Adaptive recommendation explanations

Abstract

A method and system for generating adaptive explanations for associated recommendations is disclosed. The adaptive explanations comprise a syntactical structure and associated phrases that are selected in accordance with usage behaviors and/or inferences associated with usage behaviors. The phrases included in an adaptive explanation may be selected through application of a non-deterministic process. The adaptive explanations may be beneficially applied to recommendations that are associated with content, products, and people, including recommendations that comprise advertisements.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS

This formula emphasizes per object accesses, but tempers this with a square root factor associated with the absolute level of accesses by the member. The result is a table, Table 2A, of the form:

TABLE 2A
Member 1
Behaviors	Topic 1	Topic 2	Topic 3	Topic 4	. . .	Topic N
Subscriptions	1	1	0	0		1
Topic Accesses	14	3	57	0		8
Weighted Accesses	9.1	12	3.2	0.6		2.3
Weighted Saves	0.9	1.3	1.1	0	. . .	0.03

The next step is to transform Table 2A into a MTAV. In some embodiments, indexing factors, such as the following may be applied:

Topic Affinity Indexing Factors Weight

Subscribe Indexing Factor       10
Topic Indexing Factor           20
Accesses Indexing Factor        30
Save Indexing Factor            40

These factors have the effect of ensuring normalized MTAV values ranges (e.g. 0-1 or 0-100) and they enable more emphasis on behaviors that are likely to provide relatively better information on member interests. In some embodiments, the calculations for each vector of Table 1A are transformed into corresponding Table 2 vectors as follows:

• • 1. Table 3 Indexed Subscriptions for a topic by Member 1=Table 2A Subscriptions for a topic * Subscribe Indexing Factor
  • 2. Table 3 Indexed Direct Topic Accesses by Member 1=Table 2A Topic Accesses* Topic Indexing Factor
  • 3. Table 3 Indexed Accesses for a topic by Member 1=((Table 2A Weighted Accesses for a topic by Member 1)/(Max(Weighted Accesses of all Topics by Member 1)))*Accesses Indexing Factor
  • 4. Table 3 Indexed Saves for a topic by Member 1=((Table 2A Weighted Saves for a topic by Member 1)/(Max (Weighted Saves of all Topics by Member 1)))* Saves Indexing Factor
    The sum of these Table 3 vectors results in the MTAV for the associated member 200 as shown in the hypothetical example of Table 3 below:

TABLE 3
Member 1
Indexed
Behaviors	Topic 1	Topic 2	Topic 3	Topic 4	. . .	Topic N
Subscriptions	0	10	10	10		10
Topic Accesses	5	1	20	0		8
Weighted Accesses	11	1	30	12		6
Weighted Saves	0	10	40	1		2
Member 1 MTAV	16	22	100	23	. . .	26

Member-to-member affinities can be derived by comparing the MTAV's of a first member 200 and a second member 200. Statistical operators such as correlation coefficients may be applied to derive a sense of the distance between members in n-dimensional topic affinity space, where there N topics. Since different users may have access to different topics, the statistical correlation for a pair of members must be applied against MTAV subsets that contain only the topics that both members have access to. In this way, a member-to-member affinity vector (MMAV) can be generated for each member or user 200, and the most similar members, the least similar members, etc., can be identified for each member 200.

With the MTAV's and MMAV's, and Most Similar Member information, a set of candidate objects to be recommended can be generated. These candidate recommendations will, in a later processing step, be ranked, and the highest ranked to candidate recommendations will be delivered to the recommendation recipient. Recall that recommendations 250 may be in-context of navigating the system 925 or out-of-context of navigating the system 925. Following are more details on an exemplary set of steps related to generating out-of-context recommendations. At each of step the candidate objects may be assessed against rejection criteria (for example, the recommendation recipient has already recently received the candidate object may be a cause for immediate rejection) and against a maximum number of candidate objects to be considered.

• • 1. Determine if there are objects that have been related to objects at a sufficient level of relatedness that have been “saved” by the recommendation recipient since the time the recommendation recipient saved the object. This may be a good choice for a first selection step because it would be expected that such objects would be highly relevant to the recommendation recipient.
  • 2. Determine if the Most Similar Members to the recommendation recipient have “saved” (related) objects in the last 12 months to the recommendation recipient's highest affinity topics.
  • 3. Determine if the Most Similar Members to the recommendation recipient have rated objects at a level greater than some threshold over some time period, that are related to the recommendation recipient's highest affinity topics.
  • 4. Determine the most frequently accessed objects by the Most Similar Members over some period of time that are related to recommendation recipient's highest affinity topics.
  • 5. Determine the highest influence objects that are related to recommendation recipient's highest affinity topics.

A variation of the out-of-context recommendation process may be applied for in-context recommendations, where the process places more emphasis of the closeness of the objects to the object being viewed in generating candidate recommendation objects.

For both out-of-context and in-context recommendations, a ranking process may be applied to the set of candidate objects, according to some embodiments. The following is an exemplary set of input information that may be used to calculate rankings.

• • 1. Editor Rating: If there is no editor rating for the object, this value is set to a default
  • 2. Community Rating (If there is no community rating for the object, this value can be set to a default)
  • 3. Popularity: Indexed popularity (e.g., number of views) of the object.
  • 4. Change in Popularity: Difference in indexed popularity between current popularity of the object and the object's popularity some time ago
  • 5. Influence: Indexed influence of the object, where the influence of an object is calculated recursively based on the influence of other objects related to the said object, weighted by the degree of relationship to the said object, and where the initial setting of influence of an object is defined as its popularity.
  • 6. Author's Influence: Indexed influence of the highest influence author (based on the sum of the influences of the author's content) of the content referenced by the object
  • 7. Publish Date: Date of publication of the object
  • 8. Selection Sequence Type: An indicator the sequence step in which the candidate object was selected
  • 9. Object Affinity to MTAV: The indexed vector product of the Object-Topic Affinity Vector (OTAV) and the MTAV. The values of the OTAV are just the relevancies between the object and each topic. Here is an example of the OTAV-MTAV vector product.

A ranking is then developed based on applying a mathematical function to some or all or input items listed directly above, and/or other inputs not listed above. In some embodiments, user or administrator-adjustable weighting factors may be applied to the raw input values to tune the object ranking appropriately. These recommendation preference settings may be established directly by the user, and remain persistent across sessions until updated by the user, in some embodiments.

Some example weighting factors that can be applied dynamically by a user or administrator are as follows:

1. Change in Popularity (What's Hot” factor)

2. Recency Factor

3. Object Affinity to MTAV

These weighting factors could take any value (but might be typically in the 0-5 range) and could be applied to associated ranking categories to give the category disproportionate weightings versus other categories. They can provide control over how important change in popularity, freshness of content, and an object's affinity with the member's MTAV are in ranking the candidate objects.

The values of the weighting factors are combined with the raw input information associated with an object to generate a rating score for each candidate object. The objects can then be ranked by their scores, and the highest scoring set of X objects, where X is a defined maximum number of recommended objects, can be selected for deliver to a recommendation recipient 200. In some embodiments, scoring thresholds may be set and used in addition to just relative ranking of the candidate objects. The scores of the one or more recommended objects may also be used by the computer-based system 925 to provide to the recommendation recipient a sense of confidence in the recommendation. Higher scores would warrant more confidence in the recommendation of an object than would lower scores.

Recommendation Explanation Generation

In addition to delivering a recommendation 250 for an object, the computer-based application 925 may deliver a corresponding explanation 250c of why the object was recommended. This can be very valuable to the recommendation recipient 200 because it may give the recipient a better sense of whether to bother to read or listen to the recommended content, without committing significant amount of time. For recommendations that comprise advertising content, the explanation may enhance the persuasiveness of the ad.

In some embodiments, variations of the ranking factors may be applied in triggering explanatory phrases. For example, the following table illustrates how the ranking information can be applied to determine both positive and negative factors that can be incorporated within the recommendation explanations. Note that the Ranking Value Range is the indexed attribute values before multiplying by special scaling factors Ranking Category Weighting Factors such as the “What's Hot” factor, etc.

TABLE 2E
	2
	Ranking		4	5
1	Value	3	1st	2nd	6
Ranking	Range	Transformed	Positive	Positive	Negative
Category	(RVR)	Range	Threshold	Threshold	Threshold
Editor Rating	0-100	RVR	60	80	20
Community Rating*	0-100	RVR	70	80	20
Popularity	0-100	RVR	70	80	10
Change in	−100-100	RVR	30	50	−30
Popularity
Object Influence	0-100	RVR	50	70	5
Author's Influence	0-100	RVR	70	80	.01
Publish Date	−Infinity-0	100-RVR	80	90	35
Object Affinity to	0-100	RVR	50	70	20
MTAV

An exemplary process that can be applied to generate explanations based on positive and negative thresholds listed in 2E is as follows:

Step 1: First Positive Ranking Category—subtract the 1^st Positive Threshold column from the Transformed Range column and find the maximum number of the resulting vector (may be negative). The associated Ranking Category will be highlighted in the recommendation explanation.

Step 2: Second Positive Ranking Category—subtract the 2^nd Positive Threshold column from the Transformed Range column and find the maximum number of the resulting vector. If the maximum number is non-negative, and it is not the ranking category we already selected, then include this second ranking category in the recommendation explanation.

Step 3: First Negative Ranking Category—subtract the Negative Threshold column from the Transformed Range column and find the minimum number of the resulting vector. If the minimum number is non-positive this ranking category will be included in the recommendation explanation as a caveat, otherwise there will be no caveats.

Although two positive and one negative thresholds are illustrated in this example, and unlimited number of positive and negative thresholds may be applied as required for best results.

In some embodiments explanations are assembled from component phrases and delivered based on a syntax template or function. Following is an example syntax that guides the assembly of an in-context recommendation explanation. In the syntactical structure below phrases within { } are optional depending on the associated logic and calculations, and “+” means concatenating the text strings. Other detailed syntactical logic such as handling capitalization is not shown in this simple illustrative example.
{[Awareness Phrase (if any)]}+{[Sequence Number Phrase (if any)]+[Positive Conjunction]}+[1^st Positive Ranking Category Phrase]+{[Positive Conjunction]+[2^nd Positive Ranking Category Phrase (if any)]}+{[Negative Conjunction]+[Negative Ranking Category Phrase (if any)]}+{[Suggestion Phrase (if any)]}

The following section provides some examples of phrase tables or arrays that may be used as a basis for selecting appropriate phrases for a recommendation explanation syntax. Note that in the following tables, when there are multiple phrase choices, they are selected probabilistically. “NULL” means that a blank phrase will be applied. [ ] indicates that this text string is a variable that can take different values.

System Awareness Phrases

Trigger Condition      Phrase

Apply these phrase     1) I noticed that
alternatives if any of 2) I am aware that
the 4 Sequence         3) I realized that
Numbers was triggered  4) NULL

Out-of-Context Sequence Number Phrases

Trigger Condition Phrase

Sequence 1        1) other members have related [this object] to
                  [saved object name], which you have saved,
Sequence 2        1) members with similar interests to you have
                  saved [this object]
Sequence 3        1) members with similar interests as you have
                  rated [this object]highly
                  2) Members that have similarities with you
                  have found [this object] very useful
Sequence 4        1) [this object] is popular with members that
                  have similar interests to yours
                  2) Members that are similar to you have often
                  accessed [this object]

Positive Ranking Category Phrases

Trigger Category     Phrase

Editor Rating        1) [it] is rated highly by the editor
Community Rating*    1) [it] is rated highly by other members
Popularity**         1) [it] is very popular
Change in Popularity 1) [it] has been rapidly increasing in popularity
Object Influence     1) [it] is [quite] influential
Author's Influence   1) the author is [quite] influential
                     2) [author name] is a very influential author
Publish Date         1) it is recently published
Object Affinity to   1) [it] is strongly aligned with your interests
MTAV (1)             2) [it] is related to topics such as [topic name]
                     that you find interesting
                     3) [it] is related to topics in which you have an
                     interest
Object Affinity to   4) I know you have an interest in [topic name]
MTAV (2)             5) I am aware you have an interest in [topic
                     name]
                     6) I have seen that you are interested in [topic
                     name]
                     7) I have noticed that you have a good deal of
                     interest in [topic name]

Positive Conjunctions

Phrase

1) and

Negative Ranking Category Phrases

Trigger
Category        Phrase

Editor Rating   1) it is not highly rated by the editor
Community       1) it is not highly rated by other members
Rating
Popularity      1) it is not highly popular
Change in       1) it has been recently decreasing in popularity
Popularity
Object          1) it is not very influential
Influence
Author's        1) the author is not very influential
Influence       2) [author name] is not a very influential author
Publish Date    1) it was published some time ago
                2) it was published in [Publish Year]
Object Affinity 1) it may be outside your normal area of interest
to MTAV         2) I'm not sure it is aligned with your usual interest areas

Negative Conjunctions

Phrase

1) , although
2) , however
3) , but

Suggestion Phrases (Use Only if No Caveats in Explanation)

Phrase

1) , so I think you will find it relevant
2) , so I think you might find it interesting
3) , so you might want to take a look at it
4) , so it will probably be of interest to you
5) , so it occurred to me that you would find it of interest
6) , so I expect that you will find it thought provoking
7) NULL

The above phrase array examples are simplified examples to illustrate the approach. In practice, multiple syntax templates, accessing different phrase arrays, with each phrase array many different phrases and phrase variations are required to give the feel of human-like explanations.

As mentioned above, a sense of confidence of the recommendation to the recommendation recipient can also be communicated within the recommendation explanation. The score level may contribute to the confidence level, but some other general factors may be applied, including the amount of usage history available for the recommendation recipient on which to base preference inferences and/or the inferred similarity of the user with one or more other users for which there is a basis for more confident inferences of interests or preferences.

Recommendation explanations are one type of behavioral-based communications 250c that the one or more computer-based applications 925 may deliver to users 200. Other types of adaptive communications 250c may be delivered to a user 200 without necessarily being in conjunction with the recommendation of an object or item of content. For example, a general update of the activities of other users 200 and/or other trends or activities related to people or content may be communicated.

Adaptive communications 250c may also comprise one or more phrases that communicate an awareness of behavioral changes in the user 200 over time, and inferences thereof. These behavioral changes may be derived, at least in part, from an evaluation of changes in the user's MTAV affinity values over time. In some cases, these behavioral patterns may be quite subtle and may otherwise go unnoticed by the user 200 if not pointed out by the computer-based system 925. Furthermore, the one or more computer-based systems may infer changes in interests or preferences of the user 200 based on changes in the user's behaviors over time. The communications 250c of these inferences may therefore provide the user 200 with useful insights into changes in his interest, preferences, and tastes over time. This same approach can also be applied by the one or more computer-based systems to deliver insights into the changes in interests, preferences and tastes associated with any user 200 to another user 200. These insights, packaged in an engaging communications 250c, can simulate what is sometimes referred to as “a theory of mind” in psychology.

The adaptive communications 250c in general may apply a syntactical structure and associated probabilistic phrase arrays to generate the adaptive communications in a manner similar to the approach described above to generate explanations for recommendations. The phrase tendencies of the adaptive communications 250c over a number of generated communications can be said to constitute a “personality” associated with the one or more computer-based applications 925. The next section describes how in some embodiments of the present invention the personality can evolve and adapt over time, based at least in part, on the behaviors of the communication recipients 200.

Adaptive Personalities

FIG. 9 is a flow diagram of the computer-based adaptive personality process 1000 in accordance with some embodiments of the present invention. A user request for a communication 1010 initiates a function 1020 that determines the syntactical structure of the communication 250c to the user 200. The communication 250c to user 200 may be an adaptive recommendation 250, an explanation associated with a recommendation, or any other type of communication to the user. The communication 250c may be in written format, or may be an audio-based format.

In accordance with the syntactical structure that is determined 1020 for the communication, one or more phrases are probabilistically selected 1030 based on frequency distributions 3030 associated with an ensemble of phrases to generate 1040 a communication 930 to the user.

User behaviors 920, which may include those described by Table 1 herein, are then evaluated 1050 after receipt of the user communication. Based, at least in part, on these evaluations 1050, the frequency distributions 3030 of one or more phrases that may be selected 1030 for future user communications are then updated 1060. For example, if the user communication 250c is an the explanation associated with an adaptive recommendation 250, and it is determined that the recommendation recipient reads the corresponding recommended item of content, then the relative frequency of selection of the one or more phrases comprising the explanation of the adaptive recommendation 250 might be preferentially increased versus other phrases that we not included in the user communication. Alternatively, if the communication 250c elicited one or more behaviors 920 from the communication recipient 200 that were indicative of indifference or a less than positive reaction, then the relative frequency of selection of the one or more phrases comprising the communication might be preferentially decreased versus other phrases that we not included in the user communication.

In FIG. 11, an illustrative data structure 3000 supporting the adaptive personality process 1000 according to some embodiments is shown. The data structure may include a designator for a specific phrase array 3010. A phrase array may correspond to a specific unit of the syntax of an overall user communication. Each phrase array may contain one or more phrases 3040, indicated by a specific phrase ID 3020. Associated with each phrase 3040 is a selection frequency distribution indicator 3030. In the illustrative data structure 3000 this selection frequency distribution of phrases 3040 in a phrase array 3010 is based on the relative magnitude of the value of the frequency distribution indicator. In other embodiments, other ways to provide selection frequency distributions may be applied. For example, phrases 3040 may be selected per a uniform distribution across phrase instances in a phrase array 3010, and duplication of phrase instances may be used to as a means to adjust selection frequencies.

Communication of Self-Awareness

FIG. 10 is a flow diagram of the computer-based adaptive self-awareness communication process 2000 in accordance with some embodiments of the present invention. The process 2000 begins with an evaluation 2010 of phrase frequency distribution 3030 changes over time. Then the appropriate syntactical structure of the communication 250c of self-awareness is determined 2020. One or more phrases 3040 that embody a sense of self-awareness are then selected in accordance with the syntactical structure requirements and changes in phrase frequency distributions over time.

Returning to FIG. 11, in some embodiments, phrase attributes that are associated with specific phrases may be used as a basis for self-aware phrase selection. Two example phrase attributes 3050, 3060 whose values are associated with specific phrases 3040 are shown. An unlimited number of attributes could be used as to provide as nuanced a level of self-awareness as desired.

When changes in phrase frequency distributions 3030 are evaluated 2010, the corresponding attributes 3050, 3060 are also evaluated. These attributes map to attributes 4050, 4060 that are associated with self-aware phrases 4040 in self-aware phrase data structure 4000. For example, if phrases 3040 that have the attribute value “humorous” have been increasing in frequency, then self-aware phrases that reference “humorous” may be appropriate to include in generating 2040 a communication of self-awareness 250c to a user 200. As is the case of any other communication 250c, the behaviors 920 of the recipient 200 of the communication may be evaluated 2050, and the self-aware phrase frequency distributions 4030 of the self-aware phrases 4040 may be updated 2060 accordingly. This recursive evaluation and updating of phrase frequency distributions can be applied without limit.

FIG. 12 depicts the major functions associated with a computer based system 925 that exhibits an adaptive personality, and optionally, a self-aware personality. Recall that in some embodiments, the computer-based system 925 comprises an adaptive system 100.

A request 6000 for a communication to a user 200 is made. The request 6000 may be a direct request from a user 200, or the request may be made by another function of the computer-based system 925. In some embodiments the request 6000 for a communication to the user may be initiated by a function that generates 240 an adaptive recommendation. A communication to the user is then generated 7000. This generation is done by first determining the appropriate syntactical rules or structure 7500 for the communication. In some embodiments, the syntax rules 7500 are of an “If some condition, Then apply a specific phrase array 3010” structure. Once the appropriate syntax is established and associated phrase arrays 3010 are determined, specific phrases are probabilistically retrieved from the phrase array function 5000 based on selection frequency distributions associated with the corresponding phrase arrays. The communication 250c is then assembled and delivered to a user 200.

User behaviors 920 of the communication recipient 200 are then monitored 8000. Based on inferences from these behaviors 920, the phrase array frequency distributions of the phrase array function 5000 are updated 9000 appropriately.

Computing Infrastructure

FIG. 13 depicts various computer hardware and network topologies on which the one or more computer-based applications 925 may operate.

Servers 950, 952, and 954 are shown, perhaps residing at different physical locations, and potentially belonging to different organizations or individuals. A standard PC workstation 956 is connected to the server in a contemporary fashion, potentially through the Internet. It should be understood that the workstation 956 can represent any computer-based device, mobile or fixed, including a set-top box. In this instance, the one or more computer-based applications 925, in part or as a whole, may reside on the server 950, but may be accessed by the workstation 956. A terminal or display-only device 958 and a workstation setup 960 are also shown. The PC workstation 956 or servers 950 may be connected to a portable processing device (not shown), such as a mobile telephony device, which may be a mobile phone or a personal digital assistant (PDA). The mobile telephony device or PDA may, in turn, be connected to another wireless device such as a telephone or a GPS receiver.

FIG. 13 also features a network of wireless or other portable devices 962. The one or more computer-based applications 925 may reside, in part or as a whole, on all of the devices 962, periodically or continuously communicating with the central server 952, as required. A workstation 964 connected in a peer-to-peer fashion with a plurality of other computers is also shown. In this computing topology, the one or more computer-based applications 925, as a whole or in part, may reside on each of the peer computers 964.

Computing system 966 represents a PC or other computing system, which connects through a gateway or other host in order to access the server 952 on which the one or more computer-based applications 925, in part or as a whole, reside. An appliance 968, includes software “hardwired” into a physical device, or may utilize software running on another system that does not itself the one or more computer-based applications 925. The appliance 968 is able to access a computing system that hosts an instance of one of the relevant systems, such as the server 952, and is able to interact with the instance of the system.

While the present invention has been described with respect to a limited number of embodiments, those skilled in the art will appreciate numerous modifications and variations therefrom. It is intended that the appended claims cover all such modifications and variations as fall within the scope of this present invention.

Claims

1. A computer-based recommendation method comprising:

generating an affinity vector between a first user of a computer-based system and a plurality of computer-based objects based, at least in part, on the first user's behaviors;

generating a similarity metric between the first user and a second user of the computer-based system based, at least in part, on the affinity vector of the first user and an affinity vector of the second user;

generating a recommendation for delivery to the first user based, at least in part, on the affinity vector of the first user and the similarity metric; and

generating an explanation for the recommendation comprising one or more phrases, wherein the selection of the one or more phrases is based, at least in part, on a plurality of user behaviors and in accordance with a computer-implemented syntactical structure.

2. The method of claim 1 wherein generating an affinity vector between a first user of a computer-based system and a plurality of computer-based objects based, at least in part, on the first user's behaviors comprises:

generating affinities between the user and a plurality of topic objects.

3. The method of claim 1 wherein generating a recommendation for delivery to the first user based, at least in part, on the affinity vector of the first user and the similarity metric comprises:

recommending the second user to the first user.

4. The method of claim 1 wherein generating a recommendation for delivery to the first user based, at least in part, on the affinity vector of the first user and the similarity metric comprises:

generating a recommendation responsive to the a user's search request.

5. The method of claim 1 wherein generating a recommendation for delivery to the first user based, at least in part, on the affinity vector of the first user and the similarity metric comprises:

generating a recommendation based on a recommendation preference setting established by associated with the user.

6. The method of claim 1 wherein generating an explanation for the recommendation comprising one or more phrases, wherein the selection of the one or more phrases is based, at least in part, on a plurality of user behaviors and in accordance with a computer-implemented syntactical structure comprises:

selecting phrases for inclusion in the explanation based on a frequency distribution.

7. The method of claim 1 wherein generating an explanation for the recommendation comprising one or more phrases, wherein the selection of the one or more phrases is based, at least in part, on a plurality of user behaviors and in accordance with a computer-implemented syntactical structure comprises:

selecting phrases for inclusion in the explanation in accordance with the calculated confidence level of the recommendation.

8. A computer-based recommendation system comprising:

means to generate an affinity vector between a first user of a computer-based system and a plurality of computer-based objects based, at least in part, on the first user's behaviors;

means to generate a similarity metric between the first user and a second user of the computer-based system based, at least in part, on the affinity vector of the first user and an affinity vector of the second user;

means to generate a recommendation for delivery to the first user based, at least in part, on the affinity vector of the first user and the similarity metric; and

means to generate an explanation for the recommendation comprising one or more phrases, wherein the selection of the one or more phrases is based, at least in part, on a plurality of user behaviors and in accordance with a computer-implemented syntactical structure.

9. The system of claim 8 wherein means to generate an affinity vector between a first user of a computer-based system and a plurality of computer-based objects based, at least in part, on the first user's behaviors comprises:

means to generate affinities between the user and a plurality of topic objects.

10. The system of claim 8 wherein means to generate a recommendation for delivery to the first user based, at least in part, on the affinity vector of the first user and the similarity metric comprises:

means to recommend the second user to the first user.

11. The system of claim 8 wherein means to generate a recommendation for delivery to the first user based, at least in part, on the affinity vector of the first user and the similarity metric comprises:

means to generate a recommendation responsive to the a user's search request.

12. The system of claim 8 wherein means to generate a recommendation for delivery to the first user based, at least in part, on the affinity vector of the first user and the similarity metric comprises:

means for generating a recommendation based on a recommendation preference setting established by associated with the user.

13. The system of claim 8 wherein means to generate an explanation for the recommendation comprising one or more phrases, wherein the selection of the one or more phrases is based, at least in part, on a plurality of user behaviors and in accordance with a computer-implemented syntactical structure comprises:

means to select phrases for inclusion in the explanation based on a frequency distribution.

14. The system of claim 8 wherein means to generate an explanation for the recommendation comprising one or more phrases, wherein the selection of the one or more phrases is based, at least in part, on a plurality of user behaviors and in accordance with a computer-implemented syntactical structure comprises:

means to select phrases for inclusion in the explanation in accordance with the calculated confidence level of the recommendation.

15. A computer-based recommendation explanation system comprising:

means to generate a recommendation based, at least in part, on a plurality of usage behaviors;

a syntactical structure for an explanation of an associated for the recommendation;

a plurality of phrase arrays associated with the syntactical structure, wherein each phrase array comprises a plurality of phrases;

a mapping of usage behaviors and corresponding phrase arrays appropriate to apply; and

means to generate an the explanation for the recommendation, wherein the explanation is generated in accordance with the syntactical structure and associated phrase arrays, and the mapping of usage behaviors with the corresponding phrase arrays appropriate to apply.

16. The system of claim 15 wherein the means to generate an explanation for the recommendation, wherein the explanation is generated in accordance with the syntactical structure and associated phrase arrays, and the mapping of usage behaviors with the corresponding phrase arrays appropriate to apply comprises:

means to probabilistically select from alternative syntactical structures.

17. The system of claim 15 wherein the means to generate an explanation for the recommendation, wherein the explanation is generated in accordance with the syntactical structure and associated phrase arrays, and the mapping of usage behaviors with the corresponding phrase arrays appropriate to apply comprises:

means to select phrases from a frequency distribution for inclusion in the explanation.

18. The system of claim 15 wherein the means to generate an explanation for the recommendation, wherein the explanation is generated in accordance with the syntactical structure and associated phrase arrays, and the mapping of usage behaviors with the corresponding phrase arrays appropriate to apply comprises:

means to apply behavioral thresholds to trigger appropriate phrase arrays.

19. The system of claim 15 wherein the means to generate an explanation for the recommendation, wherein the explanation is generated in accordance with the syntactical structure and associated phrase arrays, and the mapping of usage behaviors with the corresponding phrase arrays appropriate to apply comprises:

means to adjust the selection of phrases for inclusion in the explanation based on the phrase composition of previously generated explanations.

20. The system of claim 15 wherein the means to generate an explanation for the recommendation, wherein the explanation is generated in accordance with the syntactical structure and associated phrase arrays, and the mapping of usage behaviors with the corresponding phrase arrays appropriate to apply comprises:

means to adjust the selection of phrases for inclusion in the explanation based on usage behaviors of recipients of previously generated explanations.

21. A method comprising:

generating, by a processing device, a recommendation of a computer-based object for delivery to a first user of a computer-based system based on first user behavior associated with a plurality of computer-based objects; and

generating, by the processing device, an explanation for the recommendation comprising one or more phrases based on a plurality of other user behaviors and in accordance with a computer-implemented syntactical structure.

22. The method of claim 21, further comprising,

generating, by the processing device, a first affinity vector associated with the first user based on the first user behavior associated with the plurality of computer-based objects;

generating, by the processing device, a similarity metric between the first user and a second user of the computer-based system based on the first affinity vector and a second affinity vector associated with the second user;

wherein, the generating the recommendation is based on the first affinity vector and the similarity metric.

23. The method of claim 22 wherein the second affinity vector is based on the second user behavior associated with the plurality of computer-based objects.

24. The method of claim 22 wherein the second affinity vector is based on the second user behavior associated with a different plurality of computer-based objects.

25. The method of claim 22 wherein the generating the affinity vector further comprises generating, by the processing device, affinities between the user and a plurality of topic objects.

26. The method of claim 22 wherein the recommendation further comprises recommending the second user to the first user.

27. The method of claim 21 wherein the recommendation is responsive to a search request associated with the first user.

28. The method of claim 21 wherein the recommendation is based on a recommendation preference setting associated with the first user.

29. The method of claim 21 wherein the generating the explanation further comprises selecting, by the processing device, phrases for inclusion in the explanation based on a frequency distribution.

30. The method of claim 21 the generating the explanation further comprises selecting, by the processing device, phrases for inclusion in the explanation in accordance with the calculated confidence level of the recommendation.

31. An apparatus comprising:

means for generating a recommendation of a computer-based object for delivery to a first user of a computer-based system based on first user behavior associated with a plurality of computer-based objects; and

means for generating an explanation for the recommendation comprising one or more phrases based on a plurality of other user behaviors and in accordance with a computer-implemented syntactical structure.

32. The apparatus of claim 31, further comprising,

means for generating a first affinity vector associated with the first user based on the first user behavior associated with the plurality of computer-based objects;

means for generating a similarity metric between the first user and a second user of the computer-based system based on the first affinity vector and a second affinity vector associated with the second user;

wherein, the generating the recommendation is based on the first affinity vector and the similarity metric.

33. The apparatus of claim 32 wherein the second affinity vector is based on the second user behavior associated with the plurality of computer-based objects.

34. The apparatus of claim 32 wherein the second affinity vector is based on the second user behavior associated with a different plurality of computer-based objects.

35. The apparatus of claim 32 wherein the means for generating the affinity vector further comprises means for generating affinities between the user and a plurality of topic objects.

36. The apparatus of claim 31 wherein the recommendation further comprises recommending the second user to the first user.

37. The apparatus of claim 31 wherein the recommendation is responsive to a search request associated with the first user.

38. The apparatus of claim 31 wherein the recommendation is based on a recommendation preference setting associated with the first user.

39. The apparatus of claim 31 wherein the means for generating the explanation further comprises means for selecting phrases for inclusion in the explanation based on a frequency distribution.

40. The apparatus of claim 31 wherein the means for generating the explanation further comprises means for selecting phrases for inclusion in the explanation in accordance with the calculated confidence level of the recommendation.

41. An apparatus, comprising:

computer hardware configured to:

store instructions associated with an application program; and

execute the stored instructions to:

generate a recommendation based, at least in part, on a plurality of usage behaviors;

associate a plurality of phrase arrays with a syntactical structure;

map the plurality of usage behaviors and corresponding phrase arrays selected from the plurality of phrase arrays; and

generate an explanation for the recommendation based on the map, the explanation comprising a phrase associated with one of the corresponding phase arrays in accordance with the syntactical structure.

42. The apparatus of claim 41, wherein the corresponding phrase array comprises a plurality of phrases.

43. The apparatus of claim 41 wherein the computer hardware is further configured to select alternative syntactical structures based on probabilistic determination to generate the explanation.

44. The apparatus of claim 41 wherein the computer hardware is further configured to select the phrase from a frequency distribution for inclusion to generate the explanation.

45. The apparatus of claim 41 wherein the computer hardware is further configured to apply behavioral thresholds to map the plurality of usage behaviors and corresponding phrase arrays selected from the plurality of phrase arrays.

46. The apparatus of claim 41 wherein the computer hardware is further configured to adjust the selection of the phrase for inclusion in the explanation based on the phrase composition of previously generated explanations.

47. The apparatus of claim 41 wherein the computer hardware is further configured to adjust the selection of the phrase for inclusion in the explanation based on usage behaviors of recipients of previously generated explanations.

48. computer hardware having instructions stored thereon that, in response to execution by a processing device, cause the processing device to perform operations comprising:

generating a recommendation based, at least in part, on a plurality of usage behaviors;

associating a plurality of phrase arrays with a syntactical structure;

mapping the plurality of usage behaviors and corresponding phrase arrays selected from the plurality of phrase arrays; and

generating an explanation for the recommendation based on the map, the explanation comprising a phrase associated with one of the corresponding phase arrays in accordance with the syntactical structure.

49. The computer-readable-memory device of claim 48, wherein the corresponding phrase array comprises a plurality of phrases.

50. The computer hardware of claim 48 wherein the operations further comprise selecting alternative syntactical structures based on probabilistic determination to generate the explanation.

51. The computer hardware of claim 48 wherein the operations further comprise selecting the phrase from a frequency distribution for inclusion to generate the explanation.

52. The computer hardware of claim 48 wherein the operations further comprise applying behavioral thresholds to map the plurality of usage behaviors and corresponding phrase arrays selected from the plurality of phrase arrays.

53. The computer hardware of claim 48 wherein the operations further comprise adjusting the selection of the phrase for inclusion in the explanation based on the phrase composition of previously generated explanations.

54. The computer hardware of claim 49 wherein the operations further comprise adjusting the selection of the phrase for inclusion in the explanation based on usage behaviors of recipients of previously generated explanations.