CN105590237A - Application of high utility sequential pattern with negative-profit items in electronic commerce business decision making - Google Patents

Abstract

The invention relates to application of high utility sequential patterns (HUSP) with negative-profit items in electronic commerce business decision making. A depth pruning strategy is first used for determining whether a node is a leaf node and returning back to a parent node if yes. Then a width pruning strategy is used. Once connection items are collected, futureless items are deleted from respective lists. A child, namely a new sequence, is constructed through I-connection and S-connection. Sequences just including negative-profit items are pruned according to a negative sequence pruning strategy. An HUSPNIV algorithm outputs HUSPs complying with a condition and then calls itself recursively to construct a deeper LQS-Tree. Sequential patterns are mined according to the algorithm and the purchase behaviors of customers are analyzed by the sequential patterns. Thus, a seller may predict a product selling state in the future according to a current product selling state so as to well establish a marketing strategy, arrange the placement of products, and increase sale volumes.

Description

The application of the effective sequence pattern of the negative profit item of band in ecommerce decision-making

Technical field

The present invention relates to the application in ecommerce decision-making of effective sequence pattern with negative profit item, belong to orderThe technical field of row mode excavation application.

Background technology

Be penetrated into more and more all sidedly the every field such as human society, politics, economy, culture along with Internet technology, numberWord business has become everyone, each tissue must be faced, and ecommerce, shopping at network have been no longer a kind of choosingsSelect, and become inevitable. In recent years, shopping at network is because it is convenient, economic dispatch advantage has obtained all over the world popularizing rapidly and sending outExhibition, data volume is also explosive growth simultaneously, all increases with several grades of geometry every year, meanwhile a lot of large-scale e-business networksStand, as the Taobao under Amazon, Alibaba and day cat store, Jingdone district etc. have all accumulated a large amount of client trading data. WithIncreasing of shopping website number, the intensified competition between each network businessman. How to make full use of these data, obtain client'sShopping mode, can also make businessman's profit maximum, thereby improve better net when client is carried out to personalized commercial product recommendingThe service quality of standing and economic benefit are ecommerce problems in the urgent need to address.

Different from traditional management style, the businessman of ecommerce can not remove to understand client, the dependency number obtaining intuitivelyAccording to limited (such as client's log-on message, purchaser record etc.). Analyze and excavate by the client's purchaser record to a large amount of,Find often access and the larger sequence pattern of businessman's profit of client, for different client properties and shopping online stepSuddenly, adopt different commercial product recommending forms, in good time to the appropriate commodity of lead referral, and optimize e-business network site commodityPutting position, preferentially recommends businessman to invigorate moving commodity, can effectively increase client's Transaction apparatus meeting, and website browsing person is turnedBecome buyer, improve cross-selling ability, improve client's loyalty, and improve service quality and the economic effect of websiteBenefit.

The personalized commercial of ecommerce is recommended, and does not need to pay very large cost, only needs the content of website according to everyIndividual client's feature is carried out suitable adjustment, makes more personalized commercial recommend net according to each client's consumption preferencePage, provides more selection to client. So also set up on a network for the client on each network with regard to being equivalent toShop, carries out pointed commercial product recommending to each client, helps client from huge goods catalogue, to pick out reallyBe applicable to the commodity of oneself needs.

Carry out personalized commercial recommendation based on sequence pattern analysis, normally adopt minimum support threshold value to find dataFrequent sequence in storehouse in certain a period of time, what within this time period, which commodity can be bought by client is many. But byCan not show commercial value and impact in excavating by support threshold value the Frequent Sequential Patterns obtaining, so the large portion in themDivide and all can not provide Useful Information for the decision-making of business. In many situations, a lot of real interested valuable mouldsFormula because its support lower and screened fall, and the high pattern profit of support is not necessarily high, as household electrical applianceProfit high more a lot of than the profit of a bag nut, but it is more much lower than nut to sell the support of household electrical appliance, and everyA kind of price, profit of commodity are all different, and some clients may buy several for same commodity. In order to solve thisKind of problem, profit (effectiveness) is introduced in during Frequent Sequential Patterns excavates, and for excavating, user is interested has a higher realityWith the sequence pattern being worth (being called for short effective sequence pattern, high value sequence pattern or high profit sequence pattern) (HighUtilitySequentialPatterns, HUSP). The expression side that effectiveness can be liked with cost, profit or other userFormula is measured. Frequent Sequential Patterns Mining Problems is redefined as effective sequential mode mining, and minimum support threshold value alsoBe replaced by minimum effectiveness threshold value, each item has 2 attributes, i.e. individual event profit and quantity purchases.

There were in recent years a lot of scholars to obtain one at the interested sequence pattern this respect with profit item of digging userFixed achievement in research, but major part is all the positive profit of only having considered item, and do not mention negative profit, and negative profit is in realityEqually very important in application. In actual applications, there are some large-scale businessmans in order to attract client, push up sales, do sales promotionActivity, such as full how much amount of money freight free, full how much amount of money can be given commodity, send reward voucher etc. to client, and these are givenCommodity, reward voucher, the freight charges that send are exactly negative profit concerning the income of businessman, if finally find businessman do advertising campaign meeting than withMore toward profit. This is a kind of sales tactics, illustrates that the item of the negative profit of band is very for effective sequential mode miningUseful. Not also a lot of about the mode excavation of negative profit item at present, only have several sections of articles to discuss and excavate the negative profit item of bandThe method of frequent item set, still, also do not find so far any sequence pattern about how excavating the negative profit item of bandMethod. Therefore, how excavating user's effective sequence pattern interested and the negative profit item of band is the pass that needs solutionKey problem.

