CN112330362A - Rapid data intelligent analysis method for internet mall user behavior habits - Google Patents

Abstract

The invention relates to the technical fields of network marketing and artificial intelligence, in particular to a fast data intelligent analysis method for user behavior habits of Internet malls, including a sequence of user behavior patterns of Internet malls, a coding module, a space pool module and a time module. According to the user's current behavior, more targeted sales recommendations are given; using the advantage that the micro-column in HTM contains multiple cells, the input location information can be distinguished, and the input active cell set can be reset to the learning cell set , so that when HTM is learning online, it can learn for the current input sequence and improve the learning efficiency; in the process of learning repeated sequences, reducing the number of active cell sets can also effectively reduce the number of self-associated cells and reduce the possibility of loop prediction. The learning effect of the HTM is improved; the present invention sets the synapse value in the newly added dendritic branch above the connection threshold, thereby improving the learning efficiency of the HTM.

Description

Fast data intelligent analysis method for user behavior habits of Internet malls

technical field

The invention relates to the technical fields of network marketing and artificial intelligence, in particular to a fast data intelligent analysis method for the behavior habits of users in Internet malls.

Background technique

Internet mall user behavior habits are carried out in specific scenarios, users also recognize products through scenarios, and have different needs in different scenarios. If the selling point of the product is connected with the user's needs, and the user's pain point is effectively touched by the use of the scene, the emotional resonance of the consumer is aroused, and the desire to buy can be stimulated. For initial users or users with little consumption information, due to insufficient user information, it is difficult to dig out the consumption habits of users, and it is difficult to provide users with targeted recommendations and improve the transaction rate of products. Scenario marketing can build the display order of products based on the causal relationship or function, and then give more targeted recommendations based on the user's current selection. Therefore, how to quickly build the memory of the scene and use the memory as the follow-up user behavior habits The basis of statistics becomes the basis for fast intelligent analysis.

Brain-like learning is a hot research topic in the field of artificial intelligence and machine learning. Hierarchical Temporal Memory HTM (Hierarchical Temporal Memory) is a machine learning technology that simulates the information processing mechanism of the human brain by simulating the organization and mechanism of cerebral cortex cells. HTM is essentially a memory-based system. The HTM network is trained by a large amount of temporal data, stores a large number of pattern sequences, and predicts the next possible input through the memorized pattern sequence.

Different from the existing artificial neural network, HTM takes cells as the basic unit and is managed in a hierarchical manner; first, several cells are formed into a micro-column, and then these micro-columns constitute the HTM network space. The spatial pooling algorithm and the temporal pooling algorithm are two important steps in training HTM. First, the spatial pooling algorithm is used to select some activated micropillars from all the micropillars to correspond to the current input. Then use the time pooling algorithm to select partially activated cells from these microcolumns to express the location information of the input, and by adjusting the relevant dendritic branches on these active cells, build the correlation between the input and the input, and learn; at the same time, use the active cells and the constructed dendritic branches to make predictions about the input at the next moment.

The current time pooling algorithm only uses the simple Hebbian rule to establish the association between the active cells at two adjacent moments before and after by adjusting the connection value of the synapses in the dendritic branch, and learn the characteristics of the sequence; and only in the dendritic branch The sequence of learning tasks can only be completed after the middle-connected synapses accumulate to a certain threshold. If you want to build a fast learning model, you must design a new type of time pooling algorithm to improve the efficiency and effect of the HTM learning sequence.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is: in order to solve the problems existing in the above background technology, an improved fast data intelligent analysis method for Internet mall user behavior habits is provided, which solves the problem of low learning efficiency during HTM (Hierarchical Temporal Memory) fast training. and poor learning outcomes.

The technical solution adopted by the present invention to solve the technical problem is as follows: a fast data intelligent analysis method for user behavior habits of Internet malls, comprising a sequence of user behavior patterns of Internet malls, a coding module, a space pool module and a time module. The temporal module includes acquiring the set of micropillars activated by the input, generating the learning cell set, adjusting the dendritic branches, adjusting the active cell set and the prediction unit, including the following steps:

Step 1, according to the scenario of the Internet mall, construct the user behavior sequence, and form different user behavior combination timing patterns;

Step 2, for different Internet shopping mall marketing scenarios, use the user behavior combination mode with time sequence characteristics as the training object of the HTM fast training model;

Step 3, use the space pool algorithm to select some micro-pillars from all the micro-pillars for activation, and make the activated micro-pillars correspond to a certain commodity in the current user behavior pattern;

Step 4: Use the input position information to generate a learning cell set and a temporary active cell set on the activated micropillars, which enables the learning process to be carried out for the sequence of the current position, improves the accuracy of HTM learning, and in the process of learning repeated sequences , effectively reducing the number of self-associated cells, reducing the possibility of loop prediction, and improving the learning effect of HTM;

Step 5: Adjust the dendritic branches associated with adjacent inputs on the learning cell, and set the synaptic value in the newly added dendritic branch as the connectivity value according to the characteristics of online learning, so that the time pooling algorithm can pass one training. Quickly form memory and learning of sequences, and improve the learning efficiency of HTM;

Step 6, use the adjusted active cell set to predict the commodity at the next moment, and set it as the recommended commodity.

