A short video live marketing task recommendation method based on deep learning
Through the deep learning-based recall and sorting model combined with user behavior and demographic characteristics, the problem of inaccurate recommendations in the short video live broadcast marketing platform is solved, accurate live broadcast marketing task recommendations are achieved, and business value and user experience are enhanced.
Patent Information
- Application Number
- CN202210475374.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-29
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-04-29
AI Technical Summary
The existing short video live broadcast marketing platform cannot accurately provide accurate product recommendations for streamers, and the traditional collaborative filtering algorithm cannot effectively utilize the live broadcast data of the streamers, resulting in poor recommendation results.
Using a deep learning-based method, a recall model and sorting model is constructed, combined with user behavior sequences and demographic characteristics, a list of tasks to be recommended is obtained, and a new task recommendation is optimized through cold start strategies to achieve accurate recommendations.
It improves the recommendation accuracy of the short video live broadcast marketing platform, meets the streamer's needs for selling goods, exerts the platform's commercial value, and enhances user experience.
Smart Images

Figure CN114862511B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of short video live broadcast marketing recommendation, and in particular relates to a short video live broadcast marketing task recommendation method based on deep learning. Background Art
[0002] With the rapid development of information technology, short video social platforms have become an indispensable part of people's daily lives. Short video social platforms are popular with the public for their rich content, novel dissemination methods and convenient access. In recent years, short videos have also developed rapidly in the field of e-commerce, and their commercial value has become increasingly prominent. More and more live streamers sell products through live broadcasts as a marketing carrier, and have achieved good results. As a link between short video live streamers and product sellers, short video live marketing platforms play a pivotal role. However, in short video live marketing platforms, there are still deficiencies in recommending related products to live streamers. The existing traditional collaborative filtering algorithms cannot accurately provide recommendation tasks to target users, which makes it impossible for live streamers to accurately select related products suitable for their own products. How to mine suitable products for live streamers in the marketing platform through live stream data and then achieve accurate recommendations has become an urgent problem to be solved. Summary of the invention
[0003] The purpose of the present invention is to provide a short video live marketing task recommendation method based on deep learning to address the problems existing in the short video live marketing platform in the prior art. The present invention collects user behavior sequences and live marketing sequences, obtains a list of tasks to be recommended through a recall model, and combines the user list to be recommended with the user behavior sequence and the demographic sequence, and obtains a recommended task list through a sorting model.
[0004] The objective of the present invention is achieved through the following technical solution: A short video live broadcast marketing task recommendation method based on deep learning, comprising the following steps:
[0005] (1) Obtaining the live broadcast marketing data of the broadcaster user and the user's historical behavior data;
[0006] (2) Construct a recall model. The data obtained in step (1) is used as the input of the model, and the output is the set of tasks to be recommended.
[0007] (3) When a new task appears, a cold start strategy is implemented; a ranking model is constructed, the input of which includes the set of recommended tasks, demographic characteristics, and user behavior characteristics, and the output is the click probability of the task;
[0008] (4) Select F tasks with the highest click probability for recommendation.
[0009] Furthermore, in step (1), the live broadcast marketing data of the broadcaster user User's historical behavior data Among them, X s It represents the behavioral feature set of live streamers’ sales data sorted by sales volume. represents the behavioral characteristics of the k-th live streamer’s live streaming data, X b Represents a set of user historical behavior features arranged in chronological order. Represents the kth historical behavior feature in the sequence.
[0010] Furthermore, in step (1), the live broadcast information of the host user and the user's historical behavior data are obtained by means of a web crawler.