Data source taking the website user's purchase order data in e-commerce platform as excavating.

Transaction taking 5 clients in 3 months is as example, if table 1 is the unit effectiveness of each, also can claim individual event profit,Table 2 is client's purchase sequence databases of having put in order, and this database is relevant with the quantity purchase of item, so added one aboveIndividual " q-" prefix. That letter represents is commodity ID.

Table 1 unit effectiveness

Table 2Q-sequence library

Table 3Q-sequence effectiveness

A client all transaction records within certain time period form an orderly quantity series or q-sequence, canBe expressed as s=< l₁l₂...l_m>, be the ordered list of q-item collection, wherein l_k(1≤k≤m) is q-item collection. In sequence ,/Collection is sequential, a kind of commodity of each Xiang Dou representative transaction, and element refers to that this client is in the some concrete timeAll commodity of the disposable purchase of point, represent with { } or (), this client may buy same commodity in the different time periods,An item may occur in the different elements of a sequence. Every commodity (item) i_kHaving two attributes, is outside effectivenessAnd inner effectiveness (internalutility) (externalutility). Outside effectiveness refers to the unit profit of commodity, canTo be positive or negative, be called for short p (i_k). Inner effectiveness refers to the quantity purchase of commodity, is positive, is called for short quantity item(quantitativeitem) be or q-item, an ordered pair (i_k, q), wherein i_kRepresent a kind of commodity, q is that positive number is representingCommodity i_kQuantity purchase. Quantity item collection or q-item collection can be expressed as l=[(i₁,q₁)(i₂,q₂)...(i_n,q_n)], wherein (i_k,q_k) be a q-item and 1≤k≤n. In the time that q-item collection only has a q-item, bracket can omit. The large I table of Q-sequence sBe shown size (s), refer to the total quantity of q-item collection in s. A q-sequence library S is one group of components<sid, s>set, itsMiddle sid is the sequence_id of q-sequence, and s is a data sequence.

In table 2, s₁(sid=1) that in, show is commodity a, b, and d, e and their quantity purchase are respectively 4,2,4,2.In table 1, show commodity a, b, d, the unit effectiveness of e is 5 ,-3,2,4, but commodity b is disutility, loses. At s₁In(e, 2), (a, 4), (b, 2), (d, 4) are q-items; [(a, 4) (b, 2)] are q-item collection, comprise two q-items.

Summary of the invention

For the deficiencies in the prior art, the effective sequence pattern that the invention provides the negative profit item of band is determined in ecommerceApplication in plan formulation.

The efficient recursive algorithm that proposes a HUSPNIV by name in the present invention is excavated the negative profit of the interested band of userEffective sequence pattern, the main thought of described algorithm is first to adopt depth pruning strategy, judges that whether a node isLeafy node, if so, turns back to parents' node. Adopt again width beta pruning, once after having collected connection item, lack of prospectsWill from list separately, delete. Connect to be connected with S-by I-respectively and build child, i.e. new sequence. Generate newAfter sequence, cut the sequence that only comprises negative profit item according to negative sequence Pruning strategy. As the unit of each in infructescenceEffectiveness is all born, and this sequence will be cut so. HUSPNIV algorithm is exported qualified effective sequence pattern, thenRecursively call and self remove to build darker LQS-Tree.

Excavate and obtain sequence pattern by this algorithm, utilize these sequence patterns to analyze client's buying behavior, makeSeller can predict later buying and selling of commodities situation according to current buying and selling of commodities situation, thereby can better formulate battalionPutting of pin strategy, arrangement commodity, improves offtake.

Technical scheme of the present invention is as follows:

The application of the effective sequence pattern of the negative profit item of band in ecommerce decision-making, concrete steps comprise:

(1) definition

Definition q-item effectiveness: a q-item effectiveness is single q-item (i_j,q_j) effectiveness, be expressed as u (i_j,q_j)：

u(i_j,q_j)＝p(i_j)×q_j(Ⅰ)

In formula (I), i_jRepresent a kind of commodity; q_jBe positive number, representing commodity i_jQuantity purchase; P (i_j) refer to commodity i_jOutside effectiveness, i.e. commodity i_jUnit profit, be positive or negative; For example, the effectiveness of q-item (b, 2) is: u (b, 2)=-3×2＝-6。

Definition Q-item collection effectiveness: be q-item collection l=[(i₁,q₁)(i₂,q₂)...(i_j,q_j)...(i_n,q_n)] effectiveness, tableBe shown u (l):

$\begin{array}{cc}u\left(l\right)=\underset{j=1}{\overset{n}{\&Sigma;}}u(i_j,q_j)& \left(II\right)\end{array}$

In formula (II), 1≤j≤n; For example, q-item collection l=[(a, 4) (b, 2)] effectiveness be: u (l)=5 × 4+ (3) ×2＝14。

Sequence effectiveness in definition Q-sequence library S: suppose given sequence t=< t₁t₂...t_n>, the q-order corresponding with tRow s=< l₁l₂...l_m> utility schedule be shown u (t, s):

In formula (III), t=< t₁t₂...t_n>and s=<l₁l₂...l_m> referring to: a client is all within certain time periodThe transaction record orderly quantity series or the q-sequence that form;

The utility schedule of t in database S is shown u (t):

