A session recommendation method based on global item long short-term conversion information weight graph

Abstract

The application discloses a kind of conversation recommendation methods based on global Item long-term and short-term conversion information weight graph, comprising the following steps: the Item clicked by the conversation in conversation data set is handled into the form of conversation corresponding Item sequence according to time sequence;Local long-term and short-term Item conversion information weight graph L-Graph is modeled according to conversation sequence;Global long-term and short-term Item conversion information graph weight graph G-Graph is constructed by superimposing local long-term and short-term Item conversion information weight graph L-Graph, and corresponding global weight graph matrix M is obtained;The next Item recommendation is carried out to conversation by global weight graph M.The method of the application can comprehensively consider long-term and short-term conversion information of Item in each conversation to predict the next Item click, not only considers the Item conversion information in single conversation, but also combines global Item conversion information, can provide more information for sparse interaction matrix, and can well utilize global weight graph information for Item recommendation for sparse interaction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of conversation recommendation, and in particular to a conversation recommendation method based on global Item long-term and short-term conversion information weight graph. BACKGROUND

[0002] Conversation recommendation is a branch of recommendation system, which solves the recommendation problem of anonymous users and can generate timely recommendations in a session established in the network. In the application scenario of conversation recommendation, the conversion relationship of Items in the conversation is a very important factor for accurate recommendation.

[0003] The conversion relationship between Items in the conversation includes both short-term conversion relationship between Items such as front and rear Items and long-term conversion relationship between Items with several Items in between. Incomplete capture of the conversion relationship between Items cannot provide sufficient information for accurate next-hop recommendation, and cannot achieve more accurate personalized recommendation.

[0004] Many existing conversation recommendation methods do not combine the long-term and short-term conversion information between Items in the conversation, and also rarely combine all conversation information, so the results of conversation recommendation are often unsatisfactory. In order to more comprehensively utilize the Item conversion information in the conversation and combine the information of each conversation, therefore, an effective conversation recommendation method combining Item long-term and short-term conversion information is urgently needed. SUMMARY

[0005] To solve the problems in the prior art, the present application provides a conversation recommendation method based on global Item long-term and short-term conversion information weight graph, which can more comprehensively utilize the Item conversion information in the conversation and combine the information of each conversation, thereby solving the problems mentioned in the background.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solution: a conversation recommendation method based on global Item long-term and short-term conversion information weight graph, comprising the following steps:

[0007] S1, processing the Items clicked by the conversation in the conversation data set into the form of Item sequence corresponding to the conversation in time sequence;

[0008] S2, modeling a local long-term and short-term Item conversion information weight graph L-Graph according to the conversation sequence of step S1;

[0009] S3, constructing a global long-term and short-term Item conversion information graph weight graph G-Graph by superimposing the local long-term and short-term Item conversion information weight graph L-Graph, and obtaining a corresponding global weight graph matrix M;

[0010] S4, next item recommendation for the session by the global weight graph M.

[0011] Preferably, the step S1 specifically includes: converting the multiple items involved in the same session in the time sequence of clicking the items into the sequence form of the session, and then filtering out the session sequences and items that do not meet the requirements; the session sequences and items that do not meet the requirements refer to the sessions with a length less than 5 and the items with an occurrence frequency less than five.

[0012] Preferably, the step S2 specifically includes the following:

[0013] S21, according to the session sequences of the step S1, taking the different items appearing in each sequence as a node of a local weight graph L-Graph;

[0014] S22, traversing the items in the list corresponding to the session sequences, connecting a directed edge between the target item and any item clicked after the target item;

[0015] S23, let n be the length of the list corresponding to the session, I i is the i-th clicked item, and the sequence of the clicked items is: I1, I2, I3, …, In; n , then I i is the directed edge weight of the item I j pointing to the item clicked after it.

[0016]

[0017] wherein α and β are parameters, the weight of the directed edge between any two nodes is inversely proportional to the distance between the nodes in the sequence, and the weight of the directed edge between the nodes is controlled by the two parameters α and β.

