CN114862511A - Short video live broadcast marketing task recommendation method based on deep learning - Google Patents
Short video live broadcast marketing task recommendation method based on deep learning Download PDFInfo
- Publication number
- CN114862511A CN114862511A CN202210475374.3A CN202210475374A CN114862511A CN 114862511 A CN114862511 A CN 114862511A CN 202210475374 A CN202210475374 A CN 202210475374A CN 114862511 A CN114862511 A CN 114862511A
- Authority
- CN
- China
- Prior art keywords
- task
- marketing
- live
- model
- user
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 31
- 238000013135 deep learning Methods 0.000 title claims abstract description 19
- 230000006399 behavior Effects 0.000 claims abstract description 54
- 238000012163 sequencing technique Methods 0.000 claims description 27
- 239000013598 vector Substances 0.000 claims description 25
- 230000006870 function Effects 0.000 claims description 18
- 238000012549 training Methods 0.000 claims description 17
- 238000000605 extraction Methods 0.000 claims description 8
- 239000011159 matrix material Substances 0.000 claims description 8
- 230000003542 behavioural effect Effects 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 4
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- 238000005457 optimization Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 2
- 238000001914 filtration Methods 0.000 abstract description 3
- 239000000969 carrier Substances 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000002085 persistent effect Effects 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0639—Item locations
-
- 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
- G06F16/951—Indexing; Web crawling techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- 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
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Probability & Statistics with Applications (AREA)
- Development Economics (AREA)
- Economics (AREA)
- Marketing (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a short video live broadcast marketing task recommendation method based on deep learning, which comprises the steps of firstly defining behavior sequence characteristics of input users, then designing a recall model based on live broadcast data and user behavior data for recalling recommended live broadcast marketing tasks, then providing a ranking model based on demographic characteristics and user behavior characteristics for ranking the recommended tasks, and for the cold start problem of a system, providing a related solution calculated through cosine similarity. According to the method, a deep learning-based recommendation model is combined with a collaborative filtering idea to recall and sort related live marketing tasks, the live marketing tasks with the top ranking are finally selected as recommendation sequences to be accurately recommended, the deep learning and recommendation system is combined, the recommendation accuracy and the user experience are improved, and the method has important significance in the field of marketing recommendation.
Description
Technical Field
The invention belongs to the technical field of short-video live broadcast marketing recommendation, and particularly relates to a short-video live broadcast marketing task recommendation method based on deep learning.
Background
With the rapid development of information technology, short video social platforms are increasingly becoming an indispensable part of people's daily life. The short video social platform is popular among the public by virtue of rich content, a novel propagation mode and a convenient acquisition way, in recent years, the short video is rapidly developed in the field of e-commerce, the commercial value of the short video is increasingly highlighted, more and more carriers with goods sell products by taking live broadcast as marketing carriers, and a good effect is achieved. The short video live broadcast marketing platform plays a very important role as a link between a short video band goods broadcaster and a product seller. However, in a short video live broadcast marketing platform, the recommendation of related products to a tape-carrier is still insufficient, and the conventional traditional collaborative filtering algorithm cannot accurately provide a recommendation task for a target user, so that the tape-carrier cannot accurately select related products suitable for the tape-carrier, and how to dig out commodities suitable for the tape-carrier in the marketing platform through live broadcast data of the tape-carrier so as to realize accurate recommendation becomes a problem to be solved urgently.
Disclosure of Invention
The invention aims to provide a short video live broadcast marketing task recommendation method based on deep learning, aiming at the problems of short video live broadcast marketing platforms in the prior art. According to the method, a task list to be recommended is obtained through a recall model by collecting a user behavior sequence and a live broadcasting marketing sequence; and combining the user list to be recommended with the user behavior sequence and the demographic sequence, and obtaining a recommended task list through a sequencing model.
The purpose of the invention is realized by the following technical scheme: a short video live broadcast marketing task recommendation method based on deep learning comprises the following steps:
(1) acquiring live broadcasting marketing data of a broadcasting main user and historical behavior data of the user;
(2) constructing a recall model, wherein the data acquired in the step (1) is used as the input of the model, and the output is a set of tasks to be recommended;
(3) when a new task appears, implementing a cold start strategy; constructing a sequencing model, inputting a to-be-recommended task set, demographic characteristics and user behavior characteristics by the model, and outputting click probability of a task;
(4) and F tasks with the highest click probability are selected for recommendation.