$\begin{array}{cc}u\left(t\right)=\underset{s\&Element;S}{\&cup;}u(t,s)& \left(IV\right)\end{array}---\left(4\right)$

In formula (IV), a q-sequence library S is one group of components<sid, s>set, sid is q-sequenceSequence_id, s is a data sequence; For example, sequence t=<ba>, at the s of table 2₄In, the effectiveness of t is u (t, s₄)＝{u(<(b, 1) (a, 3)>), u (<(b, 1) (a, 4)>) }={ 12,17}, the effectiveness of t in S is u (t)={ u (t, s₂),u(t,s₄),u(t,s₅)={ { 2,7}, { 12,17}, { 9,4}}. May there are multiple values with the q-sequence of sequences match, as t=just now <Ba >, it is at s₄In have two value of utilities, be respectively 12 and 17. This be different from Frequent Sequential Patterns a bit, frequent sequence mouldFormula only has a support value.

Definition Q-sequence effectiveness: the effectiveness u (s) of Q-sequence s refers to that unit effectiveness in s is the effectiveness summation of positive item, that is:

That for example, in table 3, show is the q-sequence s after defining₁，s₂，s₃，s₄，s₅Effectiveness.

The weighting sequence effectiveness of sequence t in definition Q-sequence library S, i.e. SequenceWeightedUtilization, is called for short SWU, is expressed as SWU (t), that is:

For example, SWU (<ca>)=u (s₂)+u(s₄)＝104。

(2) excavate effective sequence pattern

In multiple value of utilities in the q-sequence corresponding with t, choose maximum value of utility u_max(t) conduct is in this q-orderThe effectiveness of row;

$\begin{array}{cc}{u}_{max}\left(t\right)=\underset{s\&Element;S}{\&Sigma;}max\left\{u(t,s)\right\}& \left(VII\right)\end{array}$

Work as u_max(t) >=when ξ, sequence t is an effective sequence pattern; ξ is the given minimum effectiveness threshold value of user(minimumutility，minutil)；

To find all effective sequence patterns with the effective sequential mode mining of negative profit item, its unit effectPositive also comprising negatively with both comprising, if only comprise positive unit effectiveness, is the effective sequential mode mining with profit item.

For example, sequence<ba>Effectiveness be u_max(<ba>)=7+17+9=33. If minimum effectiveness threshold xi is 30, soThis sequence is effective sequence pattern, because u_max(s)＝33≥ξ。

(3) utilize HUSPNIV algorithm to excavate the effective sequence pattern of the negative profit item of the interested band of user

1. first adopt depth pruning strategy, judge whether a node is leafy node, if so, turn back to parents' knotPoint; Otherwise, enter step 2.; It is darker that the leafy node of depth pruning strategy in setting by identification avoids that algorithm walks.

2. adopt width beta pruning, collection can connect item, deletes the item lacking of prospects in list separately, then passes through respectively I-Connect and be connected structure child with S-, i.e. new sequence, after generating new sequence, only cuts and comprises according to negative sequence Pruning strategyThe sequence of negative profit item, bears as the unit effectiveness of each in infructescence, and this sequence will be cut so, excavatesThe effective sequence pattern of the negative profit item of the interested band of user, recursively calls and self removes to build darker LQS-Tree;

It should be noted that the list of the item in effective sequence pattern with the effective sequential mode mining of negative profit itemPosition effectiveness may be born, but has at least one to be positive, if entirely born, needs to cut, because there is no actual meaningJustice.

For example, suppose that B and C are the items with negative profit, if<BC>be candidate sequence,<BC>can be cut, becauseIn sequence, each is negative profit item. If contain positive profit item in candidate sequence, and the value of utility of this sequence is greater than orEqual minutil threshold value, this sequence is effective sequence pattern.

(4) utilize the effective sequence pattern of the negative profit item of the interested band of user that step (3) excavates, analyze clientBuying behavior, make seller predict later buying and selling of commodities situation according to current buying and selling of commodities situation, formulate marketingPutting of strategy, arrangement commodity.

Preferred according to the present invention, described LQS-Tree, i.e. Q-sequence lexicographic tree LQS-TreeT (LexicographicQ-sequenceTree, LQS-Tree), refer to: all sequences are all arranged in the structure of LQS-TreeT, LQS-The root node of TreeT is empty, and other each node is a sequence and the effectiveness of having deposited sequence, any nodeChild or be that I-connects, or be that S-connects, in LQS-TreeT arbitrarily the child of node all by the lexicographic order row who increases progressivelyRow; Suppose k-sequence t, the operation that the end of adding a new item to t forms (k+1)-sequence is called connection, if tSize does not change, and claims this I-of being operating as to connect, if the size of t how one, be referred to as so S-and connect;

In Frequent Sequential Patterns excavates, closure property plays an important role in sequential mode mining downwards. ButThat this characteristic is not adapting to effective sequential mode mining. In table 2, u_max(<ea>)＝28+27＝55，u_max(<e>)=8+12+12=32, visible sequence<e>value of utility be less than the value of utility of its father's sequence. What is more important, if weCheck the maximum utility of a paths in a complete LQS-Tree, sequence of calculation pattern<(ab)>,<(ab) a>,<(ab)(ac)>,<(ab) (acd)>and<(ab) (acd) a>, their value of utility is respectively 50,22,18,24 and 34. In effective orderIn row mode excavation, they also no longer meet antimonotone characteristic. Therefore, given minimum effectiveness threshold xi > 0, effective sequence mouldFormula can't generate a complete LQS-Tree sometimes, and this is very normal.