[0018] S24, obtaining the local long-term and short-term item transition information weight graph L-Graph according to each node in the session sequence and the weight of the directed edge between any two nodes.

[0019] Preferably, the step S3 specifically includes the following:

[0020] S31, calculating the weight of the directed edge in the local weight graph corresponding to each session according to the directed edge weight calculation formula, determining the weight W i of the directed edge by the head node I j and the tail node I Ii-Ij of the directed edge, and adding the weight of the directed edge to the weight M ij of the corresponding position of the matrix, wherein M ij represents the weight of the item I i pointing to the item I j .

[0021] S32, add the directed edge weights in the local long and short term conversion information graph L-Graph corresponding to all sessions to the weight matrix to obtain the weight matrix M corresponding to the global long and short term conversion information weight graph G-Graph.

[0022] Preferably, the step S4 specifically comprises the following:

[0023] S41, construct a Session-Item matrix S-I;

[0024] S42, multiply the matrix S-I with the obtained global weight matrix M to obtain an Item scoring matrix S;

[0025] S43, obtain an Item recommendation candidate set for the next click of the session according to the Item scoring matrix S;

[0026] S44, sort the Items in the candidate set according to the Item weights to perform the final Item recommendation, and the larger the Item corresponding weight is, the greater the possibility of becoming the next clicked Item of the session is.

[0027] Preferably, the matrix S-I is obtained by multiplying the total number of sessions and the total number of Items, each row of the matrix S-I represents the Items clicked by the session corresponding to the row index and the number of clicks, and the value not equal to 0 indicates that the session has clicked the corresponding Item, and the value size indicates the number of times the session has clicked the Item.

[0028] Preferably, the Item sequence form specifically comprises:

[0029] First, store the list composed of the Items corresponding to the session and the time of clicking the Item in the empty list session = [] according to the subscript index, session[0] = [[Item1, time1], [Item2, time2], [Item3, time3], [Item4, time4], [Item5, time5]] is the Items corresponding to the session 0 and the time of clicking the Item;

[0030] Then, sort the list corresponding to each session according to the time, and arrange the Item clicked at the earliest time in front of the sequence, thereby constructing the sequence form of the behavior of continuously clicking several Items of the session in sequence;

[0031] Finally, after sorting, traverse each session corresponding list, delete the list Item corresponding time, keep the session clicked Item, and arrange session[0] into session[0]=[Item1, Item2, Item3, Item4, Item5].

[0032] The beneficial effects of the present application are:

[0033] 1) The method of the present application combines global session information and long-term and short-term conversion information between Items in the session, first filters the data set and processes the session clicked Item into a sequence pattern; then controls the weight of the directed edge between Items according to the distance in the sequence; finally, the weight of the directed edge between Items in the directed graph obtained from all sessions is filled into the corresponding adjacency matrix, which can provide more abundant and comprehensive information for the next session clicked Item recommendation, so as to improve the accuracy of the model when recommending the next Item.

[0034] 2) The method of the present application can more comprehensively consider global session information and long-term and short-term conversion information between Items in the session, and can also make more accurate recommendations for sessions that provide less information. The method of the present application not only considers the Item conversion information in a single session, but also combines global Item conversion information, which can provide more information for sparse interaction matrices and can well utilize global weight graph information for Item recommendation. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The figure is a step flowchart of the method of the present application;

[0036] Figure 2 The figure is a session recommendation method model diagram of the global long-term and short-term conversion information weight graph;

[0037] Figure 3 The figure is a local session graph L-Graph corresponding to the session sequence. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0039] Embodiment 1

[0040] The application provides a technical scheme: a session recommendation method based on a global Item long-term and short-term conversion information weight graph, a step flow as shown in the figure Figure 1 includes the following steps:

[0041] S1, the session clicked by the session data set Item is processed into the session corresponding Item sequence form in time sequence;

[0042] The step S1 specifically includes: the multiple Items involved in the same session are sorted and converted into the sequence form of the session according to the time sequence of the clicked Item, and then the session sequence and Item that do not meet the requirements are filtered out; the session sequence and Item that do not meet the requirements refer to the session with a length less than 5 and the Item with an occurrence frequency less than five times.