Further, in the step (1), relevant live marketing data of the main user is broadcastedHistorical behavioral data of a userWherein, X s A set of behavioral characteristics representing tape shipper tape data sorted by sales volume,representing behavioral characteristics of kth tape-carrier tape-stock data, X b Representing a set of historical behavior features of the user arranged in chronological order,representing the kth historical behavior signature in the sequence.
Further, in the step (1), live broadcast information of the broadcast main user and historical behavior data of the user are obtained in a web crawler mode.
Further, the step (2) comprises:
(2.1) input layer: the sequence of inputs includes a sequence of user behaviorsLive marketing data sequence
(2.2) embedding layer: embedding high-dimensional input into low-dimensional dense feature vectors; initializing the embedded matrix to obtain the embedded user behavior sequenceAnd live marketing data sequences
(2.3) extraction layer: processing the embedded sequence by adopting a basic attention mechanism to obtain a processed user behavior sequenceAnd the live marketing sequence isSplicing is carried out;
(2.4) the hidden layer is responsible for training the recall model, the output vector of the extraction layer is used as input, a deep neural network is used for model training, the network has 3 layers, and a well-trained model is obtained by using a supervised learning mode;
(2.5) an output layer which is divided into two parts of training and online service;
a training part: normalizing the output of the recall model hidden layer through a Softmax function to obtain a weight vector representing each task type; then, through classification, a typical task set of the marketing task types of the users to be recommended is obtained
The online service part comprises the following steps: using nearest neighbor algorithm, according to L 1 And calculating the top N tasks most similar to the typical tasks of each type from a live marketing task database, and using the top N tasks as a task set of the corresponding type to be recommended to the user.
Further, the step (2.3) includes:
(2.3.1) respectively calculating the correlation degree of the feature vector and the query vector in the two embedded sequences by using a dot product method as a weight w k ;
(2.3.2) use Softmax function to weight w k Normalization is carried out to obtain the normalized weight
(2.3.3) based on normalized weightsTo the characteristicsCalculating a weighted sumWherein,correspond toOr Outputting results for attention mechanism corresponding to user behavior sequenceOr outputting results of attention mechanism corresponding to live marketing data sequence
Further, in step (3), when a new task occurs, implementing a cold start policy, specifically including: judging the result obtained in step 2Whether a newly-appeared live task t exists in the target tasks to be recommended new :
(3.1.1) if a new live task occurs, t is directed to the occurrence of the new task new Task type of, computing task t new Cosine similarity s with other tasks in live marketing task database cos (ii) a Selecting and t new K with highest similarity 1 Each task is added into a set of corresponding task types; after a cold start strategy is implemented on all new tasks, an obtained quasi-recommendation task set is used as an input t of the sequencing model;
and (3.1.2) if no new live broadcast task appears, directly taking the to-be-recommended task obtained in the step (2) as t and inputting the t into the sequencing model.
Further, in step (3.1.1), K 1 ≤N。
Further, in step (3), constructing a ranking model, including:
(3.1) input layer: the demographic characteristics p, the target task t and the user behavior characteristics X of the user are combined b As an input;
(3.2) embedding layer: embedding high-dimensional input into low-dimensional dense feature vectors; initializing the embedding matrix to obtain the embedded demographic characteristicsUser behavior characteristicsTarget task
(3.3) attention layer: obtaining different user behavior characteristics using a basic attention mechanismImportance to task prediction as a weightBased onTo the characteristicsCalculating a weighted sumWill be provided withSplicing, and inputting a hidden layer;
(3.4) hidden layer: using Relu activation function as the input part of the next layer, and using two-classification cross entropy loss function as the optimization function L of the ranking model because the ranking model is a two-classification ranking model 2 ;
(3.5) output layer: since the ranking model is a binary model, the predicted click probability is obtained using Softmax.
The invention has the beneficial effects that: according to the method, a recalling model and a sequencing model based on broadcasting owner marketing data and behavior data are designed based on collected live broadcast commodity data of a user, classical YouTube DNN is used as a reference model and the characteristics of short video live broadcast marketing data are combined. The method considers the quantity (daily sales volume and sales volume trend) of the live items of the broadcaster, the types of the goods with the goods, the related after-sale evaluation, the related behavior data of the user in the short-video live broadcast marketing platform and the like, determines the live broadcast marketing product to be recommended through two main steps of recall and sequencing, and has great significance for meeting the accurate recommendation of the short-video live broadcast marketing platform and playing the maximum commercial value of the short-video live broadcast marketing platform.
