【摘要】：現(xiàn)在我們處在信息急速爆炸的時(shí)代,這時(shí)候很難做到為用戶提供符合心意的有用信息。因?yàn)樗阉饕娴某霈F(xiàn),用戶減少了部分信息過(guò)載壓力,但存在結(jié)果單一性問(wèn)題,無(wú)法提供差異性的可以滿足用戶偏好的服務(wù)。具體的,推薦系統(tǒng)通過(guò)探究其核心的相關(guān)信息,即用戶的行為、偏好和環(huán)境上下文等因素,篩選掉與用戶喜好無(wú)關(guān)的信息,從而為用戶推薦滿足個(gè)性化需求的服務(wù)。協(xié)同過(guò)濾是目前眾多推薦方法中應(yīng)用范圍最廣的。它的基本思想是挖掘用戶行為背后信息,篩選到相似用戶,依據(jù)相似用戶對(duì)某一具體資源的偏好來(lái)推斷目標(biāo)用戶對(duì)具體資源的喜好程度,依照其值順序推薦。實(shí)踐證明,此算法可提高電子商務(wù)領(lǐng)域中用戶由網(wǎng)頁(yè)瀏覽者到物品選購(gòu)者的轉(zhuǎn)化率。盡管協(xié)同過(guò)濾算法取得了不錯(cuò)的成績(jī),但傳統(tǒng)協(xié)同過(guò)濾算法僅通過(guò)單一評(píng)分來(lái)挖掘相似用戶(物品),推薦的效果并不占優(yōu)勢(shì)。不少學(xué)者將時(shí)間、地點(diǎn)、標(biāo)簽等上下文信息融合到協(xié)同過(guò)濾推薦算法中,以期提高個(gè)性化推薦的質(zhì)量。通過(guò)大量的與協(xié)同過(guò)濾算法相關(guān)的文獻(xiàn)閱讀、資料總結(jié)與內(nèi)容討論,本文在經(jīng)典的算法基礎(chǔ)上進(jìn)行了創(chuàng)新和改進(jìn),大量的模擬實(shí)驗(yàn)結(jié)果證明了新算法的可行性與優(yōu)越性。具體的工作總結(jié)為以下幾部分:(1)將時(shí)間上下文信息加入到協(xié)同過(guò)濾推薦算法中。利用用戶先后購(gòu)買同一物品的時(shí)間關(guān)系來(lái)衡量用戶間相似度,得到用戶特征向量;利用物品先后被同一用戶購(gòu)買的時(shí)間關(guān)系來(lái)衡量物品間相似度,可以計(jì)算得到該物品的特征向量;最后,將前面得到的特征向量融合到概率矩陣分解模型中并不斷的對(duì)其進(jìn)行優(yōu)化來(lái)降低誤差。(2)將標(biāo)簽上下文信息加入到協(xié)同過(guò)濾推薦算法中。利用標(biāo)簽信息來(lái)豐富用戶(物品)信息,提出了一種基于用戶(物品)標(biāo)簽特征向量的建模方法。通過(guò)用戶-標(biāo)簽、物品-標(biāo)簽二部圖求出用戶間相似度和物品間的相似度。將用戶評(píng)分的時(shí)間上下文因素考慮進(jìn)來(lái),對(duì)最近鄰模型進(jìn)行優(yōu)化,動(dòng)態(tài)發(fā)現(xiàn)對(duì)當(dāng)前用戶(物品)影響最大的鄰居集合。(3)提出一種融合時(shí)間上下文和標(biāo)簽上下文的協(xié)同過(guò)濾推薦算法。通過(guò)時(shí)間上下文來(lái)計(jì)算用戶相似度,通過(guò)標(biāo)簽上下文來(lái)計(jì)算物品相似度,最后融合到矩陣分解模型中。(4)提出融合多種上下文的推薦系統(tǒng)框架,并給出上下文數(shù)據(jù)采集、用戶興趣偏好提取,上下文感知推薦生成的具體方法。
[Abstract]:Now we are in the era of information explosion, it is difficult to provide users with the right useful information. Because of the emergence of search engines, users reduce the pressure of partial information overload, but there is a problem of single results, which can not provide different services that can meet users' preferences. Specifically, the recommendation system through exploring its core information, that is, user behavior, preferences and environmental context and other factors, screening out information independent of user preferences, so as to recommend users to meet the needs of personalized services. Collaborative filtering is the most widely used recommendation method. Its basic idea is to mine the information behind the user's behavior, filter out the similar user, infer the target user's preference for a specific resource according to the preference of the similar user to a specific resource, and recommend it according to its value order. It is proved that this algorithm can improve the conversion rate from page viewer to item buyer in the field of electronic commerce. Although the collaborative filtering algorithm has achieved good results, the traditional collaborative filtering algorithm only uses a single score to mine similar users (items), and the recommended results are not dominant. In order to improve the quality of personalized recommendation, many scholars fuse the contextual information such as time, place and label into collaborative filtering recommendation algorithm. By reading a lot of literatures, summarizing the data and discussing the contents of the collaborative filtering algorithm, this paper innovates and improves on the basis of the classical algorithm, and a large number of simulation results prove the feasibility and superiority of the new algorithm. The specific work is summarized as follows: (1) time context information is added to collaborative filtering recommendation algorithm. The similarity between users is measured by the time relationship between the same items, and the user feature vector is obtained, and the similarity between the items is measured by the time relationship between the items purchased by the same user. The eigenvector of the item can be calculated. Finally, the former eigenvector is fused into the probabilistic matrix decomposition model and optimized continuously to reduce the error. (2) the label context information is added to the collaborative filtering recommendation algorithm. This paper presents a modeling method based on user (item) label feature vector to enrich user (object) information by label information. The similarity between user and item is obtained by user-label, item-label two-part graph. Taking into account the time context of the user rating, the nearest neighbor model is optimized. (3) A collaborative filtering recommendation algorithm combining time context and label context is proposed. The similarity of users is calculated by time context, and the similarity of items is calculated by label context. Finally, it is fused into matrix decomposition model. (4) the framework of recommendation system is proposed, and the context data collection is given. User interest preference extraction, context-aware recommendation generation of specific methods.
【學(xué)位授予單位】：山東師范大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2016
【分類號(hào)】：TP391.3
，

本文編號(hào)：2259565