[0043] Further, the Item sequence form specifically includes:

[0044] First, the list composed of the Items involved in the corresponding session and the time of clicking the Item is stored in the empty list session = [] according to the subscript number, and session [0] = [[Item1, time1], [Item2, time2], [Item3, time3], [Item4, time4], [Item5, time5]] is the Item involved in the session 0 and the time corresponding to the clicked Item;

[0045] Then, the list corresponding to each session is sorted according to time, and the Item with the earliest clicking time is arranged in front of the sequence, thereby constructing the sequence form of the session continuous clicking several Items in sequence;

[0046] Finally, after sorting, the list corresponding to each session is traversed, the time corresponding to the Item in the list is deleted, the Item clicked by the session is retained, and session [0] is arranged into session [0] = [Item1, Item2, Item3, Item4, Item5].

[0047] S2, according to the session sequence of step S1, a local long-term and short-term Item conversion information weight graph L-Graph is modeled;

[0048] Further, step S2 specifically includes the following:

[0049] S21, according to the session sequence of step S1, different Items appearing in each sequence are taken as nodes of a local weight graph L-Graph;

[0050] S22, traverse the Item in the list corresponding to the session sequence, connect a directed edge between the target Item and any one Item clicked after the target Item;

[0051] S23, let n be the length of the list corresponding to the session, I i is the i th clicked Item, and the clicked Item sequence is: I1, I2, I3, …, I n , then I i points to the I j clicked after it;

[0052]

[0053] Wherein α, β are parameters, the weight of the directed edge between any two nodes is inversely proportional to the distance between the nodes in the sequence, and the weight of the directed edge between the nodes is controlled by two parameters α, β;

[0054] S24, according to each node in the session sequence and the weight of the directed edge between any two nodes, obtain the local long-short term Item conversion information weight graph L-Graph.

[0055] S3, construct a global long-short term Item conversion information graph weight graph G-Graph by superimposing the local long-short term Item conversion information weight graph L-Graph, and obtain the corresponding global weight graph matrix M;

[0056] Further, step S3 specifically includes the following:

[0057] S31, according to the directed edge weight calculation formula, calculate the weight of the directed edge in the local weight graph corresponding to each session, and determine that the weight of the directed edge is added to the corresponding position weight M i of the matrix, M j represents the weight of the Item I pointing to the Item I ij ; ij i j

[0058] S32, add the weight of the directed edge in the local long-short term conversion information graph L-Graph corresponding to all sessions to the weight matrix, and obtain the weight matrix MS4, recommend the next Item for the session according to the global weight graph M.

[0059] Further, the step S4 specifically includes the following:

[0060] S41, construct a Session-Item matrix S-I; ​​​

[0061] S42, matrix S-I is multiplied with the obtained global weight matrix M to obtain an item scoring matrix S;

[0062] S43, obtaining an item recommendation candidate set for the next click of the session according to the item scoring matrix S;

[0063] S44, performing final item recommendation on the items in the candidate set according to the item weight size, and the larger the item corresponding weight is, the greater the possibility of becoming the item clicked next in the session is.

[0064] Further, the matrix S-I is obtained by multiplying the total number of sessions and the total number of items, each row of the matrix S-I represents the item clicked by the session corresponding to the row number and the number of clicks, and the value not equal to 0 indicates that the session clicks the corresponding item, and the size of the value indicates the number of times the session clicks the item.

[0065] The application scheme combines global session information and long-term and short-term conversion information between items in the session, filters the data set first, and processes the items clicked by the session into a sequence mode; then controls the weight of the directed edge between the items in the sequence mode according to the distance; finally, the weight of the directed edge between the items in the directed graph obtained from all sessions is filled into the corresponding adjacency matrix, which can provide more rich and comprehensive information for the item recommendation for the next session click, so as to improve the accuracy of the model when performing the next item recommendation.

