1

Topic: Simple recommendatory system

I welcome. At once I speak - a subject I understand superficially. I.e. a principle of operation and any titles of algorithms I know, but is deeper, in mathematics, did not plunge. Therefore I ask to answer too, "on fingers". In general the idea of implementation of simple recommendatory system came to me for a long time. 1. The user receives the message. The message consists of words. All words are brought in basis if they in it missed. Except for the unions, pretexts, digits, etc. 2. The user estimates the message - "plus" or "minus": 2.1. If "plus" to each word of the message in basis the rating raises. 2.2. If "minus", on the contrary: to each word of the message in basis the rating is reduced. 3. The system receives the following message and on the basis of words of this message and a rating of these words from basis, forms a message rating. 3.1. We admit, the message rating is formed by stupid addition. Then, if the rating is more than zero the message is shown to the user if it is less is not present. 3.2. Let's admit, the message rating too is formed by stupid addition. All unread messages are built linearly depending on a rating: from high to the low. In this case the user can adjust "" messages. Is better should work on news sites. For the sake of experiment it would be possible and to implement, but laziness. How think, it will work?

2

Re: Simple recommendatory system

Hello, Real 3L0, you wrote: R3> 1. The user receives the message. The message consists of words. All words are brought in basis if they in it missed. Except for the unions, pretexts, digits, etc. R3> 2. The user estimates the message - "plus" or "minus": R3> 2.1. If "plus" to each word of the message in basis the rating raises. R3> 2.2. If "minus", on the contrary: to each word of the message in basis the rating is reduced. Nich-that I do not understand! () In sense, than is longer the message (more precisely, the in it more than significant words), the above or more low there can be a rating of all message? That is it is supposed to calculate or write off a rating for each word? If yes, what for so to do? After all there are cases when the short message is more useful and more readable than the long. And here suddenly it turns out that the rating for the short message certainly cannot be above, than for long...

3

Re: Simple recommendatory system

Hello, Real 3L0, you wrote: I Am afraid such simple algorithm will not guess though somehow well. Words can be same, but, as they say, with nuance... ps: here that you want to implement actually - https://en.wikipedia.org/wiki/Bag-of-words_model

4

Re: Simple recommendatory system

Hello, ZevS, you wrote: ZS> I Am afraid such simple algorithm will not guess though somehow well. Words can be same, but, as they say, with nuance... With the same success it is possible to compare human transfer to the machine. I think, coincidence of 90 % of significant words generally is possible.

5

Re: Simple recommendatory system

Hello, Lazytech, you wrote: L> In sense, than is longer the message (more precisely, the in it more than significant words), the above or more low there can be a rating of all message? Yes. But there are two variants. 1. Quite probably that if this long message the keyword will meet more often, therefore as a whole it does not affect a rating of the message. 2. It is possible to add any normalizing coefficient in algorithm of a rating. L> that is it is supposed to calculate or write off a rating for each word? Yes. L> If yes, what for so to do? After all there are cases when the short message is more useful and more readable than the long. And here suddenly it turns out that the rating for the short message certainly cannot be above, than for long... As I wrote, a basis are news sites. On them such normally does not happen. And the most important thing: it not a usefulness estimation. It is an estimation .

6

Re: Simple recommendatory system

R3> As think, it will work? I think that for such purpose it will be better to work the simplified variant: 1. The pool of the read messages will be clustered under the text (with selection of roots of words and the registration of word combinations), k-means or hierarchically 2. Depending on estimations each cluster is estimated on . 3. For each new message its cluster and  is defined. 4. After an estimation of the new message passage in point 1.

7

Re: Simple recommendatory system

Hello, Real 3L0, you wrote: R3> As think, it will work? As a whole, you described a principle of operation of that is called Content Based Filtering. But described superficially. Therefore to work as a whole will be, but it is possible is much better. It will be better to work, if words in the message are not stupid for pushing in basis, and to "weigh", for example, by means of TF-IDF metrics. A word, with a bicycle, you, the barin!

8

Re: Simple recommendatory system

Hello, De-Bill, you wrote: DB> 1. The pool of the read messages will be clustered under the text (with selection of roots of words and the registration of word combinations), k-means or hierarchically Clustering implies small division, for example, on "Windows setting" and "Windows reinstallation"? One can be interesting, another - is not present.

9

Re: Simple recommendatory system

R3> For the sake of experiment it would be possible and to implement, but laziness. If, how it was clarified, the algorithm is known, can eat implementation examples? And examples of the given algorithm, without the registration of estimations of other users.

10

Re: Simple recommendatory system

R3> Clustering implies small division, for example, on "Windows setting" and "Windows reinstallation"? One can be interesting, another - is not present. Depends how you define roots. If to cut only the terminations setting and reinstallation will be different words.