[0011] Further, step (2) comprises:
[0012] (2.1) Input layer: The input sequence includes the user behavior sequence Live Marketing Data Series
[0013] (2.2) Embedding layer: embed high-dimensional input into low-dimensional dense feature vectors; initialize the embedding matrix to obtain the embedded user behavior sequence and live marketing data series
[0014] (2.3) Extraction layer: Use the basic attention mechanism to process the embedded sequence and obtain the processed user behavior sequence And the live marketing sequence is And splicing;
[0015] (2.4) Hidden layer, responsible for training the recall model, takes the output vector of the extraction layer as input, and uses a deep neural network for model training. The network has 3 layers and uses supervised learning to obtain a trained model;
[0016] (2.5) Output layer, which is divided into two parts: training and online serving;
[0017] Training part: Normalize the output of the hidden layer of the recall model through the Softmax function to obtain the weight vector representing each task type; then classify it to obtain the typical task set of the marketing task type to be recommended to the user.
[0018] Online service part: Use the nearest neighbor algorithm and L1 to calculate the top N tasks that are most similar to each type of typical tasks from the live marketing task database as the task set of the corresponding type to be recommended to users.
[0019] Further, step (2.3) comprises:
[0020] (2.3.1) Using the dot product method, we calculate the correlation between the feature vector and the query vector in the two embedded sequences, and use it as the weight w k ;
[0021] (2.3.2) Use the Softmax function to adjust the weight w k Normalize and get the normalized weight
[0022] (2.3.3) Based on the normalized weight Features Calculate weighted sum in, correspond or Output results of the attention mechanism corresponding to the user behavior sequence Or the output of the attention mechanism corresponding to the live marketing data sequence
[0023] (2.3.4) and Concatenate and get vector
[0024] Furthermore, in step (3), when a new task appears, a cold start strategy is implemented, specifically including: determining whether there is a new live broadcast task among the target tasks to be recommended obtained in step 2 new :
[0025] (3.1.1) If a new live broadcast task appears, new The task type is to calculate the task t new Cosine similarity s with other tasks in the live marketing task database cos ; Select and t new The K1 tasks with the highest similarity are added to the set of corresponding task types. After implementing the cold start strategy for all new tasks, the set of proposed recommended tasks is used as the input t of the sorting model.
[0026] (3.1.2) If no new live broadcast task appears, the proposed recommended task obtained in step (2) is directly used as t and input into the sorting model.
[0027] Furthermore, in step (3.1.1), K1≤N.
[0028] Furthermore, in step (3), constructing a ranking model includes:
[0029] (3.1) Input layer: The user's demographic characteristics p, target task t and user behavior characteristics Xb As input;
[0030] (3.2) Embedding layer: embed high-dimensional input into low-dimensional dense feature vectors; initialize the embedding matrix to obtain the demographic characteristics after embedding User behavior characteristics Objectives and tasks
[0031] (3.3) Attention layer: Use the basic attention mechanism to obtain different user behavior characteristics The importance of task prediction as a weight based on Features Calculate weighted sum Will Concatenate them together and input into the hidden layer;
[0032] (3.4) Hidden layer: Use the Relu activation function as the input of the next layer. Since the sorting model is a binary sorting model, the binary cross entropy loss function is used as the optimization function L2 of the sorting model.
[0033] (3.5) Output layer: Since the ranking model is a binary classification model, Softmax is used to obtain the predicted click probability.
[0034] The beneficial effects of the present invention are as follows: based on the collected user live broadcast commodity data, the present invention uses the classic YouTubeDNN as a benchmark model, and combines the characteristics of short video live broadcast marketing data to design a recall model and a sorting model based on the broadcaster's marketing data and behavior data. The present invention takes into account multiple aspects such as the number of goods brought by the broadcaster (daily sales, sales trends), the types of goods brought, relevant after-sales evaluations, and relevant behavior data of users in the short video live broadcast marketing platform, and determines the live broadcast marketing products that need to be recommended through the two main steps of recall and sorting, which is of great significance for satisfying the precise recommendation of the short video live broadcast marketing platform and maximizing the commercial value of the short video live broadcast marketing platform. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 It is a workflow diagram of the short video live broadcast marketing platform of the present invention;
[0036] Figure 2 It is a schematic diagram of the overall process of the present invention;
[0037] Figure 3 is a schematic diagram of a recall model of a recommendation system according to an embodiment of the present invention;
[0038] Figure 4 is a flowchart of a recall model and a cold start strategy according to an embodiment of the present invention;
[0039] Figure 5 Schematic diagram of a ranking model of a recommendation system according to an embodiment of the present invention. DETAILED DESCRIPTION
[0040] The present invention is further described below in conjunction with specific embodiments, but the protection scope of the present invention is not limited thereto.