Preferred according to the present invention, described connection, refers to: adopt I-to connect the effect of the mode being connected with S-based on father nodeGenerate the value of utility of child's node by value, the utility schedule of q-sequence is shown utility matrix, i.e. utilitymatrix, Mei GeyuanElement is all a tuple in matrix; What first value showed is the effectiveness of q-item, second value demonstration be to remove in q-sequenceRemaining effectiveness outside q-item and, that is: surplus utility, the effectiveness of the item not occurring in q-sequence is 0;

We are with the q-sequence s in table 2₄For example, Q-sequence S₄Utility matrix as shown in table 4, other sequence can adoptSame mode is calculated.

Table 4

Be recorded as example with b in table 4, first introduce I-and connect. Only have the Xiang Caineng larger than b to participate in I-connection, as formIn from c1 (q-item collection 1 c) to d3 be all can connect item. More precisely, only have in q-item collection 1 and 3 and contain b, thatC1=(6,40) to item corresponding in c3=(2,0) can be used for forming with<(bc)>the q-subsequence that mates.<(bc)>s₄In effectiveness just equal u (<b>, s₄)={-3, the effectiveness of-3} adds that the item c1=(6,40) newly increasing is to c3='s (2,0)Effectiveness, u (<(bc)>, s₄)＝{-3+6,-3+2}＝{3,-1}。

S-connects a little a little complicated, continue taking<(bc)>as example. In utility matrix, do not have again other item can with<(bc)>carrying out I-has connected. But in matrix from a1 to d3 can with<(bc)>form q-subsequence, sequence with S-connected mode<(bc) a>,<(bc) b>,<(bc) c>and,<(bc) d>be all candidate sequence.<[(b, 1) (c, 6)] (a, 3)>and<[(b, 1) (c,6)] (a, 4)>matching sequence<(bc) a all>, its effectiveness be u (<(bc) a>, s₄)={ 3+15,3+20}={18,23}. ForBased on s₄In<(bc)>generation<(bc) a>effectiveness, need to from<(bc)>know following information: c1 and c3 be coupling<(bc)>Last q-item of q-subsequence, as for which q-item coupling b, this is not so important. We claim that c1 is axle(pivot), because it is first coupling<(bc)>the q-item of end item c, same c3 is called end q-item.

Set sequence t and S, the child's of t and t maximum utility is no more than

Wherein, i is the axle of t in s, and s ' is the most left subsequence that mates t in s, i ∈ s ' and

Based on this, if the upper bound of effectiveness, the effectiveness of residual term add the most left subsequence effectiveness and, be less than ξ, soWe can stop deeper search, turn back to search procedure.

We illustrate with an example, positive and negative if the unit effectiveness of database middle term comprises, USpan soThis Pruning strategy of algorithm is unaccommodated. Given minimum effectiveness threshold value minutil is 10. Sequence<(cf)>only appear at SIn s₃. According to character 3, u_rest(f,s₃) and u (<[(c, 4) (f, 6)]>)=10-3=7's and be less than minutil, returnIts parents' node. But, true really not so, sequence u_maxThe d of (<(cf)>)=16. Therefore, if comprise negative profit in residual termProfit, USpan can not find all effective sequence patterns so, may lose some patterns. Described USpan algorithmReferring to J.F.Yin, Z.G.ZhengandL.B.Cao, " USpan:Anefficientalgorithmformininghighutilitysequentialpatterns.”inProc.ofthe18thACMSIGKDDInternationalConferenceonKnowledgeDiscoveryandDataMining2012,pp.660–668.

Therefore, we have redefined the computational process of surplus utility in utility matrix, only calculate positive profit item value. Table 5That show is the q-sequence s after redefining₄Utility matrix.

Table 5

Preferred according to the present invention, adopt width beta pruning, collect and connect item, in deletion list separately, lack of prospects, refer to: the pattern lacking of prospects for fear of generation, for scanning the width Pruning strategy of subprogram, based on weighting sequenceDownward closure property judge whether an item is promising (promising); Set a k-sequence t, one newItem is connected to t and generates (k+1)-sequence t ', if SWU (t ') >=ξ, we claim that an i is promising for t, otherwiseItem i is unpromising; The downward closure property of described weighting sequence refers to: suppose q-sequence library S and two sequencest₁And t₂, wherein t₂Comprise t₁, SWU (t so₂)≤SWU(t₁)。

Preferred according to the present invention, described HUSPNIV algorithm false code is as follows:

Input: sequence t, the effectiveness u (t) of t, the sequence library S based on effectiveness, minimum effectiveness threshold xi;

Output: all effective sequence patterns;

1) what describe is depth pruning strategy, judges whether a node is leafy node, sets sequence t and S, t and t'sChild's maximum utility is no more thanWherein, i is the axle of t in s, and s ' mates t in sThe most left subsequence, i ∈ s ' andThis result of calculation and ξ are compared to show whether this node is leafy node, asIt is less than ξ fruit, turns back to parents' node;

2) what-4) describe is the line width beta pruning of going forward side by side of scanning subprogram, once after having collected and connecting, lack of prospectsItem will be deleted from list separately;

7) what with 13) describe is to be connected to be connected with S-by I-respectively to build child, i.e. new sequence;

8) and 14) describe: after generating new sequence, cut and only comprise negative profit according to negative sequence beta pruning PNS strategyThe sequence of profit item, bears as the unit effectiveness of each in infructescence, and this sequence will be cut so, and HUSPNIV calculatesMethod is exported qualified effective sequence pattern, then recursively calls and self removes to build darker LQS-Tree.