Further, the space pool module is used to obtain an activated micro-pillar set, and the time pool module includes a generation learning cell set unit, and the input end of the learning cell set unit is connected to the output end of the space pool module, and the learning The output end of the cell set unit is connected to the adjustment dendritic branch unit and the prediction unit in turn.

The beneficial effects of the present invention are:

1. For the scenario of the Internet mall, build a correlation model of user behavior. For users who cannot obtain historical information, more targeted sales recommendations can be given based on the user's current behavior;

2. When generating learning cells, the present invention utilizes the advantage that the micro-columns in HTM contain multiple cells, and can distinguish the input position information, reset the input active cell set as the learning cell set, and use it to express when predicting. The input current position information facilitates the association between the subsequent learning content and the input of the current position, so that when HTM learns online, it can learn for the current input sequence and improve the learning efficiency; in the process of learning the repeated sequence, the number of active cell sets is reduced, It can also effectively reduce the number of cells associated with itself, reduce the possibility of loop prediction, and improve the learning effect of HTM;

3. The present invention only adjusts the dendritic branches associated with adjacent inputs, and according to the characteristics of rapid training, the synapse value in the newly added dendritic branches is set above the connectivity threshold, so that the algorithm can be trained through one training. Form the memory and learning of the pattern sequence, and improve the learning efficiency of HTM.

Description of drawings

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

Fig. 1 is the flow chart of the analysis method of the present invention.

Detailed ways

The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, and only illustrate the basic structure of the present invention in a schematic manner, so they only show the structures related to the present invention.

Example 1:

The fast data intelligent analysis method for Internet mall user behavior and habits shown in FIG. 1 includes a sequence of Internet mall user behavior patterns, a coding module, a space pool module, and a time module, and the time module includes a collection of micro-columns that obtains input activation. , generate learning cell sets, adjust dendritic branches, adjust active cell sets and make predictions in six steps:

Step 1. Use the scene to construct the sequence of commodities and form different timing patterns of commodity combinations.

Step 2, for different marketing scenarios, the user behavior combination mode with time sequence characteristics is used as the training object of the HTM fast training model;

In step 3, the set of micro-columns activated by the input can be generated by the space pool algorithm, and some micro-columns are selected from all the micro-columns for activation, and the activated micro-columns are corresponding to the current input;

Step 4, the generation of the learning cell set depends on the predicted cell set at the previous moment. For the unpredicted activated micropillars, the selected learning cells can express the input current position information, so that the learning process is more specific to the sequence at the current position; specifically ,

Step 4.1, the method of generating the learning cell set is:

Step 41.1, if the cell on the activated micro-column is predicted by the input at the previous moment, the cell is set as the current input learning cell, which is expressed as:

Step 4.1.2, if all cells on the activated micro-column are not predicted by the input at the previous moment, then randomly select the cell with the least number of dendritic branches among all the cells on the micro-column as the learning cell, which is expressed as:

为t时刻第j个微柱上第i个细胞被选为学习细胞，i为微柱上细胞的编号，j是微柱的编号，W^t为t时刻输入激活的微柱集合，

为t-1时刻的预测细胞矩阵；min_j(the segment′s number of cell_i，j)为第j个微柱上包含最少树突分支的细胞编号；The learning cells selected by the above method constitute a learning cell set; wherein,

is the i-th cell on the j-th micro-column at time t is selected as the learning cell, i is the number of the cell on the micro-column, j is the number of the micro-column, W ^t is the set of activated micro-columns at time t,

is the predicted cell matrix at time t-1; min _j (the segment's number of cell _{i, j} ) is the number of the cell containing the least dendritic branches on the jth micropillar;

Step 4.2, in the activated microcolumns screened by the spatial pool algorithm, use the following rules to generate active cells. If there are predicted cells on the activated microcolumn, the cell is set as an active cell. With no predicted cells, all cells on the microcolumn are set as active cells.