Drawings
FIG. 1 is a flow chart of the short video live broadcast marketing platform of the present invention;
FIG. 2 is a schematic overview of the process of the present invention;
FIG. 3 is a recall model schematic of a recommendation system in accordance with an embodiment of the present invention;
FIG. 4 is a flow chart of a recall model and cold start strategy according to an embodiment of the present invention;
FIG. 5 is a diagram of a ranking model of a recommendation system in accordance with an embodiment of the present invention.
Detailed Description
The invention will be further described with reference to specific examples, but the scope of the invention is not limited thereto.
As shown in FIG. 1, the short video live broadcasting marketing task recommendation method based on deep learning of the invention includes defining behavior sequence characteristics of an input user, designing a recall model based on live broadcasting data and user behavior data for recalling a to-be-recommended live broadcasting marketing task, and then providing a ranking model based on demographic characteristics and user behavior characteristics for ranking the to-be-recommended tasks.
As shown in fig. 2, the method comprises the following steps:
step 1: and collecting the data of the broadcast main user live broadcast room, and preprocessing and persisting the data.
Step 1.1: collecting live broadcast data of a user live broadcast room in a web crawler mode, persistently storing the live broadcast data, and extracting the data of the user live broadcast roomCombining historical behavior data of users in short video live-broadcast marketing platformWherein, X s A set of behavioral characteristics representing tape shipper tape data sorted by sales volume,representing the behavior characteristics of the kth tape broadcaster tape cargo data, wherein K is 1-K; x b Representing a set of historical behavior features of the user arranged in chronological order,representing the kth historical behavior feature in the sequence, and if the number of the features is larger than K, intercepting the latest K features.
Step 2: and obtaining a recalled task according to the recall model of the live broadcast data and the historical behavior data of the user.
Step 2.1: design of a recall model, the invention designs the recall model by combining a DNN neural network with an attention mechanism, and a basic recall model is shown in FIG. 3. The recall model comprises an input layer, an embedded layer, an extraction layer, a hidden layer and an output layer.
Recalling the input layer of the model, and sequencing the historical behavior data of the user according to time to obtain a user behavior sequenceSequencing the user marketing data of the live broadcast room according to the sales volume to obtain a live broadcast marketing data sequenceThen the input sequence X ═ X b ,X s ]。
Embedding layer of recall model:
embedding high-dimensional user behavior sequence into low-dimensional dense feature vectorThen, the model is used as the input of the recall model; initializing an embedded matrix A b The embedding formula is as follows:
Embedding high-dimensional live marketing data sequence into low-dimensional dense feature vectorThen, the model is used as the input of the recall model; initializing an embedded matrix A s The embedding formula is as follows:
The extraction layer of the recall model uses a basic attention mechanism and comprises the following specific steps:
(i) and respectively calculating the correlation degree of the feature vector and the query vector in the two embedded sequences by using a dot product method:
(ii) Using SoftmaxFunction pair of the above weights w k And (3) carrying out normalization:
(iii) Will be provided withAnd features ofMultiplying and accumulating to obtain the final output result of the basic attention mechanism
Wherein,outputting results for attention mechanism corresponding to user behavior sequenceOr outputting results of attention mechanism corresponding to live marketing data sequence
(iv) The user behavior sequence output by the recall model extraction layer isLive marketing program with recall model extraction layer outputIs listed asWhere K represents the total number of features of the user behavior sequence and the live marketing sequence after the basic attention mechanism, which would be done using the Concat functionAndsplicing to obtain vectors
The hidden layer of the recall model is responsible for training the recall model and extracting the output vector of the layerAnd as an input of the hidden layer, a deep neural network is used for recalling model training, the network has 3 layers, and a trained recall model is obtained by using a supervised learning mode. The training process is as follows:
wherein u represents the final learned user vector; MLP (-) represents a multi-layered perceptron in which, after each layer of the network, the Relu activation function is used as an input to the next layer.
The output layer of the recall model is divided into two parts of training and online service serving.
In the training part, the output of the recall model hidden layer is normalized by a Softmax function to obtain a weight vector E ═ E representing each task type 1 ,...,e P ]Wherein P represents the total number of live task types; then, through classification, a typical task set of the marketing task types of the users to be recommended is obtainedN 1 Number of types of tasks to be recommended to the user, N 1 ≤P。
The online service part is used in the actual application stage, adopts a nearest neighbor algorithm and outputs according to the recall modelAnd calculating the top N tasks most similar to the typical tasks of each type from a live marketing task database, and using the top N tasks as a task set of a corresponding type to be recommended to the user.