11

Re: Simple recommendatory system

R3> If how it was clarified, the algorithm is known, can eat implementation examples? And examples of the given algorithm, without the registration of estimations of other users. TfidfVectorizer in sklearn.feature_extraction.text

12

Re: Simple recommendatory system

Hello, De-Bill, you wrote: R3>> If how it was clarified, the algorithm is known, can eat implementation examples? And examples of the given algorithm, without the registration of estimations of other users. DB> TfidfVectorizer in sklearn.feature_extraction.text In sense, for the user, instead of the programmer?

13

Re: Simple recommendatory system

Hello, Real 3L0, you wrote:  forgot to fasten! Without  now does not fly up!... <<RSDN@Home 1.0.0 alpha 5 rev. 0>>

14

Re: Simple recommendatory system

Hello, Real 3L0, you wrote: R3> 3.1. We admit, the message rating is formed by stupid addition. Then, if the rating is more than zero the message is shown to the user if it is less is not present. No, no, no, no, no! At first the system of estimations is set. For example, badly, neutrally, well. Or five-ball , where neutrally (well) as a matter of fact triple, and four and five, well and perfectly,  and two, disgustingly and badly. Or ten-ball on any basis and so on. Further each person estimates messages. On the basis of these estimations of messages makes the list of adherents, that is people which as estimated messages and at which the difference in estimations is minimum. On the basis of the general difference in estimations ordering of adherents is formed. And already  the individual list of recommendations. This standard decision, itself such never did, but used. The majority recommendation otherwise turns out. But as the similar system will produce everyone I get on, many do not begin to estimate simply. As a result the majority making recommendations will be actually minority from total of visitors. That that at you is described it  when its creator  user group it  and made to itself(himself) the infinite karma. But it only meant that the similar system of estimations failed. The minority can squeeze out the majority, and it already estimates nothing this majority. By the way, to tell, the system of estimations RSDN in the technical plan too became outdated. She does not help to search for adherents or the useful comments. And still as a whole about estimations, personally it is not pleasant to me, when from me hide messages. It is necessary to tear them, sometimes some times because some administrators guessed to displace an amount of times depending on an amount of the bad estimations. It looks extremely stupidly. And on other resources it is necessary to tear the message and there the text changes color, and than worse an estimation, especially it merges with a background. It is necessary to strain eyes that it to read, or to select, to read on contrast of selection. As administrators guessed to such  I do not know, or  the message, or do not hide, what for to spoil to visitors of an eye. Thus the democratic system of estimations does not carry generally any favor.

15

Re: Simple recommendatory system

Hello, velkin, you wrote: V> And already  the individual list of recommendations. This standard decision, itself such never did, but used. I used many recommendatory systems. V> the majority recommendation otherwise turns out. But as the similar system will produce everyone I get on, many do not begin to estimate simply. And whence in opposite/my case there is a majority? After all all estimations is only my estimations. When I thought of implementation if I will begin this experiment I will program this all without server usage. V> by the way to tell, the system of estimations RSDN in the technical plan too became outdated. She does not help to search for adherents or the useful comments. She never also did not allow. If only most not to remember those who "plus" delivered you. By the way, as a variant, on the basis of RSDN it is possible to test not bad rating . V> And still as a whole about estimations, personally it is not pleasant to me, when from me hide messages.... V> and on other resources it is necessary to tear the message and there the text changes color, and than worse an estimation, especially it merges with a background.... +1 if it is implemented in comments. If it is implemented for the root message any more all so is bad.

16

Re: Simple recommendatory system

Hello, velkin, you wrote: V> Hello, Real 3L0, you wrote: R3>> 3.1. We admit, the message rating is formed by stupid addition. Then, if the rating is more than zero the message is shown to the user if it is less is not present. V> Is not present, no, no, no, no! V> the system of estimations At first is set. For example, badly, neutrally, well. Or five-ball  where it is neutral (well) as a matter of fact triple, and four and five, well and perfectly,  and two, disgustingly and badly. Or ten-ball on any basis and so on. V> Further each person estimates messages. On the basis of these estimations of messages makes the list of adherents, that is people which as estimated messages and at which the difference in estimations is minimum. On the basis of the general difference in estimations ordering of adherents is formed. [skipped] here it was the surface description so-called Collaborative Filtering. It is other approach to recommendations. But too legitimate and quite to itself spread. Especially in cases, when the document (for example, the film) is impossible (or it are very inconvenient) for breaking into making sense components (for example, words). Here and so "Is not present, no, no, no, no!" To shout it the rights does not give.  it is possible to recommend differently. Methods . : I  here to steam of heads from the  on recommendatory systems for introduction (especially .2). Only here the publisher, a reptile, does not resolve. Here if was  any input... It would be possible to organize. The profit would be all.