[0041] like Figure 1 As shown, the present invention is a short video live broadcast marketing task recommendation method based on deep learning. First, the behavioral sequence characteristics of the input user are defined, and then a recall model based on the live broadcast data and the user behavior data is designed to recall the live broadcast marketing tasks to be recommended. Then, a sorting model based on demographic characteristics and user behavior characteristics is proposed to sort the tasks to be recommended. For the cold start problem of the system, the present invention also provides a related solution through cosine similarity calculation.
[0042] like Figure 2 As shown, the following steps are included:
[0043] Step 1: Collect the broadcaster's live broadcast room data, and pre-process and persist the data.
[0044] Step 1.1: Collect the live broadcast data of the user's live broadcast room through web crawlers, store it persistently, and extract the user's live broadcast room data Combined with historical behavior data of users in short video live marketing platform Among them, X s It represents the behavioral feature set of live streamers’ sales data sorted by sales volume. represents the behavioral characteristics of the k-th live streamer’s live streaming data, k=1~K; X b Represents a set of user historical behavior features arranged in chronological order. Represents the kth historical behavior feature in the sequence. If the number of features is greater than K, the most recent K features are taken.
[0045] Step 2: Obtain the recall task based on the recall model of live broadcast data and user historical behavior data.
[0046] Step 2.1: Design of recall model. The present invention designs the recall model by combining DNN neural network with attention mechanism. The basic recall model is as follows: Figure 3 As shown in Figure 2, the recall model includes input layer, embedding layer, extraction layer, hidden layer, and output layer.
[0047] The input layer of the recall model sorts the user's historical behavior data by time to obtain the user behavior sequence Sort the live broadcast room user marketing data by sales volume to obtain the live broadcast marketing data sequence Then the input sequence X=[X b ,X s ].
[0048] Embedding layer of the recall model:
[0049] Embed high-dimensional user behavior sequences into low-dimensional dense feature vectors After that, it is used as the input of the recall model; initialize the embedding matrix A b , the embedding formula is as follows:
[0050]
[0051] in, for The low-dimensional dense features after embedding.
[0052] Embed the high-dimensional live marketing data sequence into a low-dimensional dense feature vector After that, it is used as the input of the recall model; initialize the embedding matrix A s , the embedding formula is as follows:
[0053]
[0054] in, For the corresponding The low-dimensional dense features after embedding.
[0055] The extraction layer of the recall model uses the basic attention mechanism, and its specific steps are:
[0056] (i) Using the dot product method, calculate the correlation between the feature vector and the query vector in the two embedded sequences:
[0057]
[0058] in, correspond or q represents the query vector, w k Representation characteristics The correlation weight between q.
[0059] (ii) Use the Softmax function to adjust the weight w k Normalize:
[0060]
[0061] in, Represents the normalized related weight w k . i1=1~K, used for summation.
[0062] (iii) With features Multiply and accumulate to get the final output of the basic attention mechanism
[0063]
[0064] in, Output results of the attention mechanism corresponding to the user behavior sequence Or the output of the attention mechanism corresponding to the live marketing data sequence
[0065] (iv) The user behavior sequence output by the recall model extraction layer is: The live marketing sequence output by the recall model extraction layer is Among them, K represents the total number of features of the user behavior sequence and the live marketing sequence after the basic attention mechanism. The Concat function is used to convert and Concatenate and get vector
[0066] The hidden layer of the recall model is responsible for training the recall model and extracting the output vector of the layer As the input of the hidden layer, a deep neural network is used to train the recall model. The network has 3 layers and a supervised learning method is used to obtain a trained recall model. The training process is:
[0067]
[0068] Where u represents the final learned user vector; MLP(·) represents a multi-layer perceptron, where after each layer of the network, the Relu activation function is used as the input of the next layer.