By giving relevant live marketing data X s Historical behavior data X b And marketing tasks selected by the user history are used as a training data set of the recall model. In the training data set of the embodiment, 60% are positive samples, namely, live marketing tasks selected by a user; 40% is a negative example, i.e., a live marketing mission that the user dislikes.
And step 3: as shown in fig. 5, the method for sequencing the to-be-recommended tasks obtained in step 2 according to the sequencing model and the cold start strategy includes the following steps:
step 3.1: as shown in fig. 4, it is determined whether there is a newly-appeared live task t in the target task to be recommended obtained in step 2 new
(3.1.1) if a new live task occurs, implementing a cold start policy. For the appearance of a new task t new Task type of, computing task t new Cosine similarity s with other tasks in live marketing task database cos :
Wherein, t j For the jth other task in the live marketing task database, | | | | | represents solving a two-norm. For the obtained cosine similarity s cos Sorting and selecting the K with the highest similarity 1 Each task is added to a set of corresponding task types, and K 1 N is less than or equal to N. And after a cold start strategy is implemented on all the new tasks, the obtained quasi-recommended task set is used as the input t of the sequencing model. At this time, the number of tasks corresponding to each task type may be different, but is not less than N.
(3.1.2) if no new live broadcast task appears, directly taking the quasi-recommendation task obtained in the step 2 as t and inputting the t into the sequencing model.
Step 3.2: and constructing a sequencing model based on the demographic characteristics and the user behavior characteristics, and sequencing the obtained recall tasks. The sequencing model comprises an input layer, an embedded layer, an attention layer, a hidden layer and an output layer.
Input layer of ranking model, user's demographic characteristicsTarget task t to be recommended and user behavior characteristic X b As an input. Wherein the demographic characteristics comprise characteristics of region, gender, age, occupation, and the like,denotes the n-th 2 Individual demographic characteristics, N 2 Representing a total number of demographic characteristics; the user behavior sequence comprises K user behavior characteristics which are the latest of the user, and each characteristic comprises a task characteristic and a time sequence characteristic.
The embedding layer of the sequencing model is similar to the embedding layer of the recall model, and the input of high dimension and the user behavior characteristic X are input b The demographic characteristics p and the target task t are respectively embedded into low-dimensional dense characteristic vectors and then used as the input of the attention layer of the sequencing model; initializing an embedded matrix A 1 、A 2 、A 3 The embedding formula is as follows:
wherein the embedded demographic feature vector setEmbedded user behavior features Is the embedded target task.
And (4) an attention layer of the sequencing model, wherein the importance of different user behavior characteristics to task prediction is obtained by using a basic attention mechanism, and the user behavior characteristics output by the attention layer of the sequencing model are calculated
Wherein,representing the extracted kth attention feature; i all right angle 2 1-K for summation; w is a (k) And b (k) For the kth learned parameter, tanh is the hyperbolic tangent function. K represents the number of user behavior features.
Splicing features using the Concat function as input ζ to the hidden layer of the ranking model:
wherein,a user behavior feature representing the attention layer output,representing the demographic characteristics after the embedding,is an embedded proposed task.
The hidden layer of the sequencing model uses a Relu activation function as an input part of the next layer, and the sequencing model is a two-class sequencing model, and uses a two-class cross entropy loss function as an optimization function L of the sequencing model 2 :
Wherein U is the number of samples in the training set,for the ith in the training set 3 The true click probability of a user for a sample,ith prediction for ranking model 3 Click probability of individual samples.
And (3) an output layer of the sorting model, wherein the sorting model is a two-classification sorting model, and the click rate prediction of the output layer is obtained by using a Softmax function:
wherein,the click probability predicted for the ranking model. W is a learnable parameter matrix, D is an output value of a hidden layer of the sequencing model, and b is a bias parameter.
And 4, step 4: and recommending the live broadcast task for the target broadcast owner according to the sequenced tasks.
Specifically, all tasks to be recommended are sequentially ranked from large to small according to the click probability of the predicted broadcasting main user output by the ranking model, and F highest live broadcasting tasks are selected according to the requirement and serve as recommendation tasks of the short-video live broadcasting marketing platform to the broadcasting owner.