[0066] Embodiment 2

[0067] A session recommendation method based on a global item long-term and short-term conversion information weight graph, the method model is as shown in Figure 2 The method comprises the following steps:

[0068] 1) The Diginetica data set contains 982961 clicks, 719470 training sessions, 60858 test sessions, 43097 items, the average session length is 5.12, and the items clicked by the session contain time. We can process the items involved in each session in the Diginetica data set into a sequence form according to the time sequence and filter the sessions and items that do not meet the requirements, and the process is as follows:

[0069] 1-1) In order to better perform the recommendation and obtain more valuable information, we filter out the sessions with a length of 1 and the items with an occurrence frequency less than five times in the Diginetica data set;

[0070] 1-2) Convert the multiple Items involved in the same session into the sequence form according to the time sequence of clicking the Items: first, store the list composed of the Items involved in the corresponding session and the time of clicking the Items in the empty list session = [] according to the subscript sequence number, for example, session[0] = [[Item1, time1], [Item2, time2], [Item3, time3], [Item4, time4], [Item5, time5]] is the Item and the corresponding time of clicking the Item involved in session 0; second, sort the list corresponding to each session according to the time, and arrange the Item with the earliest clicking time in front of the sequence, so as to build the sequence form of the behavior of continuously clicking several Items in the session; finally, after sorting, traverse the list corresponding to each session, delete the time corresponding to the Item in the list, and only keep the Item clicked by the session, for example, the above session[0] is arranged into session[0] = [Item1, Item2, Item3, Item4, Item5].

[0071] 2) Build a local long-short term Item transition information weight graph L-Graph from the session sequence, and the L-Graph schematic diagram is as shown in Figure 3 The process includes:

[0072] 2-1) Take the different Items appearing in each sequence as a node of a local weight graph L-Graph, and the connection between the nodes is a directed weight edge. Traverse each node to connect a directed edge between the current node and each node after the current node, and the directed edge indicates that the session clicks the tail node after clicking the head node, and the weight of the directed edge represents the possibility of clicking the tail node after clicking the head node;

[0073] 2-2) The weight of the directed edge between any two nodes is inversely proportional to the distance between the nodes in the sequence. Generally speaking, the farther the distance between the Items in the sequence, the weaker the correlation between the Items. Here, the weight of the directed edge between the nodes is controlled by two parameters a, b; the specific method is as follows: let n be the length of the list corresponding to the session, I i is the i-th clicked Item, and the sequence of the clicked Items is: I1, I2, I3, …, In. n is the i-th clicked Item, and the sequence of the clicked Items is: I1, I2, I3, …, In. i is the i-th clicked Item, and the sequence of the clicked Items is: I1, I2, I3, …, In. j The calculation formula of the directed edge weight of I

[0074]

[0075] Wherein, a, b are parameters, control the influence degree of distance on weight and the variance of weight value, from the weight formula can be seen that the weight is related to the interval distance between Item and Item, with the increase of the ordering distance between I i With the increase of the ordering distance between I j , the numerator (n-j+1) decreases, the denominator (n-i) is unchanged, when a, b take positive values, the weight decreases with the increase of the ordering distance between I i With the increase of the ordering distance between I j , we traverse the interval (0, 2) with 0.3 interval to assign a, b, that is, the possible values of a, b are (0.3, 0.6, 0.9, 1.2, 1.5, 1.8), and the best a and b value combination is found by the Item recommendation accuracy.