[0069] The output layer of the recall model is divided into two parts: training and online serving.
[0070] In the training part, the output of the hidden layer of the recall model is normalized by the Softmax function to obtain the weight vector E = [e1,...,e P ], where P represents the total number of live broadcast task types; then through classification, the typical task set of the marketing task type to be recommended to the user is obtained N1 is the number of task types to be recommended to the user, N1≤P.
[0071] The online service part is used in the actual application stage, using the nearest neighbor index nearest neighbor algorithm, according to the output of the recall model From the live marketing task database, calculate the top N tasks that are most similar to the typical tasks of each type as the task set of the corresponding type to be recommended to users.
[0072] By giving relevant live marketing data X s 、Historical behavior dataX b The marketing tasks selected by the user in the past are used as the training data set of the recall model. In the training data set of this embodiment, 60% are positive samples, that is, the live marketing tasks selected by the user; and 40% are negative samples, that is, the live marketing tasks that the user does not like.
[0073] Step 3: If Figure 5 As shown, according to the sorting model and the cold start strategy, the proposed recommended tasks obtained in step 2 are sorted, including the following steps:
[0074] Step 3.1: If Figure 4 As shown, it is determined whether there is a new live broadcast task t in the target task to be recommended obtained in step 2. new
[0075] (3.1.1) If a new live broadcast task appears, a cold start strategy is implemented. new The task type is to calculate the task t new Cosine similarity s with other tasks in the live marketing task database cos :
[0076]
[0077] Among them, t j is the jth other task in the live marketing task database, and || || represents the calculation of the bi-norm. The cosine similarity s cos Sorting is performed, and K1 tasks with the highest similarity are selected and added to the set of corresponding task types, and K1≤N. After implementing the cold start strategy for all new tasks, the set of recommended tasks is used as the input t of the sorting model. At this time, the number of tasks corresponding to each task type may be different, but no less than N.
[0078] (3.1.2) If no new live broadcast task appears, the proposed recommended task obtained in step 2 is directly used as t and input into the sorting model.
[0079] Step 3.2: Based on demographic characteristics and user behavior characteristics, a ranking model is constructed to rank the obtained recall tasks. The ranking model includes an input layer, an embedding layer, an attention layer, a hidden layer, and an output layer.
[0080] The input layer of the ranking model takes the user’s demographic characteristics The target task t and user behavior characteristics X to be recommended b As input. Among them, demographic characteristics include region, gender, age, occupation and other characteristics. represents the n2th demographic feature, N2 represents the total number of demographic features; the user behavior sequence contains the user's most recent K user behavior features, each feature contains task features and time series features.
[0081] The embedding layer of the sorting model is similar to the embedding layer of the recall model. It transforms the high-dimensional input, user behavior feature X b , demographic features p and target tasks t are embedded into low-dimensional dense feature vectors and then used as inputs of the attention layer of the sorting model. The embedding matrices A1, A2, and A3 are initialized, and the embedding formula is as follows:
[0082]
[0083]
[0084]
[0085] Among them, the set of embedded demographic feature vectors User behavior characteristics after embedding is the target task after embedding.
[0086] The attention layer of the sorting model uses the basic attention mechanism to obtain the importance of different user behavior features for task prediction and calculate the user behavior features output by the attention layer of the sorting model
[0087]
[0088]
[0089] in, represents the kth attention feature extracted; i2 = 1~K, used for summation; w (k) and b (k) is the kth learned parameter, tanh is the hyperbolic tangent function, and K represents the number of user behavior features.
[0090] Use the Concat function to concatenate features as the input ζ of the hidden layer of the sorting model:
[0091]
[0092] in, represents the user behavior characteristics output by the attention layer, represents the demographic characteristics after embedding, is the proposed recommendation task after embedding.
[0093] The hidden layer of the sorting model uses the Relu activation function as the input of the next layer. Since the sorting model is a binary sorting model, the binary cross entropy loss function is used as the optimization function L2 of the sorting model:
[0094]
[0095] Among them, U is the number of samples in the training set, is the actual click probability of the user of the i3th sample in the training set, The click probability of the i3th sample predicted by the ranking model.
[0096] The output layer of the sorting model. Since the sorting model is a binary sorting model, the Softmax function is used to obtain the click rate prediction of the output layer:
[0097]
[0098] in, is the click probability predicted by the ranking model. W is the learnable parameter matrix, D is the output value of the hidden layer of the ranking model, and b is the bias parameter.
[0099] Step 4: Recommend live broadcast tasks to the target broadcaster based on the sorted tasks.
[0100] Specifically, according to the predicted click probability of the broadcaster user output by the sorting model, all the proposed recommended tasks are sorted from large to small, and the F highest live broadcast tasks are selected according to demand as the recommended tasks for the broadcaster by the short video live broadcast marketing platform.
[0101] The present invention uses a recommendation model based on deep learning combined with the idea of collaborative filtering to recall and sort related live marketing tasks, and finally selects the top-ranked live marketing tasks as the recommendation sequence for accurate recommendation, combining deep learning with the recommendation system to solve the recommendation problem in the field of short video live marketing. The present invention can be applied to the field of short video live marketing, and can also be used in other related fields, improving the accuracy of recommendations and user experience, and is of great significance in the field of marketing recommendations.
[0102] The above description is a specific embodiment of the present invention and the technical principles used. If the changes made according to the concept of the present invention do not exceed the spirit covered by the description and drawings, they should still fall within the scope of protection of the present invention.
Claims
1. A short video live marketing task recommendation method based on deep learning, characterized in that: The steps include: (1) Obtaining relevant live broadcast marketing data of the broadcaster user and the user's historical behavior data, wherein the relevant live broadcast marketing data of the broadcaster user is: The user's historical behavior data is: X s It represents the behavioral feature set of live streamers’ sales data sorted by sales volume. represents the behavioral characteristics of the k-th live streamer’s live streaming data, X b Represents a set of user historical behavior features arranged in chronological order. Represents the kth historical behavior feature in the sequence; (2) Construct a recall model by combining a DNN neural network with an attention mechanism. The recall model includes an input layer, an embedding layer, an extraction layer, a hidden layer, and an output layer. Input layer of the recall model: The input sequence includes the user behavior sequence Live Marketing Data Series Embedding layer of the recall model: embed the high-dimensional user behavior sequence into a low-dimensional dense feature vector and initialize the embedding matrix A b , get the embedded low-dimensional dense user behavior feature vector sequence Embedding matrix A b The formula is as follows: in, for The low-dimensional dense features after embedding; Embed the high-dimensional live marketing data sequence into a low-dimensional dense feature vector and initialize the embedding matrix A s , get the embedded low-dimensional dense live marketing data feature vector sequence Embedding matrix A s The formula is as follows: in, for The low-dimensional dense features after embedding; Extraction layer of the recall model: (i) Using the dot product method, calculate the correlation between the feature vector and the query vector in the two embedded feature vector sequences: in, correspond or q represents the query vector, w k Represents the correlation weight between the feature vector and the query vector q; (ii) Use the Softmax function to adjust the weight w k Normalize and get the normalized weight (iii) With features Multiply and accumulate to get the final output of the basic attention mechanism in, Output result of the attention mechanism corresponding to the low-dimensional dense user behavior feature vector sequence Or the output result of the attention mechanism corresponding to the low-dimensional dense live marketing data feature vector sequence (iv) The user behavior sequence output by the recall model extraction layer is: The live marketing sequence output by the recall model extraction layer is Among them, K represents the total number of features of the user behavior sequence and the live marketing sequence after the basic attention mechanism. The Concat function is used to convert and Concatenate and get vector Hidden layer of the recall model: responsible for training the recall model. The output vector of the extraction layer is used as the input of the hidden layer. The recall model is trained using a deep neural network. The network has three layers and uses supervised learning to obtain a trained recall model. The training process is: Where u represents the final learned user vector; MLP(·) represents a multi-layer perceptron, where after each layer of the network, the Relu activation function is used as the input part of the next layer; The output layer of the recall model: divided into two parts: training and online serving; In the training part: the output of the hidden layer of the recall model is normalized by the Softmax function to obtain the weight vector E = [e1,...,e P ], where P represents the total number of live broadcast task types; then through classification, the typical task set of the marketing task type to be recommended to the user is obtained N1 is the number of task types to be recommended to the user, N1≤P; Online service part: used in the actual application stage, the nearest neighbor algorithm is used to calculate the top N tasks most similar to each type of typical tasks from the live marketing task database according to L1, which are used as the target task set of the corresponding type to be recommended to users; (3) When a new task appears, a cold start strategy is implemented; a ranking model is constructed, the input of which includes the set of target tasks to be recommended, demographic characteristics, and user behavior characteristics, and the output is the click probability of the task; When a new task appears, a cold start strategy is implemented, specifically including: determining whether there is a new live broadcast task in the target task set to be recommended to the user obtained in step (2) new ; If a new live broadcast task appears, a cold start strategy is implemented to new Task type, calculate the new live broadcast task t new Cosine similarity s with other tasks in the live marketing task database cos ; Among them, t j is the jth other task in the live marketing task database, || || means finding the binary norm; The cosine similarity s obtained cos Sort and select new The K1 tasks with the highest similarity are added to the set of corresponding task types. After implementing the cold start strategy for all new tasks, the set of target tasks to be recommended is used as the input t of the ranking model. If no new live broadcast task appears, the target task set to be recommended to the user obtained in step (2) is directly used as t and input into the sorting model; The ranking model is constructed, and the input of the model includes the target task set to be recommended, demographic characteristics and user behavior characteristics, and the output is the click probability of the task, which specifically includes: Input layer of the ranking model: user demographic characteristics The target task set t and user behavior characteristics X to be recommended b As input, represents the n2th demographic characteristic, and N2 represents the total number of demographic characteristics; The embedding layer of the ranking model: the high-dimensional input user behavior features X b , demographic features p and the target task set t to be recommended are respectively embedded into low-dimensional dense feature vectors, and the embedding matrices A1, A2, and A3 are initialized to obtain the embedded low-dimensional dense demographic feature vector set Embedded low-dimensional dense user behavior feature vector set Embedded low-dimensional dense set of target tasks to be recommended in As input to the attention layer of the ranking model; The formulas for embedding matrices A1, A2, and A3 are as follows: Attention layer of the sorting model: Use the basic attention mechanism to obtain the importance of different user behavior features for task prediction and calculate the user behavior features output by the attention layer of the sorting model in, represents the kth attention feature extracted; i2 = 1~K, used for summation, K represents the number of user behavior features; w (k) and b (k) is the kth learning parameter, tanh is the hyperbolic tangent function; Use the Concat function to concatenate features as the input ζ of the hidden layer of the sorting model: in, Represents the user behavior characteristics output by the attention layer; Hidden layer of the sorting model: Use the Relu activation function as the input of the next layer. Since the sorting model is a binary sorting model, the binary cross entropy loss function is used as the optimization function L2 of the sorting model. Among them, U is the number of samples in the training set, is the actual click probability of the user of the i3th sample in the training set, The click probability of the i3th sample predicted by the ranking model; Output layer of the sorting model: Since the sorting model is a binary classification model, Softmax is used to get the predicted click probability: in, is the click probability predicted by the ranking model, W is the learnable parameter matrix, D is the output value of the hidden layer of the ranking model, and b is the bias parameter; (4) Select F tasks with the highest click probability for recommendation.
2. The short video live marketing task recommendation method based on deep learning according to claim 1 is characterized in that: The selection and new Among the K1 tasks with the highest similarity, K1≤N.