Beneficial effect of the present invention is:

The present invention is applied in and carries out in customers buying behavior analytic process, has not only considered that the quantity purchase of commodity is askedTopic, but also considered the profit problems of commodity, filter out client in certain a period of time by minimum effectiveness threshold value and buy ratioCommodity more than more and businessman gets a profit, but also can help businessman to formulate marketing strategy, strive for more profit. Visitor like thisFamily, in the time buying commodity, utilizes the present invention to recommend some other clients can buy and and this product correlation ratio to himLarger, there are again the commodity of preferential activity simultaneously, thereby increase client's Transaction apparatus meeting, change website browsing person into buyer,Improve cross-selling ability, improve client's loyalty, and improve the economic benefit of website.

Detailed description of the invention

Below in conjunction with embodiment, the present invention is further qualified, but is not limited to this.

Embodiment 1

The application of the effective sequence pattern of the negative profit item of band in ecommerce decision-making, concrete steps comprise:

(1) definition

Definition q-item effectiveness: a q-item effectiveness is single q-item (i_j,q_j) effectiveness, be expressed as u (i_j,q_j)：

u(i_j,q_j)＝p(i_j)×q_j(Ⅰ)

In formula (I), i_jRepresent a kind of commodity; q_jBe positive number, representing commodity i_jQuantity purchase; P (i_j) refer to commodity i_jOutside effectiveness, i.e. commodity i_jUnit profit, be positive or negative; For example, the effectiveness of q-item (b, 2) is: u (b, 2)=-3×2＝-6。

Definition Q-item collection effectiveness: be q-item collection l=[(i₁,q₁)(i₂,q₂)...(i_j,q_j)...(i_n,q_n)] effectiveness, tableBe shown u (l):

$\begin{array}{cc}u\left(l\right)=\underset{j=1}{\overset{n}{\&Sigma;}}u(i_j,q_j)& \left(II\right)\end{array}$

In formula (II), 1≤j≤n; For example, q-item collection l=[(a, 4) (b, 2)] effectiveness be: u (l)=5 × 4+ (3) ×2＝14。

Sequence effectiveness in definition Q-sequence library S: suppose given sequence t=< t₁t₂...t_n>, the q-order corresponding with tRow s=< l₁l₂...l_m> utility schedule be shown u (t, s):

In formula (III), t=< t₁t₂...t_n>and s=<l₁l₂...l_m> referring to: a client is all within certain time periodThe transaction record orderly quantity series or the q-sequence that form;

The utility schedule of t in database S is shown u (t):

$\begin{array}{cc}u\left(t\right)=\underset{s\&Element;S}{\&cup;}u(t,s)& \left(IV\right)\end{array}---\left(4\right)$

In formula (IV), a q-sequence library S is one group of components<sid, s>set, sid is q-sequenceSequence_id, s is a data sequence; For example, sequence t=<ba>, at the s of table 2₄In, the effectiveness of t is u (t, s₄)＝{u(<(b, 1) (a, 3)>), u (<(b, 1) (a, 4)>) }={ 12,17}, the effectiveness of t in S is u (t)={ u (t, s₂),u(t,s₄),u(t,s₅)={ { 2,7}, { 12,17}, { 9,4}}. May there are multiple values with the q-sequence of sequences match, as t=just now <Ba >, it is at s₄In have two value of utilities, be respectively 12 and 17. This be different from Frequent Sequential Patterns a bit, frequent sequence mouldFormula only has a support value.

Definition Q-sequence effectiveness: the effectiveness u (s) of Q-sequence s refers to that unit effectiveness in s is the effectiveness summation of positive item, that is:

That for example, in table 3, show is the q-sequence s after defining₁，s₂，s₃，s₄，s₅Effectiveness.

The weighting sequence effectiveness of sequence t in definition Q-sequence library S, i.e. SequenceWeightedUtilization, is called for short SWU, is expressed as SWU (t), that is:

For example, SWU (<ca>)=u (s₂)+u(s₄)＝104。

(2) excavate effective sequence pattern

In multiple value of utilities in the q-sequence corresponding with t, choose maximum value of utility u_max(t) conduct is in this q-orderThe effectiveness of row;

$\begin{array}{cc}{u}_{max}\left(t\right)=\underset{s\&Element;S}{\&Sigma;}max\left\{u(t,s)\right\}& \left(VII\right)\end{array}$

Work as u_max(t) >=when ξ, sequence t is an effective sequence pattern; ξ is the given minimum effectiveness threshold value of user(minimumutility，minutil)；

To find all effective sequence patterns with the effective sequential mode mining of negative profit item, its unit effectPositive also comprising negatively with both comprising, if only comprise positive unit effectiveness, is the effective sequential mode mining with profit item.

For example, sequence<ba>Effectiveness be u_max(<ba>)=7+17+9=33. If minimum effectiveness threshold xi is 30, soThis sequence is effective sequence pattern, because u_max(s)＝33≥ξ。

(3) utilize HUSPNIV algorithm to excavate the effective sequence pattern of the negative profit item of the interested band of user

1. first adopt depth pruning strategy, judge whether a node is leafy node, if so, turn back to parents' knotPoint; Otherwise, enter step 2.; It is darker that the leafy node of depth pruning strategy in setting by identification avoids that algorithm walks.

2. adopt width beta pruning, collection can connect item, deletes the item lacking of prospects in list separately, then passes through respectively I-Connect and be connected structure child with S-, i.e. new sequence, after generating new sequence, only cuts and comprises according to negative sequence Pruning strategyThe sequence of negative profit item, bears as the unit effectiveness of each in infructescence, and this sequence will be cut so, excavatesThe effective sequence pattern of the negative profit item of the interested band of user, recursively calls and self removes to build darker LQS-Tree;

It should be noted that the list of the item in effective sequence pattern with the effective sequential mode mining of negative profit itemPosition effectiveness may be born, but has at least one to be positive, if entirely born, needs to cut, because there is no actual meaningJustice.

For example, suppose that B and C are the items with negative profit, if<BC>be candidate sequence,<BC>can be cut, becauseIn sequence, each is negative profit item. If contain positive profit item in candidate sequence, and the value of utility of this sequence is greater than orEqual minutil threshold value, this sequence is effective sequence pattern.

(4) utilize the effective sequence pattern of the negative profit item of the interested band of user that step (3) excavates, analyze clientBuying behavior, make seller predict later buying and selling of commodities situation according to current buying and selling of commodities situation, formulate marketingPutting of strategy, arrangement commodity.

Embodiment 2

The answering in ecommerce decision-making according to the effective sequence pattern of the negative profit item of the band described in embodiment 1With, its difference is, described LQS-Tree, i.e. Q-sequence lexicographic tree LQS-TreeT (LexicographicQ-sequenceTree, LQS-Tree), refer to: all sequences are all arranged in the structure of LQS-TreeT, the root node of LQS-TreeTFor sky, other each node is a sequence and the effectiveness of having deposited sequence, arbitrarily the child of node or be that I-connectsConnect, or be that S-connects, in LQS-TreeT, the child of node arranges by the lexicographic order increasing progressively arbitrarily; Suppose k-sequence t,The operation that the end of adding a new item to t forms (k+1)-sequence is called connection, if the size of t does not change, claimsThis I-that is operating as connects, if the size of t how one, be referred to as so S-and connect;

In Frequent Sequential Patterns excavates, closure property plays an important role in sequential mode mining downwards. ButThat this characteristic is not adapting to effective sequential mode mining. In table 2, u_max(<ea>)＝28+27＝55，u_max(<e>)=8+12+12=32, visible sequence<e>value of utility be less than the value of utility of its father's sequence. What is more important, if weCheck the maximum utility of a paths in a complete LQS-Tree, sequence of calculation pattern<(ab)>,<(ab) a>,<(ab)(ac)>,<(ab) (acd)>and<(ab) (acd) a>, their value of utility is respectively 50,22,18,24 and 34. In effective orderIn row mode excavation, they also no longer meet antimonotone characteristic. Therefore, given minimum effectiveness threshold xi > 0, effective sequence mouldFormula can't generate a complete LQS-Tree sometimes, and this is very normal.

Embodiment 3

The answering in ecommerce decision-making according to the effective sequence pattern of the negative profit item of the band described in embodiment 2Be with, its difference, described connection, refers to: adopt I-to connect the value of utility of the mode being connected with S-based on father node and generateThe value of utility of child's node, the utility schedule of q-sequence is shown utility matrix, i.e. utilitymatrix, each element is in matrixIt is all a tuple; What first value showed is the effectiveness of q-item, second value demonstration be in q-sequence except q-itemRemaining effectiveness and, that is: surplus utility, the effectiveness of the item not occurring in q-sequence is 0.

We are with the q-sequence s in table 2₄For example, Q-sequence S₄Utility matrix as shown in table 4, other sequence can adoptSame mode is calculated.

Table 4

Be recorded as example with b in table 4, first introduce I-and connect. Only have the Xiang Caineng larger than b to participate in I-connection, as formIn from c1 (q-item collection 1 c) to d3 be all can connect item. More precisely, only have in q-item collection 1 and 3 and contain b, thatC1=(6,40) to item corresponding in c3=(2,0) can be used for forming with<(bc)>the q-subsequence that mates.<(bc)>s₄In effectiveness just equal u (<b>, s₄)={-3, the effectiveness of-3} adds that the item c1=(6,40) newly increasing is to c3='s (2,0)Effectiveness, u (<(bc)>, s₄)＝{-3+6,-3+2}＝{3,-1}。

S-connects a little a little complicated, continue taking<(bc)>as example. In utility matrix, do not have again other item can with<(bc)>carrying out I-has connected. But in matrix from a1 to d3 can with<(bc)>form q-subsequence, sequence with S-connected mode<(bc) a>,<(bc) b>,<(bc) c>and,<(bc) d>be all candidate sequence.<[(b, 1) (c, 6)] (a, 3)>and<[(b, 1) (c,6)] (a, 4)>matching sequence<(bc) a all>, its effectiveness be u (<(bc) a>, s₄)={ 3+15,3+20}={18,23}. ForBased on s₄In<(bc)>generation<(bc) a>effectiveness, need to from<(bc)>know following information: c1 and c3 be coupling<(bc)>Last q-item of q-subsequence, as for which q-item coupling b, this is not so important. We claim that c1 is axle(pivot), because it is first coupling<(bc)>the q-item of end item c, same c3 is called end q-item.

Set sequence t and S, the child's of t and t maximum utility is no more than

Wherein, i is the axle of t in s, and s ' is the most left subsequence that mates t in s, i ∈ s ' and

Based on this, if the upper bound of effectiveness, the effectiveness of residual term add the most left subsequence effectiveness and, be less than ξ, soWe can stop deeper search, turn back to search procedure.

We illustrate with an example, positive and negative if the unit effectiveness of database middle term comprises, USpan soThis Pruning strategy of algorithm is unaccommodated. Given minimum effectiveness threshold value minutil is 10. Sequence<(cf)>only appear at SIn s₃. According to character 3, u_rest(f,s₃) and u (<[(c, 4) (f, 6)]>)=10-3=7's and be less than minutil, returnIts parents' node. But, true really not so, sequence u_maxThe d of (<(cf)>)=16. Therefore, if comprise negative profit in residual termProfit, USpan can not find all effective sequence patterns so, may lose some patterns.

Therefore, we have redefined the computational process of surplus utility in utility matrix, only calculate positive profit item value. Table 5That show is the q-sequence s after redefining₄Utility matrix.

Table 5

Embodiment 4

The answering in ecommerce decision-making according to the effective sequence pattern of the negative profit item of the band described in embodiment 1With, its difference is, adopts width beta pruning, collection can connect item, deletes the item lacking of prospects in list separately, refers to: forAvoid generating the pattern that lacks of prospects, for scanning the width Pruning strategy of subprogram, based on the downward sealing of weighting sequenceCharacteristic judges whether an item is promising (promising); Set a k-sequence t, a new item is connected to t lifeBecome (k+1)-sequence t ', if SWU (t ') >=ξ, we claim that an i is promising for t, otherwise an i is without futureItem; The downward closure property of described weighting sequence refers to: suppose q-sequence library S and two sequence t₁And t₂, wherein t₂Comprise t₁, SWU (t so₂)≤SWU(t₁)。

Embodiment 5

The answering in ecommerce decision-making according to the effective sequence pattern of the negative profit item of the band described in embodiment 1Be with, its difference, described HUSPNIV algorithm false code is as follows:

Input: sequence t, the effectiveness u (t) of t, the sequence library S based on effectiveness, minimum effectiveness threshold xi.

Output: all effective sequence patterns

1) what describe is depth pruning strategy, judges whether a node is leafy node, sets sequence t and S, t and t'sChild's maximum utility is no more thanWherein, i is the axle of t in s, and s ' mates t in sThe most left subsequence, i ∈ s ' andThis result of calculation and ξ are compared to show whether this node is leafy node, asIt is less than ξ fruit, turns back to parents' node;

2) what-4) describe is the line width beta pruning of going forward side by side of scanning subprogram, once after having collected and connecting, lack of prospectsItem will be deleted from list separately;

7) what with 13) describe is to be connected to be connected with S-by I-respectively to build child, i.e. new sequence;

8) and 14) describe: after generating new sequence, cut and only comprise negative profit according to negative sequence beta pruning PNS strategyThe sequence of profit item, bears as the unit effectiveness of each in infructescence, and this sequence will be cut so, and HUSPNIV calculatesMethod is exported qualified effective sequence pattern, then recursively calls and self removes to build darker LQS-Tree.

Claims (5)

1. the application of the effective sequence pattern of the negative profit item of band in ecommerce decision-making, is characterized in that, specifically stepSuddenly comprise:

(1) definition

Definition q-item effectiveness: a q-item effectiveness is single q-item (i_j,q_j) effectiveness, be expressed as u (i_j,q_j)：

u(i_j,q_j)＝p(i_j)×q_j(Ⅰ)

In formula (I), i_jRepresent a kind of commodity; q_jBe positive number, representing commodity i_jQuantity purchase; P (i_j) refer to commodity i_jOutsidePortion's effectiveness, i.e. commodity i_jUnit profit, be positive or negative;

Definition Q-item collection effectiveness: be q-item collection l=[(i₁,q₁)(i₂,q₂)...(i_j,q_j)...(i_n,q_n)] effectiveness, be expressed as u(l)：

$u\left(l\right)=\underset{j=1}{\overset{n}{\&Sigma;}}u(i_j,q_j)---\left(II\right)$

In formula (II), 1≤j≤n;

Sequence effectiveness in definition Q-sequence library S: suppose given sequence t=< t₁t₂...t_n>, the q-sequence s corresponding with t＝<l₁l₂...l_m> utility schedule be shown u (t, s):

In formula (III), t=< t₁t₂...t_n>and s=<l₁l₂...l_m> refer to: client all friendship within certain time periodOrderly quantity series or q-sequence that easily record forms;

The utility schedule of t in database S is shown u (t):

$\begin{array}{cc}u\left(t\right)=\underset{s\&Element;S}{\&cup;}u(t,s)& \left(IV\right)\end{array}---\left(4\right)$

In formula (IV), a q-sequence library S is one group of components<sid, s>set, sid is the sequence_ of q-sequenceId, s is a data sequence;

Definition Q-sequence effectiveness: the effectiveness u (s) of Q-sequence s refers to that unit effectiveness in s is the effectiveness summation of positive item, that is:

The weighting sequence effectiveness of sequence t in definition Q-sequence library S, i.e. SequenceWeightedUtilization, letterClaim SWU, be expressed as SWU (t), that is:

(2) excavate effective sequence pattern

In multiple value of utilities in the q-sequence corresponding with t, choose maximum value of utility u_max(t) as in this q-sequenceEffectiveness;

$u_{max}\left(t\right)=\underset{s\&Element;S}{\&Sigma;}max\left\{u(t,s)\right\}---\left(VII\right)$

Work as u_max(t) >=when ξ, sequence t is an effective sequence pattern; ξ is the given minimum effectiveness threshold value of user;

(3) utilize HUSPNIV algorithm to excavate the effective sequence pattern of the negative profit item of the interested band of user

1. first adopt depth pruning strategy, judge that whether a node is leafy node, if so, turns back to parents' node;Otherwise, enter step 2.;

2. adopt width beta pruning, collection can connect item, deletes the item lacking of prospects in list separately, then connects by I-respectivelyBe connected with S-and build child, i.e. new sequence, after generating new sequence, cuts and only comprises negative profit according to negative sequence Pruning strategyThe sequence of profit item, bears as the unit effectiveness of each in infructescence, and this sequence will be cut so, excavates userThe effective sequence pattern of the negative profit item of interested band, recursively calls and self removes to build darker LQS-Tree;

(4) utilize the effective sequence pattern of the negative profit item of the interested band of user that step (3) excavates, analyze purchasing of clientBuy behavior, make seller predict later buying and selling of commodities situation according to current buying and selling of commodities situation, formulation marketing strategy,Arrange putting of commodity.

2. the effective sequence pattern of the negative profit item of band according to claim 1 answering in ecommerce decision-makingWith, it is characterized in that, described LQS-Tree, i.e. Q-sequence lexicographic tree LQS-TreeT, refers to: all sequences are all arranged atIn the structure of LQS-TreeT, the root node of LQS-TreeT is empty, and other each node is a sequence and depositsThe effectiveness of sequence, the arbitrarily child of node or be that I-connects, or be that S-connects, the child of node arbitrarily in LQS-TreeTSon is all arranged by the lexicographic order increasing progressively; Suppose k-sequence t, the end of adding a new item to t forms (k+1)-sequenceOperation be called connection, if the size of t does not change, claim this I-of being operating as to connect, if the size of t how one, soBeing referred to as S-connects.

3. the effective sequence pattern of the negative profit item of band according to claim 2 answering in ecommerce decision-makingWith, it is characterized in that, described connection, refers to: adopt I-to connect the value of utility of the mode being connected with S-based on father node and generateThe value of utility of child's node, the utility schedule of q-sequence is shown utility matrix, i.e. utilitymatrix, each element is in matrixIt is all a tuple; What first value showed is the effectiveness of q-item, second value demonstration be in q-sequence except q-itemRemaining effectiveness and, that is: surplus utility, the effectiveness of the item not occurring in q-sequence is 0.

4. the effective sequence pattern of the negative profit item of band according to claim 1 answering in ecommerce decision-makingWith, it is characterized in that, adopt width beta pruning, collect and connect item, delete the item lacking of prospects in list separately, refer to: forThe width Pruning strategy of scanning subprogram, the downward closure property based on weighting sequence judges whether an item is promising; Set a k-sequence t, a new item is connected to t and generates (k+1)-sequence t ', if SWU (t ') >=ξ, we claim an iBe promising for t, otherwise an i it is unpromising; The downward closure property of described weighting sequence refers to: supposeQ-sequence library S and two sequence t₁And t₂, wherein t₂Comprise t₁, SWU (t so₂)≤SWU(t₁)。

5. the effective sequence pattern of bearing profit item according to the arbitrary described band of claim 1-4 is in ecommerce decision-makingApplication, it is characterized in that, described HUSPNIV algorithm false code is as follows:

Input: sequence t, the effectiveness u (t) of t, the sequence library S based on effectiveness, minimum effectiveness threshold xi;

Output: all effective sequence patterns;

1)ifpisaleafnodethenreturn

2)scantheprojecteddatabaseS(u(t))onceto:

3)a).putI-Concatenationitemsintoilist,or

4)b).putS-Concatenationitemsintoslist

5)removeunpromisingitemsinilistandslist

6)foreachitemiinilistdo

7)(t’,u(t’))←I-Concatenate(p,i)

8)ifeachitem’sexternalutilityint’isnotnegativethen

9)ifu_max(t’)≥ξthen

10)outputt’

11)HUSPNIV(t’,u(t’))

12)foreachitemiinslistdo

13)(t’,u(t’))←S-Concatenate(p,i)

14)ifeachitem’sexternalutilityint’isnotnegativethen

15)ifu_max(t’)≥ξthen

16)outputt’

17)HUSPNIV(t’,u(t’))

return

1) what describe is depth pruning strategy, judges whether a node is leafy node, sets sequence t and S, the child of t and tMaximum utility be no more than

Wherein, i is the axle of t in s, and s ' is the most left subsequence that mates t in s, i∈ s ' andThis result of calculation and ξ are compared and show that whether this node is leafy node, if it is less than ξ, returnsGet back to parents' node;

2) what-4) describe is the line width beta pruning of going forward side by side of scanning subprogram, once after having collected and having connected, lack of prospects justCan from list separately, delete;

7) what with 13) describe is to be connected to be connected with S-by I-respectively to build child, i.e. new sequence;

8) and 14) describe: after generating new sequence, cut and only comprise negative profit item according to negative sequence beta pruning PNS strategySequence, bear as the unit effectiveness of each in infructescence, this sequence will be cut so, HUSPNIV algorithm is defeatedGo out qualified effective sequence pattern, then recursively call and self remove to build darker LQS-Tree.