According to the method, a deep learning-based recommendation model is combined with a collaborative filtering idea to recall and sort related live broadcast marketing tasks, finally, the live broadcast marketing tasks with the top ranking are selected as recommendation sequences to be accurately recommended, and the deep learning and recommendation system are combined, so that the recommendation problem in the field of short-video live broadcast marketing is solved. The method and the device are applicable to the field of short video live broadcast marketing and can also be used in other related fields, the recommendation accuracy and the user experience are improved, and the method and the device are of great significance in the field of marketing recommendation.
While the invention has been described in connection with specific embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.
Claims (8)
1. A short video live broadcast marketing task recommendation method based on deep learning is characterized by comprising the following steps:
(1) acquiring live broadcasting marketing data of a broadcasting main user and historical behavior data of the user;
(2) constructing a recall model, wherein the data acquired in the step (1) is used as the input of the model, and the output is a set of tasks to be recommended;
(3) when a new task appears, implementing a cold start strategy; constructing a sequencing model, inputting a to-be-recommended task set, demographic characteristics and user behavior characteristics by the model, and outputting click probability of a task;
(4) and F tasks with the highest click probability are selected for recommendation.
2. The deep learning-based short-video live-broadcast marketing task recommendation method according to claim 1, characterized in that in the step (1), related live-broadcast marketing data of a main user is broadcastedHistorical behavioral data of a userWherein, X s A set of behavioral characteristics representing tape shipper tape data sorted by sales volume,representing behavioral characteristics of kth tape-carrier tape-stock data, X b Representing a set of historical behavior features of the user arranged in chronological order,representing the kth historical behavior signature in the sequence.
3. The deep learning-based short-video live broadcast marketing task recommendation method according to claim 1, characterized in that in the step (1), live broadcast information of a broadcast main user and historical behavior data of the user are acquired in a web crawler manner.
4. The deep learning-based short video live marketing task recommendation method according to claim 1, wherein the step (2) comprises:
(2.1) input layer: the sequence of inputs includes a sequence of user behaviorsLive marketing data sequence
(2.2) embedding layer: embedding high-dimensional input into low-dimensional dense feature vectors; initializing the embedded matrix to obtain the embedded user behavior sequenceAnd live marketing data sequences
(2.3) extraction layer: processing the embedded sequence by adopting a basic attention mechanism to obtain a processed user behavior sequenceAnd the live marketing sequence isSplicing is carried out;
(2.4) the hidden layer is responsible for training the recall model, the output vector of the extraction layer is used as input, a deep neural network is used for model training, the network has 3 layers, and a well-trained model is obtained by using a supervised learning mode;
(2.5) an output layer which is divided into two parts of training and online service;
a training part: normalizing the output of the recall model hidden layer through a Softmax function to obtain a weight vector representing each task type; then, through classification, a typical task set L of the marketing task types of the users to be recommended is obtained 1 =[l 1 ,...,l N1 ];
The online service part comprises the following steps: using nearest neighbor algorithm, according to L 1 Calculating the top N tasks most similar to the typical task of each type from the live marketing task database, and using the top N tasks as the tasks to be recommended to the user of the corresponding typeAnd (5) task collection.
5. The deep learning-based short video live marketing task recommendation method according to claim 4, wherein the step (2.3) comprises:
(2.3.1) respectively calculating the correlation degree of the feature vector and the query vector in the two embedded sequences by using a dot product method as a weight w k ;
(2.3.2) use Softmax function to weight w k Normalization is carried out to obtain the normalized weight
(2.3.3) based on normalized weightsTo the characteristicsCalculating a weighted sumWherein,correspond toOr Outputting results for attention mechanism corresponding to user behavior sequenceOr attention mechanism corresponding to live marketing data sequenceOutputting the result
6. The deep learning-based short-video live-broadcast marketing task recommendation method according to claim 1, wherein in the step (3), when a new task occurs, a cold start strategy is implemented, and specifically includes: judging whether a newly appeared live broadcast task t exists in the target tasks to be recommended obtained in the step 2 new :
(3.1.1) if a new live task occurs, t is directed to the occurrence of the new task new Task type of, computing task t new Cosine similarity s with other tasks in live marketing task database cos (ii) a Selecting and t new K with highest similarity 1 Each task is added into a set of corresponding task types; after a cold start strategy is implemented on all new tasks, an obtained quasi-recommendation task set is used as an input t of the sequencing model;
and (3.1.2) if no new live broadcast task appears, directly taking the to-be-recommended task obtained in the step (2) as t and inputting the t into the sequencing model.