表示t时刻第j个微柱上第i个细胞被置为活跃细胞，i为微柱上细胞的编号，j是微柱的编号，W^t为t时刻输入激活的微柱集合，

表示第j个微柱上第i个细胞为t-1时刻的预测细胞；The active cells selected by the above method constitute a temporary active cell set; wherein

Indicates that the i-th cell on the j-th micro-column at time t is set as an active cell, i is the number of the cell on the micro-column, j is the number of the micro-column, W ^t is the set of activated micro-columns at time t,

Indicates that the i-th cell on the j-th micro-column is the predicted cell at time t-1;

Set the non-learning cells in the active cell set to the inactive state to accurately express the current position information of the input. The main operations are as follows:

代表t时刻输入对应的学习细胞矩阵；Among them, At represents the active cell matrix generated by the input at time ^t ,

Represents the input of the corresponding learning cell matrix at time t;

Step 5. In the step of adjusting dendritic branches, the dendritic branches that need to be adjusted are either active, or new dendritic branches are added to the learning cells, the association between the front and rear inputs is constructed, and the synapses in the new dendritic branches are added. The value is set to a value above the connectivity threshold to improve the efficiency of learning; specifically, the process of adjusting dendritic branches in step 3 is:

Step 5.1, if the currently processed learning cell is a cell predicted by the input at the previous moment, adjust the active dendritic branch on the cell to strengthen the correlation with the input at the previous moment, and select the active dendritic branch on the cell. The conditions are:

Step 5.2, if the currently processed learning cell is not the cell predicted by the input at the previous moment, add a dendritic branch on the cell to build a synapse with the input learning cell at the previous moment, and form a synapse with the input learning cell at the previous moment. The association between inputs, the newly added dendritic branch is expressed as:

为第j个微柱中第i个细胞上所具有的第k个树突分支的连通性矩阵，activeThreshold为树突分支的活跃阈值，

是在第j个微柱中第i个细胞上新增的树突分支，CONNECTED_PERMANANCE为树突分支中突触的连通阈值；

为t-1时刻输入对应的学习细胞矩阵。in,

is the connectivity matrix of the kth dendritic branch on the ith cell in the jth microcolumn, activeThreshold is the active threshold of the dendritic branch,

is the newly added dendritic branch on the i-th cell in the j-th microcolumn, CONNECTED_PERMANANCE is the connectivity threshold of the synapse in the dendritic branch;

Enter the corresponding learning cell matrix for time t-1.

Step 6, adjust the active cell set and make predictions, and make predictions by reducing the range of active cells, which can improve the ability of the subsequent input to generate learning cells to distinguish the position information, and at the same time facilitate the association between the subsequent learning content and the input of the current position; The characteristics of fast training improve the learning effect and efficiency of HTM on sequences.

Based on the above method, the present invention also proposes a fast data intelligent analysis method for analyzing the behavior and habits of Internet mall users, including a space pool module and a time pool module, and the space pool module is used to obtain an activated set of micropillars; the The time pooling module includes a generating learning cell set unit, the input end of the learning cell set unit is connected to the output end of the spatial pooling module, and the output end of the learning cell set unit is sequentially connected to the adjustment dendritic branch unit and the prediction unit.

Example 2:

In this embodiment, taking "abab" as an example of the input sequence in the rapid training of the present invention, first through the learning of the space pool, it is assumed that the input a will activate the micropillars 1 and 3, and the micropillars b will activate the micropillars 2 and 4. , and there are 4 cells on each microcolumn, and set the connectivity threshold of dendrites to 0.8. The following describes the process when studying online:

For the first input a in the sequence, because there is no context, it is assumed that the learning cells that the time pool generates this input are: cell 1, ₁ and cell ₃ , 1, representing the first cell of micro-column No. 1 and micro-column No. 3 respectively. The first cell of the column, and no dendritic branches to adjust, also reset the active cells to cell _1,1 and cell _3,1 .

For the second input b in the sequence, the active cell at the previous moment is not predicted to the current input, so there is no predicted cell on the activated micro-column, assuming that the learning cells that generate this input by the time pool are: cell ₂ , 1 and cell _{4 , 1} , respectively represent the first cell of No. 2 micro-column and the first cell of No. 4 micro-column, and new dendritic branches are added to these two cells, respectively [cell 1, ₁ = 0.8, cell _{3 , 1} = 0.8], indicating that both cells have established synaptic connections to cell _1,1 and cell _3,1 , and the synapses are in a connected state. Also reset the active cells to cell _2,1 and Cell _4,1 .

For the third input a in the sequence, the active cell at the previous moment is not predicted to the current input, so there is no predicted cell on the activated micro-column. It is assumed that the learning cells that generate this input by the time pool are: cell _{1, 2} and cell _{3 , 2} , respectively represent the second cell of No. 1 micro-column and the second cell of No. 3 micro-column, and new dendritic branches are added to these two cells, respectively [cell ₂ , 1 = 0.8, cell _{4 , 1} = 0.8], indicating that both cells establish synaptic connections to cell _2,1 and cell _4,1 , and the synapses are in a connected state. Also reset the active cells to cell _{1, 2} and cell ₃ , 2.

For the fourth input b in the sequence, the active cell at the previous moment is not predicted to the current input, so there is no predicted cell on the activated micro-column. It is assumed that the learning cells that generate this input by the time pool are: cell 2, ₂ and cell _{4 , 2} , respectively represent the second cell of No. 2 micro-column and the second cell of No. 4 micro-column, and new dendritic branches are added to these two cells, respectively [cell _{1, 2} = 0.8, cell _{3 , 2} = 0.8], indicating that both cells have established synapses to cell _1,2 and cell _3,2 , and the synapses are in a connected state. Also reset the active cells to cell _2,2 and cell _4,2 .

Through fast training, HTM can learn the complete content of the "abab" sequence, but the existing time pool learning algorithm can only learn the "aba" sequence, and the fourth step of learning only strengthens the first input and the second input. relationship between.

Aiming at the scenario of the Internet shopping mall, a correlation model of user behavior is constructed, and for users who cannot obtain historical information, more targeted sales recommendations can be given based on the current behavior of the user; when generating learning cells, the present invention utilizes the HTM medium microcomputer. The column contains the advantages of multiple cells, which can distinguish the input position information, reset the input active cell set to the learning cell set, and use it to express the current position information of the input during prediction, which is convenient for the subsequent learning content and the current position. The input establishes association, so that when HTM learns online, it can learn for the current input sequence and improve the learning efficiency; in the process of learning repeated sequences, reducing the number of active cell sets can also effectively reduce the number of self-associated cells and reduce the occurrence of loop prediction. It is possible to improve the learning effect of HTM; the present invention only adjusts the dendritic branches associated with adjacent inputs, and according to the characteristics of rapid training, the synapse value in the newly added dendritic branches is set above the connectivity threshold, so that The algorithm can form the memory and learning of the pattern sequence through one training, and improve the learning efficiency of HTM.

In the description of this specification, reference to the terms "one embodiment," "some embodiments," "exemplary embodiment," "example," "specific example," or "some examples", etc., is meant to incorporate the embodiments A particular feature, structure, material, or characteristic described by an example or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

Taking the above ideal embodiments according to the present invention as inspiration, and through the above description, relevant personnel can make various changes and modifications without departing from the technical idea of the present invention. The technical scope of the present invention is not limited to the contents in the specification, and the technical scope must be determined according to the scope of the claims.

Claims (2)

1. a kind of fast data intelligent analysis method that is used for Internet mall user behavior habit, comprises Internet mall user behavior pattern sequence, coding module, space pool module and time module, it is characterized in that: described time module includes the activation of acquisition input. The collection of micropillars, the generation of learning cell sets, the adjustment of dendritic branches, the adjustment of active cell sets and prediction units, including the following steps: Step 1, according to the scenario of the Internet mall, construct the user behavior sequence, and form different user behavior combination timing patterns; Step 2, for different Internet shopping mall marketing scenarios, use the user behavior combination mode with time sequence characteristics as the training object of the HTM fast training model; Step 3, use the space pool algorithm to select some micro-pillars from all the micro-pillars for activation, and make the activated micro-pillars correspond to a certain commodity in the current user behavior pattern; Step 4: Use the input position information to generate a learning cell set and a temporary active cell set on the activated micropillars, which enables the learning process to be carried out for the sequence of the current position, improves the accuracy of HTM learning, and in the process of learning repeated sequences , effectively reducing the number of self-associated cells, reducing the possibility of loop prediction, and improving the learning effect of HTM; Step 5: Adjust the dendritic branches associated with adjacent inputs on the learning cell, and set the synaptic value in the newly added dendritic branch as the connectivity value according to the characteristics of online learning, so that the time pooling algorithm can pass one training. Quickly form memory and learning of sequences, and improve the learning efficiency of HTM; Step 6, use the adjusted active cell set to predict the commodity at the next moment, and set it as the recommended commodity. 2. the fast data intelligent analysis method for Internet shopping mall user behavior habit according to claim 1, is characterized in that: described space pool module is used to obtain the micro-pillar set of activation, and described time pool module comprises generating learning The cell set unit, the input end of the learning cell set unit is connected to the output end of the spatial pool module, and the output end of the learning cell set unit is connected to the adjustment dendritic branch unit and the prediction unit in turn.

Images