[0076] 3) Construct the global long and short-term Item conversion information graph weight graph G-Graph from the local weight graph L-Graph, the process includes:

[0077] 3-1) Superimpose all the local long and short-term Item conversion information weight graphs L-Graph, that is, add the weight values of the corresponding positions in all local long and short-term conversion information graphs L-Graph to the weight matrix M, to obtain the weight matrix M corresponding to the global long and short-term conversion information weight graph G-Graph, which is represented on the graph as superimposing all local weight graphs L-Graph, the same nodes coincide, and the weights of multiple same direction directed edges between two nodes are added to obtain the global long and short-term Item conversion information weight, and multiple same direction directed edges are merged into one directed edge. The matrix representation is to add the weights in each local long and short-term Item conversion information weight graph L-Graph to the corresponding positions in the matrix one by one;

[0078] Further, the weight matrix M described in step 3-1) is:

[0079] 1) The weight matrix M has the size of Item number multiplied by Item number, and the initial value of the matrix is 0, M ij Represents the weight of Item I i Points to Item I j ;

[0080] 2) According to the weight calculation formula in 2-2, calculate the weight of the directed edge in the local weight graph corresponding to each session, determine the weight i of the directed edge by the head node I j And the tail node I ij Add the weight of the directed edge to the corresponding position M In the matrix.

[0081] 3) Add the weights of the directed edges in the local weight graph corresponding to all sessions to the corresponding positions in the matrix. This will give you the weights of clicking other items after clicking any one item, determined by all sessions. The weight matrix at this point is the global weight matrix, corresponding to the global item long-short transition information weight graph.

[0082] 4) The global weight graph M is used to recommend the next item for the session. The process includes:

[0083] 4-1) Construct a Session-Item matrix SI. The size of the matrix is ​​the total number of Sessions multiplied by the total number of Items. Each row represents a session. Non-zero values ​​indicate that the session clicked the corresponding Item. The value indicates the number of times the session clicked the Item.

[0084] 4-2) The SI matrix obtained from 4-1) and the global weight matrix M obtained from 3-2) are multiplied to obtain the Item scoring matrix S. That is, the next Item is recommended for the conversation based on the Items interacted in the conversation and the conversion information corresponding to the Items in the global long-term and short-term Item conversion information weight matrix.

[0085] 4-3) The matrix S obtained from 4-2) represents the candidate set of recommended items for the next click in the session, which is obtained based on the transformation information of the items in the existing interactive items of the session and the global long-term and short-term item transformation information weight matrix. The items in the candidate set are sorted according to their weights and the final top K items are recommended. The larger the weight of an item, the greater the probability that it will become the item for the next click in the session.

[0086] Furthermore, the Session-Item matrix SI mentioned in step 4-1) is:

[0087] 1) The size of matrix SI is the number of sessions multiplied by the number of items. All positions in the matrix are initially set to 0. Iterate through the sessions... The corresponding item list will contain the session. Increment the value at the corresponding position in the corresponding matrix row by 1. For example, when traversing to position I... j Then SI ij =SI ij +1 means that each row of matrix SI represents the item clicked and the number of clicks in the session corresponding to the row number.

[0088] Furthermore, the S matrix mentioned in step 4-2) is:

[0089] 1) S matrix is obtained by multiplying S-I matrix and M matrix, and the size is the number of sessions multiplied by the number of items, and S-I matrix and M matrix are multiplied in order to combine the information of the item clicked by the session and the weight of the item after the item clicked by all local item long-term and short-term conversion information weight summation and the weight of the other item clicked after the item, and the global session information is combined to recommend the next item for the given session sequence;

[0090] 2) The value in the matrix row vector corresponding to each session number in the S matrix represents: according to the item currently clicked by the session, the long-term and short-term conversion information of the item is used to predict the next possible item clicked by the session, and the larger the value corresponding to the session and the item in the row vector, the greater the possibility of the item being clicked by the current session.

[0091] The method can comprehensively consider the global session information and the long-term and short-term conversion information between the items in the session, and can also make more accurate recommendations for the session with less information. The method not only considers the item conversion information in a single session, but also combines the global item conversion information, can provide more information for the sparse interaction matrix, and can well utilize the global weight graph information for item recommendation for the sparse interaction.