7. The deep learning-based short video live marketing task recommendation method according to claim 6, wherein in step (3.1.1), K is 1 ≤N。
8. The deep learning-based short video live broadcasting marketing task recommendation method according to claim 1, wherein in the step (3), a ranking model is constructed, and the method comprises the following steps:
(3.1) input layer: the demographic characteristics p, the target task t and the user behavior characteristics X of the user are combined b As an input;
(3.2) embedding layer: embedding high-dimensional input into low-dimensional dense feature vectors; initializing the embedding matrix to obtain the embedded demographic characteristicsUser behavior characteristicsTarget task
(3.3) attention layer: obtaining different user behavior characteristics using a basic attention mechanismImportance to task prediction as a weightBased onTo the characteristicsCalculating a weighted sumWill be provided withSplicing, and inputting a hidden layer;
(3.4) hidden layer: using Relu activation function as input for the next layerIn part, because the ranking model is a two-class ranking model, a two-class cross entropy loss function is used as an optimization function L of the ranking model 2 ;
(3.5) output layer: since the ranking model is a binary model, the predicted click probability is obtained using Softmax.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210475374.3A CN114862511A (en) | 2022-04-29 | 2022-04-29 | Short video live broadcast marketing task recommendation method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210475374.3A CN114862511A (en) | 2022-04-29 | 2022-04-29 | Short video live broadcast marketing task recommendation method based on deep learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114862511A true CN114862511A (en) | 2022-08-05 |
Family
ID=82634675
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210475374.3A Pending CN114862511A (en) | 2022-04-29 | 2022-04-29 | Short video live broadcast marketing task recommendation method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114862511A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117788105A (en) * | 2023-12-25 | 2024-03-29 | 公安县谦合广告装饰有限公司 | Online live broadcast method of E-commerce based on Internet |
-
2022
- 2022-04-29 CN CN202210475374.3A patent/CN114862511A/en active Pending
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117788105A (en) * | 2023-12-25 | 2024-03-29 | 公安县谦合广告装饰有限公司 | Online live broadcast method of E-commerce based on Internet |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110162703B (en) | Content recommendation method, training device, content recommendation equipment and storage medium | |
CN110955826B (en) | Recommendation system based on improved cyclic neural network unit | |
CN110309195B (en) | FWDL (full Width Domain analysis) model based content recommendation method | |
CN111222332A (en) | Commodity recommendation method combining attention network and user emotion | |
Huang | A hybrid particle swarm optimization approach for clustering and classification of datasets | |
CN108563755A (en) | A kind of personalized recommendation system and method based on bidirectional circulating neural network | |
CN105893609A (en) | Mobile APP recommendation method based on weighted mixing | |
CN112307332B (en) | Collaborative filtering recommendation method and system based on user portrait clustering and storage medium | |
Sreepada et al. | Mitigating long tail effect in recommendations using few shot learning technique | |
CN114202061A (en) | Article recommendation method, electronic device and medium based on generation of confrontation network model and deep reinforcement learning | |
CN113343077A (en) | Personalized recommendation method and system integrating user interest time sequence fluctuation | |
CN115510322A (en) | Multi-objective optimization recommendation method based on deep learning | |
CN115438787A (en) | Training method and device of behavior prediction system | |
CN116680320A (en) | Mixed matching method based on big data | |
CN115080868A (en) | Product pushing method, product pushing device, computer equipment, storage medium and program product | |
CN114862511A (en) | Short video live broadcast marketing task recommendation method based on deep learning | |
CN112486467B (en) | Interactive service recommendation method based on dual interaction relation and attention mechanism | |
CN116320626B (en) | Method and system for calculating live broadcast heat of electronic commerce | |
CN112435103A (en) | Intelligent recommendation method and system for post diversity explanation | |
Kuo et al. | Integration of particle swarm optimization and immune genetic algorithm-based dynamic clustering for customer clustering | |
CN115982373A (en) | Knowledge graph recommendation method combining multi-level interactive contrast learning | |
Chae et al. | Adversarial training of deep autoencoders towards recommendation tasks | |
CN113704439A (en) | Conversation recommendation method based on multi-source information heteromorphic graph | |
CN110956528A (en) | Recommendation method and system for e-commerce platform | |
Zhu et al. | Latent Factor Model with User and Fused Item Embeddings for Recommendation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |