Email Classi?cation with Temporal Features
Svetlana Kiritchenko1,Stan Matwin1,and Suhayya Abu-Hakima2
1School of Information Technology and Engineering,University of Ottawa, Ottawa,ON,Canada K1N6N5
2AmikaNow!Corporation,700March Road,Suite203,Kanata,Ontario,Canada K2K2V9
Abstract.We propose a novel solution to the email classi?cation problem:the integration of temporal information with the traditional content-based classi?cation approaches.We discover temporal relations in an email sequence in the form of temporal sequential patterns and embed the discovered information into content-based learning methods.The new heterogeneous classi?cation system shows a good performance reducing the classi?cation error by up to22%.
1Introduction
As estimated by Ferris Research[5],spam accounts for15%to20%of in-bound email at U.S.-based corporate organizations.Half of users receive10 or more spam messages per day while some of them can get up to several hundreds unsolicited emails.Spam is not only distractive for end users,it also costs corporations a great deal of money.However,spam is not the only problem with emails.The overwhelming amount of electronic correspondence has become very di?cult to manage for many users.One could bene?t from a personal email managing system that?lters unsolicited messages,forwards, deletes,or saves messages into relevant folders,prioritizes incoming messages and forwards the urgent ones to mobile devices.At the company level,an au-tomatic assistant that could forward messages to appropriate sta?members or even automatically respond to general inquiries would be a considerable advantage.At the same time,the automatic systems have to make decisions very accurately without active interference from the user.
In this work we propose a novel solution to the email classi?cation prob-lem:incorporating temporal information in the traditional classi?cation ap-proaches.A lot of research has been recently carried out on email classi?cation proposing general and speci?c solutions to some of the problems mentioned above.Generally,those approaches concentrate on content interpretation and work only with content-based features.At the same time,email data have a temporal character:every email message has a timestamp.Therefore,we have a sequence of items each of which has a timestamp and content.This temporal information is always present,but usually ignored by researchers. We would like to explore the potential relevance of temporal information to common tasks in the email domain.The basic idea is to extract useful features from the temporal information present in the email data and combine these
2Svetlana Kiritchenko et al.
new features with the conventional content-based classi?cation approaches. Our belief is that since we work with a much richer information space,we could expect a signi?cant improvement in the classi?cation accuracy.
This paper presents a comprehensive study of employing temporal in-formation in email classi?cation.Various approaches have been implemented and tested on several email and newsgroups corpora.The results suggest that the proposed approach can improve the existing content-based techniques to a moderate degree.
The reminder of the paper is organized as follows.In the next section we give an overview of related work.Then,we start with some experiments on adding in time in the form of simple features.Next,we introduce our idea of representing temporal information in the form of temporal patterns.We give the de?nition of a temporal pattern,present an algorithm for discovering temporal patterns in a sequence,and describe several approaches to integrate temporal patterns into content-based classi?cation.After presenting and dis-cussing the results of our experiments,we conclude the paper with possible directions for future work.
2Related Work
In this work we integrate methods from two di?erent areas of research:email classi?cation and temporal data mining.
Numerous studies on email classi?cation appeared in the machine learning literature in the last few years(for an overview of the major ones see[9]).Most of the approaches put in use only word-based features,completely ignoring the temporal aspect of the email domain.At the same time,Sahami et al.[15] showed that bringing in other kinds of features(spam-speci?c features in their study)could improve the classi?cation results.To?lter spam messages,they join together two sets of features,traditional textual and non-textual such as over-emphasized punctuation,the domain type of the sender,time the message was received,etc.Our work explores the possibility of the usage of more complex time-related features in a general email classi?cation context.
In the data mining?eld there have been a lot of research on mining sequential patterns in large databases of customer transactions:from the ?rst Apriori-like algorithms AprioriSome and AprioriAll[2]to their numer-ous modi?cations and extensions[13,16,17].Similar approaches were pro-posed for discovering sequential patterns,also called frequent episodes,in sequences[12].Recently,more complex temporal sequential patterns were investigated.Works such as[7,8,11]consider interval-based sequences where events last over a period of time(as opposed to point-based events).In this study we work with point-based events but concentrate on time elapsed be-tween the events.We propose a new algorithm MINTS that?nds temporal sequential patterns consisting of not only event types,but also the time inter-
Email Classi?cation with Temporal Features3 vals between the events.Therefore,our method predicts not only the expected event in a sequence,but also when the event is likely to happen.
Work by Kleinberg[10]applies a temporal analysis in the context of email.His assumption is that the appearance of a new topic in a document stream is signaled by a”burst of activity”.He presents a formal framework for identifying such”bursts”and shows that the document structure imposed by this analysis has a natural meaning in terms of the content.
However,research on temporal data mining focuses only on the temporal aspect of data and does not take into account any content-based features. The main contribution of this paper is an attempt to integrate methods from the two areas in one powerful heterogeneous system.
3Email Classi?cation with Temporal Features
3.1Simple Temporal Features
The simplest way of incorporating temporal information in email classi?ca-tion is to extract temporal features such as the day of the week and the time of the day the message was received from the message timestamps.We add those new features to content-based features1and run one of the standard classi?cation algorithms.Unfortunately,our preliminary experiments showed that the overall e?ect of adding simple temporal features is not very positive for any learning algorithm that we tried.These results suggest one of the two possibilities:either temporal information is irrelevant to email classi?cation or temporal information is associated with classes in more complex ways,and a simple classi?cation algorithm cannot discover those associations from the timestamps by itself.It needs our help to represent those associations more explicitly,so that it can make use of them.To explore the second possibility we introduce a notion of temporal sequential patterns.From a temporal email sequence we discover temporal relations between the classes of messages in the form of temporal sequential patterns.Then,we transform these temporal relations into new features,combine them with content-based features,and feed them into one of the standard classi?cation algorithms.
3.2Temporal Sequential Patterns
De?nition.A set of emails can be viewed as an event sequence(c1,t1)→(c2,t2)→...→(c n,t n),where each event corresponds to an email and is represented as a pair(c i,t i)with c i∈C being the category of the email 1As content-based features we use words composing the subjects and the bodies of emails.This is the most general representation of the email content.We do not include any other header information since it usually requires some type of preprocessing.
4Svetlana Kiritchenko et al.
(the event type)and t i being the timestamp of the email.The events in the sequence are ordered by their timestamps:t1≤t2≤...≤t n.
A temporal sequential pattern is an ordered sequence of event types c1→c2→...→c k along with an ordered sequence of time intervals d1→d2→...→d k?1denoted as{c1→[d1]→c2→[d2]→...→c k?1→[d k?1]→
c k}.The interpretation of pattern{c1→[d1]→c2→[d2]→...→c k?1→
[d k?1]→c k}is the following:in a given event sequence there is a subse-quence2(c1,t1)→(c2,t2)→...→(c k,t k),where t1≤t2≤...≤t k and t2?t1=d1,t3?t2=d2,...,t k?t k?1=d k?1.
Mining Temporal Sequential Patterns.Our algorithm for mining tem-poral sequential patterns is based on the ideas?rst presented in the classical Apriori algorithm[1].Apriori was initially designed to mine frequent3intra-transaction patterns in a database of customer transactions.It works itera-tively producing an(n+1)-item pattern from a pair of n-item patterns that have the same pre?x4.The observation that for any pattern to be frequent all its sub-patterns have to be frequent too makes the algorithm more e?cient. So,before scanning the database to count the number of occurrences for a new(n+1)-item pattern we could check if all of its n-item sub-patterns are frequent.
To accommodate the temporal aspect of an email sequence we designed a new algorithm MINTS(MINing Temporal Sequential patterns)to mine sequential patterns along with the exact time intervals between the events in the patterns.Although there exist many extensions of the classical Apriori algorithm to deal with sequential data,to the best of our knowledge none of them deals with temporal sequences taking into account the time elapsed between the events.To add time intervals,two key modi?cations have to be made to the Apriori-like algorithm.The?rst one is when we combine pairs of1-event patterns to make2-event patterns,we have to consider every possible time interval between the events.The second modi?cation is when we combine a pair of n-event patterns to produce an(n+1)-event pattern the n-event patterns have to have the same pre?x,which includes not only the same event types,but also the same time intervals between the events.These modi?cations,especially the?rst one,signi?cantly increase the computational complexity of the algorithm.
We also have to revise the criteria by which we choose interesting patterns. Traditionally,a pattern is considered interesting if its support and con?dence exceed user-de?ned thresholds.Generally,support is de?ned as the number of occurrences of the pattern in the data,and con?dence is de?ned as the ratio 2The elements of a subsequence are not necessarily consecutive elements of the event sequence.
3A pattern is called frequent if we can?nd at least minS its occurrences in the data.minS is a user-de?ned threshold called minimum support.
4A pre?x of an n-item pattern is the?rst(n-1)items in the pattern.
Email Classi?cation with Temporal Features5 of the support of the pattern and the support of its pre?x.These conven-tional measures have several drawbacks,especially in the temporal context. First,support does not take into account the a priori probabilities of event types.Patterns formed by frequent event types would naturally occur more frequently than patterns formed by rare event types.Therefore,selecting patterns by their support can result in missing highly predictive patterns consisting of rare event types and?nding a large number of useless patterns formed by frequent event types.Second,the con?dence of a temporal pattern no longer represents the predictive power of the pattern.Consider the follow-ing example.In a given sequence we have20occurrences of event type c1and 10occurrences of temporal pattern c1→[d]→c2.Thus,the con?dence of the pattern c1→[d]→c2is50%.That means that after an event of type c1an event of type c2would happen in d time units with50%probability. However,if in the given sequence no other event types ever happen in d time units after an event of type c1,the probability of an event happened in d time units after c1to be of type c2is100%.Since we are interested in predicting the type of a new event knowing when it happened and,therefore,knowing the exact time distances between the new and previous events,we prefer a measure that would be related to the probability of an event encountered at a time point to be of a particular class and not to the probability to encounter an event of a particular class at the given time point.
To accommodate those drawbacks,we suggest an alternative measure of the predictive power of a pattern.We consider a pattern to be interesting if it occurs much more frequently than we would expect based on the frequency of the event types composing the pattern and the length of the intervals between the events.So,for each pair of event types we scan the sequence to?nd all time intervals that occur between events of those types and count the number of occurrences(n)for each time interval.Then,we would consider only those intervals for which n is signi?cantly greater than we would expect assuming the independence of events.In other words,if the probability to encounter the pattern n times is very small,according to the Poisson distribution,then we consider it interesting and keep it;otherwise,we discard it.
What kinds of patterns can one expect to?nd?Among the patterns that we discovered,some patterns seem to occur by chance,but some are quite interesting.For example,we can?nd a pattern like{c i→[d j]→c i},which is interpreted as events of type c i occur regularly each d j time units(for example,once a week).These types of patterns are common for subscription letters.Another interesting pattern is{c i→[0hours]→c i},which says that the events of type c i occur in bunches(or at least in pairs).We also can?nd patterns like{c i→[0hours]→c i→[0hours]→c i},{c i→[0hours]→c i→[0hours]→c i→[0hours]→c i},etc.,that specify how many events of type c i usually occur within a few minutes from each other.These patterns are common for some public mailing lists like MLNet.Most of the patterns,like the ones described above,consist of one event type.Yet,in some datasets we
6Svetlana Kiritchenko et al.
can?nd patterns with several event types,which suggests that those event types are related.For example,messages from the QA department to software engineers regarding some problems with a software product could be followed by an email discussion of those problems among the engineers.
Sequence Prediction.Although temporal patterns give us insights into the nature of mechanisms generating the sequence,the main application of patterns is to predict the type of a new event in the sequence.To do so,we look for patterns that are present at the end of the sequence.If for pattern {c1→[d1]→c2→[d2]→...→c k?1→[d k?1]→c k}we?nd a subsequence corresponding to its pre?x{c1→[d1]→c2→[d2]→...→c k?2→[d k?2]→c k?1},then the instance found in d k?1time units after the end of the subse-quence is expected to have type c k.We say that the pattern is applicable to that instance.
3.3Integration of Temporal and Content-based Features Temporal patterns alone will not provide us with good prediction results, but we can integrate them with content-based features,expecting that the temporal characteristics of data would enrich the information space and allow us to obtain better classi?cation performance.In this work we propose six simple methods of combining temporal and content-based information.
The following three methods transform the temporal information into new features and add them to the set of content-based features:
?”Predicted Class”
Based on the temporal patterns applicable to the instance in question, we can predict the class of that instance(see section3.2).We add the predicted class to the feature set and train a classi?er.This predicted class presents our expectations for the next message to be of a particular class.
Certainly,our expectations cannot be very reliable;they are just guesses.
We anticipate that the classi?cation system would learn for which sets of feature values it could rely on those guesses and for which it could not.?”Cascading”
This method is similar to one presented in[6].Instead of adding the class of an instance predicted by the temporal patterns,we add the probabil-ities of classes predicted by the temporal patterns.There can be several patterns applicable to the instance in question,and they can predict di?erent classes.Based on pattern unexpectedness we calculate the prob-abilities for each class to be the class of the instance in question and add those probabilities to the feature set.
?”Patterns”
Here we add all temporal patterns applicable to the instance in question to the feature set.We anticipate that the classi?cation system would learn on which patterns and in what context it could rely.
Email Classi?cation with Temporal Features7 The next two methods combine the predictions of two classi?ers,content-based and time-based:
?”Simple Vote”
This method implements the simple voting scheme.First,we train a classi?er5on content-based features(a content-based classi?er)and run it on the test set getting a set of class probabilities for each test instance.
We also run our sequence prediction algorithm(a time-based classi?er) on the same test set getting the second set of class probabilities.Then, we compare the probability sets and for each test instance pick the class with the greatest probability value.
?”Replacing Uncertainties”
As in”Simple Vote”we have two classi?ers(content-based and time-based)and two sets of class probabilities for each test instance.Since the content-based classi?er is much more accurate than the time-based one,we pick the class predicted by the content-based classi?er unless its probability is less than a certain threshold.If it is the case we pick the class predicted by the time-based classi?er.
The last method is speci?c for the Naive Bayes classi?er:
?”Replacing A Priori Probabilities”
In the well-known Bayes’formula
P(Class|Doc)=P(Class)?P(Doc|Class)
P(Doc)
the term P(Class)represents the a priori class probabilities.These prob-abilities show how frequent the classes are and how probable is for a class to be the class of an instance without any other knowledge about the instance.We can change the a priori probabilities with respect to our temporal expectations.For example,if the most frequent class is c1but the temporal patterns suggest the class c2to be the class of the next in-stance,then we put the a priori probability of c2greater than the a priori probability of c1.We do it by simply replacing the a priori probabilities with the class probabilities predicted by the temporal patterns.
3.4Results
We have tested the proposed approaches on6datasets described in Table1.
Corpus Email I consists of about1500personal emails of one of the au-thors for a period of1.5years.Messages were divided into15categories by 5Any classi?cation algorithm producing the class probabilities as the output can be used here.In these experiments we used only Naive Bayes since the imple-mentations of Decision Trees and SVM that we have chosen cannot output the probabilities.
8Svetlana Kiritchenko et al.
Table1.Characteristics of the datasets used.
Name Description#of classes Size of the Size of the Vocabulary
training set test set size Emails I Personal email1510554523119 Emails II Personal email8250710753251 Emails III”Corporate”email516377022148 News I comp.ai
comp.ai.alife
comp.ai.doc-analysis.misc514256104649
comp.ai.games
comp.ai.philosophy
News II alt.folklore.music
alt.gothic.music
alt.music.bee-gees55971491908
alt.music.beethoven
alt.music-dire-straits
News I+II News I&News II joined1019478345569 their content and senders.Corpus Email II is a much bigger set of personal emails divided into8categories.These categories share much of the vocabu-lary,which makes them hard to distinguish for a conventional content-based classi?er.Corpus Email III consists of messages sent to a company’s general email address and models the situation where an automatic assistant has to redirect the incoming messages to the appropriate sta?members.
We have also used the newsgroups messages,which are close in structure and nature to email messages.For News I corpus we have chosen5AI-related newsgroups and collected all the postings in these groups for about3months. This corpus is very hard to classify since the groups are very close in content and vocabulary.News II corpus is smaller and easier for classi?cation;the categories are more distinct.Corpus News I+II is just two previous corpora merged together.
For each dataset we have trained and tested the standard(”Words only”) and three modi?ed(”Predicted Class”,”Cascading”and”Patterns”)ver-sions of three learning algorithms,namely Decision Trees(C4.5)[14],Naive Bayes[3]and Support Vector Machines(SVM)[4].We have also run three other modi?ed versions(”Simple Vote”,”Replacing Uncertainties”and”Re-placing A Priori Probabilities”)of the Naive Bayes algorithm.Table2sum-marizes the results of these experiments6.Figure1shows the performance of the three learning algorithms and their three modi?cations with varying vocabulary size(the number of content-based features used).
The?rst observation of the results is that we have been able to reduce the classi?cation error for each dataset tested:the reduction in error for emails is up to10%,for newsgroups it is up to22%.The second observation is that no algorithm is a clear winner.From Figure1we can see that”Patterns”modi-?cation of SVM and Naive Bayes was the best algorithm for most of the vo-6In these experiments the words were stemmed,and the stems appearing in less than6documents along with the stop-words were removed.
Email Classi?cation with Temporal Features
9Table 2.Classi?cation accuracies.Dateset Naive Bayes
Decision Trees (C4.5)
Support Vector Machines (SVM)words pred.casca-pat-simple repl.
repl.words pred.casca-pat-words pred.casca-pat-only class ding terns vote uncert.
a priori only class ding terns only class ding terns prob.emails I
82.5282.7478.9883.1982.3082.7482.9681.680.381.281.078.5473.2378.5480.31emails II
49.3049.3051.9154.4248.9349.7753.9551.252.051.451.256.0049.9560.5659.44emails III
65.8165.6765.2465.6765.8166.3865.6742.242.242.242.264.8164.9664.9665.10news I
66.7266.8967.2166.8966.7274.166.8964.864.466.666.671.6459.5169.1862.95news II
82.5583.2283.8982.5582.5585.9181.8891.391.391.390.687.9254.3689.2689.93news I+II 67.5167.5167.2767.5167.5171.7067.6370.573.072.171.973.2635.7371.7066.31 72 73
74
75
76 77 78 79
80
81
0 5000 10000 15000 20000
a c c u r a c y (%)vocabulary size C4.5
words only pred. class cascading patterns 72 74 76 78 80 82 84 0 5000 10000 15000 20000a c c u r a c y (%)vocabulary size Naive Bayes words only pred. class cascading patterns 64 66 68 70 72 74 76 78 80 0 5000 10000 15000 20000a c c u r a c y (%)vocabulary size
SVM words only pred. class cascading patterns Fig.1.Performance of the three algorithms and their three modi?cations with varying vocabulary size.
cabulary sizes.At the same time,this modi?cation of Decision Trees worked very poorly for larger vocabulary.”Predicted class”performed pretty well with Decision Trees and Naive Bayes but extremely poorly with SVM.Sur-prisingly,the simple ”Replacing Uncertainties”modi?cation of Naive Bayes worked very well for all newsgroup datasets reducing the error by 13-22%.
We noticed that Naive Bayes and SVM performed better with the addition of temporal information on more datasets than Decision Trees.We attribute this to the fact that Decision Trees are univariate,which means that they can test only one attribute at a time.Temporal characteristics can be in quite complex dependencies with content-based features.When tested alone they are not much of help.They seem to do better when combined with content in more sophisticated ways such as in Naive Bayes and SVM.
4Conclusion and Future Work
In this paper we present a novel approach of combining temporal and content-based features into a heterogeneous classi?cation system.We apply this ap-proach to the email domain.We also present a new algorithm MINTS to dis-cover temporal relations in the form of temporal sequential patterns.These temporal relations are then embedded into the conventional content-based learning methods.The performance of the heterogeneous classi?cation sys-tem has been tested on several email and newsgroup datasets.The system has been able to reduce the classi?cation error by up to 22%.
10Svetlana Kiritchenko et al.
An email system is a dynamic environment.Typically,the variety of email categories and their meaning are altered frequently:some categories become unused,new ones emerge.Temporal patterns change over time.Currently,we do not have any mechanisms to accommodate those changes.In the future, we have to consider a possibility of the evolutionary adaptation of the system to handle the changing context.
The proposed algorithm is fairly general and can be applied to other do-mains that are characterized by time-related and content-related features. In fact,email is not extremely time dependent,and,yet,we were able to achieve reasonable improvements there.With domains more heavily depen-dent on time we would expect even greater improvements.As our future work we plan to investigate the possibilities of applying our approach to other do-mains.
We can view each email message as an event generated by some external event such as a product release,a conference announcement,approaching holidays,etc.An external event”causes”a temporal pattern to emerge,and then the messages of a particular class are generated according to the pattern. In this work we analyzed only the temporal patterns themselves without looking at the external events.Because the nature of external events varies from national events to minor personal events,it is not clear how to retrieve them automatically.However,including these events into our analysis could result in greater improvements.
Acknowledgements
This work was supported by Natural Sciences and Engineering Research Council of Canada,Communications and Information Technology Ontario, and AmikaNow!Corporation(https://www.doczj.com/doc/a111594780.html,).
References
1.Agrawal,R.,Srikant,R.(1994).Fast Algorithms for Mining Association Rules.
Proc.of the20th Int.Conf.on Very Large Data Bases(VLDB),487–499.Morgan Kaufmann.
2.Agrawal,R.,Srikant,R.(1995).Mining Sequential Patterns.Proc.of the11th
Int.Conf.on Data Engineering(ICDE),3–14.IEEE Computer Society Press.
3.Borgelt,C.(2002).Bayes,version2.7.
http://fuzzy.cs.uni-magdeburg.de/?borgelt/software.html#bayes.
4.Chang,C.-C.,Lin,C.-J.(2001).LIBSVM:a library for support vector machines.
https://www.doczj.com/doc/a111594780.html,.tw/?cjlin/libsvm/.
5.Ferris Research(2003).Spam Control:Problems&Opportunities.
http:://https://www.doczj.com/doc/a111594780.html,.
6.Gama,J.(1998).Combining classi?ers by constructive induction.Proc.of the
10th European Conf.on Machine Learning(ECML),178–189.Springer.
Email Classi?cation with Temporal Features11 7.H¨o ppner,F.,Klawonn,F.(2002).Finding Informative Rules in Interval Se-
quences.Intelligent Data Analysis,6,237–255.
8.Kam,P.,Fu,A.W.(2000).Discovering Temporal Patterns for Interval-based
Events.Proc.of the2nd Int.Conf.on Data Warehousing and Knowledge Dis-covery(DaWaK),317–326.Springer.
9.Kay,J.,McCreath,E.(2001).Automatic Induction of Rules for E-Mail Classi?-
cation.UM2001:8th Int.Conf.on User Modeling,Workshop on User Modeling, Machine Learning and Information Retrieval.
10.Kleinberg,J.(2002).Bursty and Hierarchical Structure in Streams.Proc.of
the8th ACM SIGKDD Int.Conf.on Knowledge Discovery and Data Mining (KDD),91–101.
https://www.doczj.com/doc/a111594780.html,xman,S.,Unnikrishnan,K.P.,Sastry,P.S.(2002).Generalized Frequent
Episodes in Event Sequences.8th ACM SIGKDD Int.Conf.on Knowledge Dis-covery and Data Mining,Workshop on Temporal Data Mining.
12.Mannila,H.,Toivonen,H.,Verkamo, A.I.(1995).Discovering Frequent
Episodes in Sequences.Proc.of the1st Int.Conf.on Knowledge Discovery and Data Mining(KDD),210-215.AAAI Press.
13.Pei,J.,Han,J.,Mortazavi-Asl,B.,Pinto,H.,Chen,Q.,Dayal,U.,Hsu,M.-
C.(2001).Pre?xSpan:Mining sequential patterns e?ciently by pre?x-projected
pattern growth.Proc.of the17th Int.Conf.on Data Engineering(ICDE),251–225.IEEE Computer Society Press.
14.Quinlan,J.R.(1992).C4.5:Programs for Machine Learning.Morgan Kauf-
mann.
15.Sahami,M.,Dumais,S.,Heckerman,D.,Horvitz,E.(1998).A Bayesian Ap-
proach to Filtering Junk E-Mail.Proc.of the AAAI Workshop on Learning for Text Categorization.
16.Srikant,R.Agrawal,R.(1996).Mining Sequential Patterns:Generalizations
and Performance Improvements.Proc.of the5th Int.Conf.on Extending Database Technology(EDBT),3–17.Springer.
17.Zaki,M.(2001).SPADE:An E?cient Algorithm for Mining Frequent Se-
quences.Machine Learning,42(1-2),31–60.
email营销邮件模板 篇一:电子邮件营销之优秀邮件模板制作秘诀 如何制作一个讨人喜欢的邮件模板，让顾客眼前一亮,是众多邮件营销人员头痛的事。因为水能载舟，亦能覆舟。好的EDM邮件模板可以轻松引导顾客注意你希望表达的重要信息，反之，将导致他们觉得不知所云，让整个电子邮件营销前功尽弃。 U-Mail邮件群发平台总结发现，制作一套优秀的电子邮件模板的秘诀，重点遵从以下规则: 一、格式编码注意事项: 1.邮件内容页面宽度在600到800px(像素)以内，长度1024px内 2.电子邮件模板使用utf-8编码 3.源代码在10kb以内 4.用Table表格进行布局 5.邮件要求居中时，在table里设定居中属性 6.文件不能由WORD直接转换而成 7.不能有外链的css样式定义文字和图片 8.不要有动态形式的图片 1 9.没使用<table</table以外的body、meta和html之类的标签 10.背景图片代码写法为:<table background=background.gif cellspacing=0 cellpadding=0 11.没有出现onMouseOut onMouseOver 12.没使用alt
13. font-family属性不为空 二、电子邮件模板文字规则: 1.邮件主题在18个字以内，没有――～??等符号 2.邮件群发主题和内容中不存在带有网站地址的信息 3.文字内容、版面简洁，主题突出 4.没有敏感及带促销类的文字 5.本次模板发送不超过20万封 三、电子邮件模板图片规则: 1.尽量使用图片，减少文字量 2.整页图片在6张以内，每张图片最大不超过15kb 3.图片地址不是本地路径 4.图片名称不含有ad字符 5.整个邮件模板不是一张图 四、电子邮件模板链接规则: 1.链接数量不超过10个 2 2.链接是绝对地址 3.链接地址的长度不超过255个字符 4.不使用map功能 5.制作了一份和邮件内容一样的web页面，在邮件顶部写: “如果您无法查看邮件内容，请点击这里”，链接指向与邮件营销内容一致的web网页。 篇二:英语外贸邮件模板大全, 各类正式外贸电子邮件开 头模板, 英文商务回复跟进email模板
1.营销的计划和目的 2017年是邮件营销突飞猛进的一年。社交媒体和移动互联网发展迅速，传统的邮件营销方式在与微博、微信、QR 码、人人网等社交网络方式的积极融合中，展现了新的营销活力，邮件营销的可拓展性得到延伸，我们企业把邮件营销列为重要营销工具之一。为推广我们名裳的服装在线营销，吸引广大客户的加入和购买我们制定了一系列计划： 1.先整理出要发送的主题和内容，配置图文最好。 2.甄选客户，最好挑一些有代表性的，可以提供建议的，新老客户都可以。 3.发送邮件，一定要确保邮件到收信人手中。 4.初步统计邮件的开启率及邮件内容链接的跟踪率‘ 5.做好统计记录，这次邮件发送出现的问题及时改进。 目的：1、提升公司及产品品牌知名度和认知度 2、建立庞大的客户信息邮件数据库，可以有计划有组织地开展营销活动 3、与客户建立良好的关系，提高客户的品牌忠诚度 4、让客户第一时间获取公司及相关行业内信息 5.能够探测市场, 找到新的市场机会和提供新产品与新服务 6、可实现一对一的销售模式 7、必须将宣传对象引到网站来，因为网站才能提供详尽的信息，才更有说服力。 2.如何获取邮件列表 1.发动注册会员收集邮件列表 邀请了人同时在网站注册，初次购买可以享受折扣、送积分等等，要顾及注册会员的利益。获得的邮件列表，再拿来群发邮件。 2.做广告吸引用户主动订阅 把邮件地址印在名片、文具、发/票、传真、印刷品上，即使在广播电视广告上也留下邮件地址，这样，便于顾客通过邮件和你联系， 3.在报刊上发布新闻 在报刊上发表人们必须通过邮件才可以接收的免费报告，或者可以通过邮件发送的产品，可以发表一些引起读者共鸣的话题，在读者回应的过程中收集其姓名和邮件地址。 4.现有客户的邮件列表 下次与客户通过邮寄或邮件联系时，可以为通过邮件回复者提供特别的服务，例如一份免费报告或者特别折扣优惠等，你必须尽可能地得到顾客的邮件地址。5.购买邮件列表 现在有很多的公司有邮件列表买，很多人也是选择的这个方法，如果选择的渠道比较好的话，收效也是相当快的，当然花费的费用也不低。这些数据一般都是有比较详细的数据，不单单是一个邮箱，有性别、年龄、职业、地区等，购买的时候最好先分析一下对自己产品感兴趣的人员的一些特征，然后筛选后购买，这样相关性会高很多，自己以后群发邮件的时候效果也会好很多。当然购买邮件列表最好能够选择一个好一些的机构，信誉要高些的，否则大部分邮件地址都无效，也就白白浪费钱了。 6.做好邮件列表管理
三星手机email营销策划方案篇一：三星手机营销策划书 XX年04月 附件： 《三星智能手机营销策划书》 目录 一、前言 ................................................ ................................................... 错误！未定义书签。 二、界定问题 ................................................ .......................................... 错误！未定义书签。 1.策划主题 .............(来自: 小龙文档网:三星手机email营销策划方案)................................................ ......................... 错误！未定义书签。 2.策划目标 ................................................ ...................................... 错误！未定义书签。
三、环境分析 ................................................ .......................................... 错误！未定义书签。 1.三星集团简介 ................................................ .............................. 错误！未定义书签。 2.国内智能手机市场分析 ................................................ ............. 错误！未定义书签。 3.三星智能手机i9100与竞争对手分析 .................................... 错误！未定义书签。 4.三星智能手机宏观环境分析 ................................................ ..... 错误！未定义书签。 四、问题点和机会点 ................................................ .............................. 错误！未定义书签。 1.优势（strengths） ..................................... .............................. 错误！未定义书签。
电子邮件营销解决方案 一、邮件营销的目的 1、提升品牌知名度和认知度 2、建立庞大的会员数据库，可以有计划有组织地开展营销活动 3、与客户建立良好的关系，提高客户的品牌忠诚度 二、邮件营销的对象 1、对公司或产品感兴趣的客户，尽量找准目标客户，提高邮件发送的准确率 2、购买过公司产品的客户，适时发送邮件关怀，提高客户的粘着度 三、软件选择——客户精灵 从淘宝网卖家后台—我购买的服务，进入客户精灵软件 该软件功能包括：1导入客户，可按时间段导入客户；2客户管理，可按下单次数、交易金额、交易时间等对客户进行筛选；3下载客户：可下载客户手机号和详细信息；4邮件营销：可建立邮件模板，实施营销任务；5智能邮件：可建立售后关怀邮件、休眠客户激活邮件、催付款邮件 四、邮件营销的具体过程 1、收集客户资料 操作：1）点击“导入客户”>“导入3个月内客户”，选择相应的分组导入。
“导入三个月前客户”，需要先登录淘宝后台，“我是卖家”>“已卖出的宝贝”，选择时间段生成CSV格式订单报表，再将表格导入到相应的客户分组中。 2）当导入客户订单后，点击“客户管理”，根据需要进行客户筛选查询 点击客户列表后面的“修改”，可以修改客户资料，点击“交易”，可以查看该客户购买的产品详情 3）当筛选出目标客户后，可以将该部分客户列表转移到其它分组下 4）进入管理客户分组页面，可以任意增加和删除分组 2、确定投放时间 为了避免客户收到邮件时活动时间已经过期，所以必须在活动前三天，最好是提前五天就要发送邮件。根据有关调查，在上午八九点时发出邮件的打开率最低，在中午十二点到下午一点钟发送的邮件打开率最高，因为人们在午饭时间通常会看看邮件放松一下。并且星期一和周末都是公认最不适合群发邮件的时间，周中是最明智的，所以在周二、周三、周四中午或下午发送的邮件产生的效果最好。
许可Email营销策划书 实验一：认识许可Email营销 1、E-mail营销有何特点，市场如何细分？ 答： E-mail营销是在用户事先许可的前提下，通过电子邮件的方式向目标用户传递有价值信息的一种网络营销手段。 首先,随着互联网的迅猛发展,我国及全球网民数量不断上升.这 就为E-mail营销提供了一个很好的平台.凭借这一优势,E-mail营销可以将广告信息发送到全球各个地方的网民面前,使自己的广告信息传播到更广的范围;其次,E-mail营销主要利用电子邮件来进行,这种方法操作起来方便快捷,不需要太多的专业知识就可以完成,且效率 极高;然后,E-mail营销通过互联网进行,这就大大节省了花费,价格十分低廉,适合各种规模的企业;最后,E-mail营销的应用范围很广泛,对广告信息没有过多的限制,适用于各行各业. E-mail营销市场可以根据网民的年龄,文化水平,消费能力和购买欲望来进行细分,满足不同消费人群的需求. 2、你会打开所有的营销邮件吗？还是有针对性的打开某些需要的营销？
3、试分析 E-mail营销的优势和局限性？ 答：优点： 1.成本低廉。电子邮件营销只要有邮件服务器，联系10个用户与联系上万个用户的成本几乎没什么区别。 2.快速。电子邮件营销只要有邮件数据在手，发送邮件几小时后就可看到效果。 3.精准。邮件只要能投放精准，客户转化率还是非常高的，远 比其他营销方式高局限性： （1）应用条件限制。由于接收条件的局限，电子邮件需要一定的上网设备才可以接收和阅读，不像传统信函那样可以随时随地的查看。 （2）市场环境不成熟。 （3）邮件传输限制：由于受到网络传输速度、用户电子邮箱空间容量等因素的限制，并不是什么信息都可以通过Email来传递。
Email营销之系列邮件营销活动策划 原文：360email 邮件营销是企业或个人重要网络营销手段之一，不同行业、不同的产品实施的邮件营销手段不同，所以想要得到最佳的Email营销效果，需要选择正确的EDM营销策略。 规划整体营销活动 对于系列邮件营销活动来说，首先需要对此次邮件营销设定一个计划目标，即要取得何种效果，想要发送几期以及每期活动的主题。如第一期Email营销主题为新品推广，那么主题与内容要紧扣其中。 收集目标用户数据 另外策划内容饱满的邮件策划活动与目标用户优化也很重要，一定要仔细分析产品目标用户群，然后使用订阅表单、网站注册等方法去收集目标用户群地址数据，最好是基于用户许可订阅的邮件地址。 分析用户群体喜好 要根据产品卖点、特点等具体情况来分析目标用户群，如目标用户年龄层分类、购买力
分析、用户上网浏览习惯与喜好。 确定营销内容撰写 明确产品或服务卖点、用户为什么要选择你、需要用户做什么、产品或服务能够帮助到用户什么，这就是一个卖点。可以在邮件的页头、页尾有一个明显号召，页头可以使用高清大图，将具有吸引力丶冲击力的广告语，吸引用户点击，页尾可以使用图片按钮、文字让用户点击。宣传内容不需过于复杂，简洁的将重点展示出来。 邮件营销时间把握 如节假日可以提前三个月就开始策划丶实施。前两个月主要是为这次节假日邮件活动造势，告诉大家有一个什么样的活动。节假日前一个月需将活动的重点突显出来，以在活动期间获得最大的营销效果。在网络营销里，节假日活动还可以利用非邮件营销以外的营销方法来提升活动的曝光率，以将活动效果最大化。 EDM营销前期策划工作非常重要，对后期具体实施起到至关重要的作用，有助于后期 围绕策划方案进行实施，并有目标、有计划的操作。 美文欣赏 1、走过春的田野，趟过夏的激流，来到秋天就是安静祥和的世界。秋天，虽没有玫瑰的芳香，却有秋菊的淡雅，没有繁花似锦，却有硕果累累。秋天，没有
email 营销邮件模板 篇一: 电子邮件营销之优秀邮件模板制作秘诀 如何制作一个讨人喜欢的邮件模板，让顾客眼前一亮 , 是众多邮件营销人员头痛的事。因为水能载舟，亦能覆舟。好的EDM W件模板可以轻松引导顾客注意你希 望表达的重要信息，反之，将导致他们觉得不知所云，让整个电子邮件营销前功尽弃。 U-Mail 邮件群发平台总结发现，制作一套优秀的电子邮件模板的秘诀，重点遵从以下规则: 一、格式编码注意事项 : 1.邮件内容页面宽度在600到800px（像素）以内，长度1024px内 2.电子邮件模板使用 utf-8 编码 3.源代码在 10kb 以内 4.用 Table 表格进行布局 5.邮件要求居中时，在 table 里设定居中属性 6.文件不能由WOR直接转换而成 7.不能有外链的 css 样式定义文字和图片 8.不要有动态形式的图片 1 9.没使用<table</table 以外的body、meta和html之类的标签 10.背景图片代码写法为 :<table background=background.gif cellspacing=0 cellpadding=0 11.没有出现 onMouseOut onMouseOver 12.没使用 alt 13.font-family 属性不为空
二、电子邮件模板文字规则 : 1.邮件主题在18个字以内，没有一一?？?等符号 2.邮件群发主题和内容中不存在带有网站地址的信息 3.文字内容、版面简洁，主题突出 4.没有敏感及带促销类的文字 5.本次模板发送不超过 20 万封 三、电子邮件模板图片规则 : 1.尽量使用图片，减少文字量 2.整页图片在 6张以内，每张图片最大不超过 15kb 3.图片地址不是本地路径 4.图片名称不含有 ad 字符 5.整个邮件模板不是一张图 四、电子邮件模板链接规则 : 1.链接数量不超过 10 个 2 2.链接是绝对地址 3.链接地址的长度不超过 255 个字符 4.不使用map功能 5.制作了一份和邮件内容一样的 web页面，在邮件顶部写：“如果您无法查看邮件内容，请点击这里”，链接指向与邮件营销内容一致的web网页。 篇二：英语外贸邮件模板大全 , 各类正式外贸电子邮件开头模板 , 英文商务回复跟进 email 模板 英语外贸邮件模板大全各类正式外贸电子邮件开头模板 , 英文商务回复跟进 email 模板
Email电子邮件营销方案 电子邮件营销（也称为Email营销或者EDM）广泛应用于新客户的获取、客户价值的提升、客户忠诚度的维护等客户生命周期的各个阶段，可以发送促销信息、销售目录、新闻列表、市场调查、市场推广活动信息等。 根据《中国互联网络发展状况统计报告》，在用户经常使用的网络服务中，电子邮箱以92.6%的比例高居榜首，可以预见Email营销前景可观。但很多企业对于电子邮件营销的理解往往停留为简单的“名单+发送”，没有对邮箱地址进行分类，没有对邮件内容和主题进行统一规划，也没有采集营销数据和深度挖掘数据，等等。 那么企业如何实施有效的电子邮件营销呢？ 一、建立邮件列表数据库 ?公司积累的客户资料：包括通过公司网站注册、邮件列表订阅、线下业务往来，等。 ?数据库租赁：市场上有不少开展这项业务的数据租赁公司，也可以和相关网站合作，选定几家与自身目标客户重合度高的，在业内有一定知名度的咨询网站，他们往往有着大量的用户数据资源。 ?采用特定程序，在网络上爬取邮箱地址：这种方法几乎无成本投入，然效果也是最差的，因为无法界定爬取来的数据的有效性和真实性。
企业应当树立起构建企业用户数据库的理念，争取让第一种来源的数据资源加大在邮件列表库中的比重，并逐渐丰富完善用户行为以及喜好的营销数据库。由于邮件列表数据库具有高度的目标针对性及忠诚度，这决定了邮件列表是企业最珍贵的无形资产之一。 二、明确Email营销目的 Email营销的目的要与企业的营销战略一致。在开展邮件营销活动前，考虑前后活动之间的关系。是为了推广品牌形象，还是具体的产品和服务，还是维护客户关系，还是拓展新客户？这是一个相对独立的邮件营销活动？或者是一个长期网络营销活动的一个组成部分？定义好活动的目的性，为整体网络营销提供有力的保障。 并且需要为此次Emai营销制定合理的营销指标：如点击转化率是多少？ ROI 是多少？此次活动在整体市场活动中的期望效果？ 三、选取Email营销平台 ?专业级的邮件平台（数据营销）公司：设有专业的邮件发送平台，发送速度快，可以实现每小时几千封到几十万封的发送速度。 同时与主要的邮件服务提供商（ISP）签订协议，对于此平台来的邮件予以放行，所以，通过这类平台发送的邮件，到达率可以保持比较高的水平。 ?小型邮件发送公司：主要集中在本土EDM服务商。他们往往拥有几台服务器，几个IP地址，几个人，靠着这几台服务器大量发送邮件。这种公司发送邮件的成本非常低廉，往往几百元钱就可
什么是邮件营销？ 邮件营销是在用户事先许可的前提下，通过电子邮件的方式向目标用户传递有价值信息的一种网络营销手段。Email营销有三个基本因素：用户许可、电子邮件传递信息、信息对用户有价值。三个因素缺少一个，都不能称之为有效的邮件营销。是利用电子邮件与受众客户进行商业交流的一种直销方式。同时也广泛的应用于网络营销领域。邮件营销是网络营销手法中最古老的一种，邮件营销比绝大部分网站推广和网络营销手法都要老。 邮件营销- 特点 1.成本低廉 电子邮件营销只要有邮件服务器，联系10个用户与联系上万个用户的成本几乎没什么区别。但然如果要发上百封以上的邮件情况就不同了，因为这需要专用的服务器及非常大的宽带。特别是一些免费注册的邮箱每天发送邮件也是有限制的。另外发送一封邮件的成本几乎等于零，因为只要找到了用户邮箱，便可花很少的时间将邮件发送出去。 2.快速 相比其他网络营销手法，电子邮件营销也十分快速。搜索引擎优化需要几个月，甚至几年的努力才能充分发挥效果。博客营销更是需要时间，以及大量的文章。SNS营销需要花时间参与社区活动，建立广泛的关系网。而电子邮件营销只要有邮件数据在手，发送邮件几小时后就可看到效果。 3.精准 邮件只要能投放精准，客户转化率还是非常高的，远比其他营销方式高。随着RSS营销、博客营销等发展。及时向户传达信息已经不限于使用电子邮件了。 我们的- 优势 确定邮件营销的任务和目的 确定邮件发送主题进行有目的性的针对准客户进行发送邮件 广告巧妙构思 巧妙的内容构思策划进行邮件信息的发送。例如，一个销售饰品的，可以提供一个季节性的潮流风格指南;一家企业软件公司可以发送一些提高办公效率的技巧。 适当频率的进行发送 送频率过于频繁研究表明，同样内容的邮件，每个月发送2 - 3 次为宜。不要错误地认为，发送频率越高，收件人的印象就越深。过于频繁的邮件"轰炸"，只会让人厌烦，如果一周重
email营销方案怎么写 随着多媒体技术的发展,多媒体技术已经在很多领域被广泛地使用。多媒体技术在Email营销中的使用,即多媒体营销信息,将逐渐成为Email营销的新趋势。那么接下来跟读者一起来了解一下email 营销方案吧。 email营销方案一 1. 在邮件标题及正文中都尽量少使用敏感的、典型垃圾邮件常使用的词汇，如英文的伟哥，贷款，色情图片，获奖，赢取，以及中文的免费，促销，发票，礼物，避税，研修班，折扣，财务等。不是说这些词本身有什么问题，也不是完全不能用，而是尽量少用，以免触发垃圾过滤算法。少使用惊叹号，减少使用夸张的颜色，尤其是加粗的红色字体。这都是典型的垃圾邮件常用的吸引眼球的方法。如果是英文邮件，不要把很多词完全用大写。 2. 邮件内容、标题、发件人姓名都不要使用明显是虚构的字符串。比如有的垃圾邮件发送者当然不会告诉别人真名实姓，就在发信人名称中随便写上几个字母。这种莫名其妙的随机字符串通常都是欲盖弥彰的垃圾邮件特征。 3. HTML 邮件代码应该简洁，尽量少使用图片。虽然HTML 邮件允许使用图片美化邮件，但是图片与文字相比应该保持在最低比例。图片越多，被打的垃圾分数可能越高。
4. 在邮件中提醒用户把你的域名以及邮件地址加入到用户自己的白名单和通讯录中。邮件客户端软件通常都在垃圾过滤器设置中有白名单选项，绝大部分免费邮件提供商，如雅虎，hotmail，gmail 也都有相应的设置。把电子邮件地址存入到通讯录中也起到相同的效果。 5. 在电子邮件中含有"退订"的超级链接。订阅者可能会要求您将他的地址从您的数据库中删除。马上去做，并将处理的结果通知他。响应他们的要求有助于避免潜在的垃圾邮件抱怨，避免他们将您列入ISP的黑名单中。而且您的公司万一被垃圾邮件抱怨调查时，它可以证明您是一个许可营销邮件公司，而不是垃圾邮件的制造者。 6. 避免利用"抄送"和"暗送"将一封邮件发给大量的接收者。确信一封邮件每次仅发给了一个Email地址。 7. 选用优秀的邮件发送软件 现在的邮件发送、邮件群发软件越来越多了，其中软件质量效果参差不齐，为了您企业的形象建议选用优秀的软件进行营销，同时建议使用注册版的软件，否则不少试用版的都在邮件下端留有版权信息，收件人一看就知道是群发，企业形象大打折扣，事与愿违。 EDM(Email Direct Marketing)也即：Email营销、电子邮件营销，是指将制作好的广告档案通过电子邮件的形式群发至目标客户手中从而达到产品宣传、介绍及推广的目的，EDM是网络营销中的重
二、推广策划 1、网络策略 （1）网站建设 网站建设是网络营销策略的重要组成部分，网站内容是网络营销信息源的基础，网站的功能决定着哪些营销方法可以被采用而哪些不能被采用。那下面分几个方面说。 1）企业网站外观设计：也就是首页，一个网站的首页，是很重要的。首页制作得好，会吸引许多潜在的顾客。首页主题应该代表公司形象的图片，建议可以用厂房图片，而且制作成FLASH动画的效果会更好。最好再有一些优惠的消息、图片要丰富。设计要简洁，突出企业的专业化，也要能够体现出产品的特性。 2）公司相册：相册最好多分点类，不仅限于一、两个相册，这样显得过于单一。 3）发布图片：图片要生动，才有吸引力；图片要清晰，才能看清产品；可以采用新鲜柠檬及出产地作为图片背景。（如图2—1） 图2—1 4）发布供应信息：发布供应信息要在被允许的情况下，信息要尽量详细。最好发布一口价信息，这样自己的供应信息才会排位靠前，才会容易被人发现并点击。信息要不断更新，哪怕更新的只是一个小小的图片，及时发布新信息。安岳柠檬描述页面，要专业化。包含：商品规格，在阅读产品信息时应该注意买的欲望。这样就有可能留住那些“半确定”以及“不确定”的客户。
5）友情链接：友情链接是要注意时刻检查一些链接，杜绝死链接的发生。可别小看友情链接的作用，她会吸引一些网民来点击自己的网站，浏览量的上升，也会使一些潜在顾客多些。 6）联系方式：多几个在线联系方式，当然微店最好的联系方式就是通过微信还有手机通讯。其他方式还包括：阿里旺旺、QQ等，还要告知顾客自己具体 在线时间。最好能写上联系人的手机号码，以便顾客随时可以很容易的就找到你。 7）在线回复：这点现阶段做的很不好做，因为微店现阶段仅限于网页链接的评论回复、微信、手机短信等方式来接触顾客。在团队中管理中一方面要安排专业人员及时回复潜在顾客的问题，抓住商机；另一方面回复信息时注意用语专业，在售前，售中时，客服回复要热情，百问不厌。EMAIL回复，时间最好在一个工作日内。 8）设关键字：分析关键字、创建有用的关键字，可能会被搜到的关键词，例如：安岳特产、柠檬、柠檬价格、柠檬批发、柠檬片等等的相关内容；邮资说明，支付方式，退换货的说明，客服在线时的说明。抓住产品特点及卖点，为商品的推广做好准备。 9）精品推荐：推荐的应该是店铺最畅销、在活动中的产品，而且推荐也要多点，并配上专业精美的语句。相信这样就有可能留住那些“半确定”以及“不确定”的客户，并有可能产生购。 10）社区：在这里有很好的商机，在这里可以把自己的产品推销出去，所以要好好利用微信这个平台。通过添加附近的人、小区附近粘贴广告（设立微信二维码扫描）增加新的客户群体。所以应该根据产品的特性，在社区里通过发贴，进行适当推广。 14）支付软件：使用安全、方便快捷的支付方式。例如：支付宝、微信支付、QQ支付、转账等支付软件，提高消费者在感知网上交易的安全性。 （2）网站推广 网站推广可以扩大知名度、吸引新用户、增加保持力和忠诚度。那下面有7种方法： 1）搜索引擎推广方法