训练图像嵌入模型和文本嵌入模型
CN111259647A (zh) 基于人工智能的问答文本匹配方法、装置、介质及电子设备
CN111666416A (zh) 用于生成语义匹配模型的方法和装置
CN112989212A (zh) 媒体内容推荐方法、装置和设备及计算机存储介质
CN114912030A (zh) 权益模型训练方法、推荐方法及电子终端和计算机介质
CN116578729B (zh) 内容搜索方法、装置、电子设备、存储介质和程序产品
CN114741583A (zh) 基于人工智能的信息推荐方法、装置及电子设备
CN116910357A (zh) 一种数据处理方法及相关装置
CN111860870A (zh) 交互行为确定模型的训练方法、装置、设备及介质
CN116957128A (zh) 业务指标预测方法、装置、设备和存储介质
US20240005170A1 (en) Recommendation method, apparatus, electronic device, and storage medium
CN116992124A (zh) 标签的排序方法、装置、设备、介质及程序产品
Feng et al. Research on Online Learners’ Course Recommendation System Based on Knowledge Atlas in Smart Education Cloud Platform
Qu et al. The application of interactive methods under swarm computing and artificial intelligence in image retrieval and personalized analysis

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20919012

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205 DATED 09/12/2023)

122 Ep: pct application non-entry in european phase

Ref document number: 20919012

Country of ref document: EP

Kind code of ref document: A1