[0092] Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can modify the technical solutions described in the foregoing embodiments, or make equivalent replacements to part of the technical features, and any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A session recommendation method based on a global item long-short-term conversion information weight graph, characterized in that, Includes the following steps: S1. Process the items clicked in the session in the session dataset into a sequence of items corresponding to the session in chronological order; S2. Based on the conversation sequence from step S1, model a local long-short-term item transition information weight graph (L-Graph); specifically including the following: S21. Based on the session sequence in step S1, treat each different item appearing in the sequence as a node of a local weighted graph (L-Graph). S22. Traverse the Items in the list corresponding to the session sequence, and connect a directed edge between the target Item and any Item clicked after the target Item; S23. Let n be the length of the list corresponding to the session, I i Let I be the i-th clicked item, and let the sequence of clicked items be: I1, I2, I3, ..., I n , then I i Point to the I that was clicked after it j The formula for calculating the weight of a directed edge is: Where α and β are parameters, the weight of the directed edge between any two nodes is inversely proportional to the distance between nodes in the sequence, and the weight of the directed edge between nodes is controlled by the two parameters α and β; S24. Based on the weights of each node in the conversation sequence and the directed edges between any two nodes, obtain the local long-short-term item transformation information weight graph L-Graph; S3. Construct a global long-short-term item transformation information weight graph G-Graph by superimposing local long-short-term item transformation information weight graphs (L-Graphs), and obtain the corresponding global weight graph matrix M; specifically including the following: S31. Calculate the weight of the directed edges in the local weight graph corresponding to each session according to the directed edge weight calculation formula, starting from the head node I of the directed edge. i Tail node I j Determine the weight of the directed edge Weight M added to the corresponding position in the matrix ij Above, M ij Indicates ItemI i Pointing to ItemI j The weights; S32. Add the weights of the directed edges in the local long short-term transition information graph (L-Graph) corresponding to all sessions to the weight matrix to obtain the weight matrix M corresponding to the global long short-term transition information weight graph (G-Graph). S4. Recommend the next item for the session based on the global weight graph M; specifically including the following: S41. Construct the Session-Item matrix SI. Each row of the matrix SI represents the item clicked in the session corresponding to the row number and the number of clicks. S42. Multiply the matrix SI by the obtained global weight matrix M to obtain the Item scoring matrix S; S43. Based on the Item scoring matrix S, obtain the recommended Item candidate set for the next click in the session; S44. Sort the items in the candidate set according to their weights and make the final item recommendation. The higher the weight of an item, the greater the likelihood that it will become the next item to be clicked in the session.

2. The session recommendation method based on a global item long-short-term conversion information weight graph according to claim 1, characterized in that: Step S1 specifically includes: sorting multiple items involved in the same session according to the time sequence corresponding to the clicked item and converting them into a session sequence form, and then filtering out session sequences and items that do not meet the requirements; the session sequences and items that do not meet the requirements refer to sessions with a session length of less than 5 and items that appear less than five times.

3. The session recommendation method based on a global item long-short-term conversion information weight graph according to claim 1, characterized in that: The matrix SI is a matrix of the total number of sessions multiplied by the total number of items. Each row of the matrix SI represents the item clicked by the session corresponding to the row number and the number of times it was clicked. A non-zero value indicates that the session clicked the corresponding item, and the value indicates the number of times the session clicked the item.

4. The session recommendation method based on the global item long-short transition information weight graph according to claim 2, characterized in that: The item sequence format specifically includes: First, in the empty list session = [], store the list of items involved in the corresponding session and the time of clicking the item by the index number. session[0] = [[Item1, time1], [Item2, time2], [Item3, time3], [Item4, time4], [Item5, time5]] is the item involved in session 0 and the time corresponding to the click of the item; Then, the list corresponding to each session is sorted by time, with the item clicked earliest placed at the beginning of the sequence, thus constructing the behavior of clicking several items consecutively in a session into a sequential sequence. Finally, after sorting, iterate through the list corresponding to each session, delete the time corresponding to the item in the list, keep the item clicked in the session, and organize session[0] into session[0] = [Item1, Item2, Item3, Item4, Item5].

Images