增量式个性化推荐系统研究

互联网个性化推荐系统(Internet personal recommender systems)是根据用户的兴趣推荐最相关的互联网信息给用户的系统。在网上信息过载矛盾越来越严重、用户信息检索的个性化需求日益增强的现状下，推荐系统已经在搜索引擎、电子商务、网上社区等互联网关键应用中起到了关键性的作用，并且越来越受到重视。 然而，在大型网站上部署一个成熟推荐系统的代价依然很大，需要大量的计算和存储资源，推荐的准确性也依然有很大提升空间和需求，这就为推荐系统的研究提供了很多挑战。在这些挑战中推荐算法的准确性和可扩展性一直是该领域最为关注的两个问题，所谓推荐的准确性是指推荐的信息中用户真正感兴趣的比例，而可扩展性指的是系统能否在可容忍的时间和空间复杂度内处理海量的数据。如何在提高算法推荐准确性的同时增强算法的可扩展性是推荐系统改进的主要研究目标。然而，目前学术界的研究更多侧重于提高推荐算法的准确性，而对于可扩展性，很多准确性很高的算法由于需要比较复杂的计算，处理大规模动态数据的能力往往比较有限，并且它们的评测实验中并没有将可扩展性纳入到评价范畴，导致这些算法目前还很难在工业界大规模应用。 本论文的研究试图解决这一问题。通过在推荐算法中借鉴增量学习(Incremental learning)的思想，即考虑最新的训练数据来更新原有的机器学习模型，不需要或仅需要参考部分旧的训练数据，相对于使用全部数据也即批量的处理方式，增量式改进可以大大降低模型更新的复杂度，从而可以大幅度提高推荐算法在遇到新的训练数据时推荐模型更新的效率，降低计算代价，使得推荐模型的更新可以更加及时，进而提高推荐结果的准确性。具体来说，我们在提出了两种新的增量式协同过滤算法的同时，采用增量式学习的方法对目前准确性最好的若干推荐算法进行加速，特别是提高这些算法面对新的训练数据的更新模型的速度和效率，从而为这些算法的大规模的应用提供了可能。另一方面，新的训练数据包含了最新的用户兴趣，因此相对于旧的训练数据，算法在做更新时应给予更高的权重，这样才能做到推荐的结果在考虑到用户长期兴趣的同时，特别考虑用户近期的兴趣，从而使得推荐结果更加准确。这两方面归纳起来，我们旨在通过增量式学习使得推荐算法在更新时更加高效和精确，真正适用于互联网上海量数据的推荐，同时对其他增量式推荐系统方面的研究也具有借鉴意义。我们的改进工作主要包括以下几个方面： 基于主题模型的增量式推荐算法。主题模型，特别是概率隐含主题模型(PLSA)是一种广泛应用于推荐系统的主流方法，在文本推荐、图像推荐以及协同过滤推荐领域都有着很好的推荐效果。目前制约PLSA算法取得更大成功的重要因素就是PLSA算法更新的复杂度过高，使得学习模型的更新只能做批量式处理，这样就导致推荐的时效性不高，也没有办法体现用户的最新的兴趣和整体的最新动态。我们提出了一种增量式学习方法，可以应用于文本分析领域和协同过滤领域，当有新的训练数据到来时，对于基于文本的推荐，增量式更新方法仅寻找最相关的用户和文本以及涉及到的单词进行主题分布的更新，并给予新的文本以更高权重；对于协同过滤，我们的方法仅对当前用户所评分过得物品以及当前物品所涉及的用户进行更新，大大降低了更新的运算复杂度，提高了新数据在推荐算法中所占的权重，使得推荐更加准确、及时。我们的算法在天涯问答文本数据集上和MovieLens电影推荐数据集、Last.FM歌曲推荐数据集、豆瓣图书推荐数据集等协同过滤数据集上取得了很好的效果。 基于蚁群算法(Ant colony algorithm)的协同过滤推荐方法。受到群体智能(Swarm intelligence)算法的启发，我们提出了一种类似于蚁群算法的协同过滤推荐方法——Ant Collaborative Filtering，初始化阶段该方法给予每个用户或一组用户以全局唯一的单位数量的信息素，当用户对物品评分或者用户表示对该物品感兴趣时，用户所携带的信息素相应的传播到该物品上，同时该物品上已有的信息素（初始化为0）也会相应的传播给该用户；此外，用户和物品所携带的信息素会随着时间的推移有一定速率的挥发，通过挥发机制，可以在推荐时更重视用户近期的兴趣；推荐阶段，按照用户和物品所携带的信息素的种类和数量，我们可以得到相应的相似度，进而通过经典的相似度比较的方法来进行推荐。基于蚁群的协同过滤方法的优势在于可以有效的降低训练数据中的稀疏性，并且推荐算法可以实时的进行更新和推荐，同时考虑了用户兴趣随着时间的变化。我们在MovieLens电影评分、豆瓣书籍推荐、Last.FM音乐推荐数据集上验证了我们的方法。最后，我们建立了一个互联网新闻推荐系统，该系统以Firefox插件形式实现，自动采集用户浏览兴趣和偏好，后端使用不同的推荐算法推荐用户感兴趣的新闻给用户。 基于联合聚类(Co-clustering)的两阶段协同过滤方法。聚类(Clustering)是一种缩小数据规模、降低数据稀疏性的有效方法。对于庞大而稀疏的协同过滤训练数据来说，聚类是一种很自然事实上也的确很有效的预处理方法。因此我们提出了一种两阶段协同过滤框架：首先通过我们提出的一种联合聚类的方法，将原始评分矩阵分解成很多维度很小的块，每一块里面包含相似的用户对相似的物品的评分，然后通过矩阵拟合的方法（我们使用了非负矩阵分解NMF和主题模型PLSA）来对这些小块中的未知评分进行预测。当用户新增了对于某物品的一条评分，我们仅需要更新该用户或该物品所处的数据块进行重新评分预估，大大加快了评分预估的速度。我们在MovieLens电影评分数据集上验证了该算法的效果。 本文的研究成果不仅可以直接应用于大型推荐系统中，而且对于增量式推荐系统的后续研究也具有一定的指导意义。首先基于PLSA的增量式推荐算法对于其他基于图模型的推荐系统具有借鉴价值，其次蚁群推荐算法为一类新的、基于群体智能(Swarm intellignece)的协同过滤算法做出了有价值的探索，最后我们提出的两阶段协同过滤框架对于提高推荐算法的可扩展性和更新效率提出了一个通用的有效解决方案。 推荐系统是一个无止尽的优化的过程，除了推荐精度的不断提高之外，推荐算法的性能随着互联网上数据量的增加也需要进一步提高，增量式学习无疑是提高推荐算法更新速度最重要的方法，本文的研究为这一方向提供了参考。

Anomalous Thermal Stability of Nd-Fe-Co-Al Bulk Metallic Glass

The thermal stability of Nd60Fe20Co10Al10 bulk metallic glass (BMG) has been studied by differential scanning calorimetry (DSC), dynamic mechanical thermal analysis (DMTA), isochronal dilatation and compression tests. The results show that the glass transition of the BMG takes place quite gradually between about 460 and 650 K at a heating rate of 0.17 K/s. Several transformation processes are observed during continuous heating with the first crystallization process beginning at about 460 K, while massive crystallization takes place near the solidus temperature of the alloy. The positive heat of mixing between the two major constituents, Nd and Fe, and, consequently, a highly inhomogeneous composition of the attained amorphous phase are responsible for the anomalous thermal stability in this system. (C) 2002 Acta Materialia Inc. Published by Elsevier Science Ltd. All rights reserved.

Interactions of CO, HCl, and Sox in Pulverised Coal Flames

The effect of HCl and SO_2 on CO oxidation in pulverised coal flames was investigated experimentally and kinetically in an entrained flow combustion reactor. Two bituminous coals (German 'Goettelborn' and a Polish coal) were used as fuels with a feeding rate of 1 or 1.5 kg/h. HCl or SO_2 is introduced into the reactor premixed with the primary air. Experimental results indicate that HCl addition may inhibit CO oxidation in coal flames and increases CO emission. Reducing temperature in the reactor will enhance the inhibitory effect of HCl on CO oxidation. The measured CO profiles along the reactor height clearly show that the addition of HCl may inhibit CO oxidation. In the experimental range of SO_2 addition. The inhibiting effect of SO_2 on CO oxidation is less significant than HCl. A detailed kinetic mechanism is used to model the reactions. And the controlling reactions are analysed.

流化床中RDF焚烧时CO、SO_2和HCl的生成

垃圾衍生燃料 (Refuse-Derived Fuels-RDF)具有热值高、易燃烧的优点。RDF的一个潜在应用是与煤进行混烧，替代一部分锅炉燃烧用煤。由于RDF挥发份相当高，因此燃烧时的污染物排放不易控制。本文在非均匀布风流化床中进行了RDF与煤的混烧试验，测量了H_2O、CO、CO_2、NO、N_2O、HCl、SO_2等污染物质的排放特性。结果表明与单纯燃烧RDF相比，混烧时的CO生成量大大下降；SO2生成浓度较低，而HCl的生成量比单纯烧煤时明显增加。

Practical Criterion for Universal Laws and Practical Methods for Construction of Co-variance Sets

给出相对论力学中普遍定律的实用判别法和协变集的实用构造法,还给出实现非普遍定律的“可导出性”的一种实用方法．

Quantum Monte Carlo study of the CO interaction with a dimer model surface for Cr(110)

The chemisorption of CO on a Cr( 110) surface is investigated using the quantum Monte Carlo method in the diffusion Monte Carlo (DMC) variant and a model Cr2CO cluster. The present results are consistent with the earlier ab initio HF study with this model that showed the tilted/ near-parallel orientation as energetically favoured over the perpendicular arrangement. The DMC energy difference between the two orientations is larger (1.9 eV) than that computed in the previous study. The distribution and reorganization of electrons during CO adsorption on the model surface are analysed using the topological electron localization function method that yields electron populations, charge transfer and clear insight on the chemical bonding that occurs with CO adsorption and dissociation on the model surface.

Studies on two-phase co-current air/non-Newtonian shear-thinning fluid flows in inclined smooth pipes

In this work. co-current flow characteristics of air/non-Newtonian liquid systems in inclined smooth pipes are studied experimentally and theoretically using transparent tubes of 20, 40 and 60 turn in diameter. Each tube includes two 10 m lone pipe branches connected by a U-bend that is capable of being inclined to any angle, from a completely horizontal to a fully vertical position. The flow rate of each phase is varied over a wide range. The studied flow phenomena are bubbly, plug flow, slug flow, churn flow and annular flow. These are observed and recorded by a high flow. stratified flow. -speed camera over a wide range of operating conditions. The effects of the liquid phase properties, the inclination angle and the pipe diameter on two-phase flow characteristics are systematically studied. The Heywood-Charles model for horizontal flow was modified to accommodate stratified flow in inclined pipes, taking into account the average void fraction and pressure drop of the mixture flow of a gas/non-Newtonian liquid. The pressure drop gradient model of Taitel and Barnea for a gas/Newtonian liquid slug flow was extended to include liquids possessing shear-thinning flow behaviour in inclined pipes. The comparison of the predicted values with the experimental data shows that the models presented here provide a reasonable estimate of the average void fraction and the corresponding pressure drop for the mixture flow of a gas/ non-Newtonian liquid. (C) 2007 Elsevier Ltd. All rights reserved.

黑体泵浦CO_2-CO,N_2O-CO,CO_2-H_2O和N_2O-H_2O激光的研究

本文对黑体辐射泵浦CO_2激光和N_2O激光由于CO和H_2O的加入所引起的激光功率变化进行了理论和实验研究.CO的加入可使的N_2O激光功率增加28%.

稳态CO激光的解析理论

为了分析CO激光的特性,通常需要对上百个方程进行数值计算,因而解析的分析就成为一个很值得探讨的方面。本文从分布函数方程和辐射强度谱分量传输方程出发,导出稳态CO激光在级联跃迁发射时的增益饱和特性,得到的强度谱分量与增益的关系式适用于均匀和非均匀加宽同时起作用的一般情况。文中所使用的振动分布函数由一个非线性积分方程求得。对两能级特例,本文结果与文献[9]一致。对均匀加宽极限的特例,本文给出的强度和强度谱分量的关系式与文献[4]的结果一致。

半焦在流化床燃烧中Cl元素影响CO氧化的试验

在非均匀布风的流化床燃烧系统中进行了半焦中添加PVC/NaCl的混烧试验,研究了Cl元素对燃烧中CO氧化的影响以及混烧时SO_x等污染物的生成规律.试验结果表明,半焦与PVC/NaCl混烧在技术上可行,添加Cl元素抑制CO氧化,加入PVC增加SO_2排放,而加入NaCl则减少SO_2排放.

卤素对于燃烧中CO氧化和NO生成的影响

卤族元素在燃烧时会影响CO的氧化和NO的生成。本文通过化学平衡方法分析了H2O和HCl对于CO氧化的协同效应,在非均匀布风流化床中进行了PVC塑料与煤或半焦的混烧试验,测量了CO和CH4的排放特性,从而从理论上和实验上证明了HCl对CO氧化的抑制作用。并给出了今后研究的方向和问题。

A Trabecular Bone Explant Model of Osteocyte–Osteoblast Co-Culture for Bone Mechanobiology

The osteocyte network is recognized as the major mechanical sensor in the bone remodeling process, and osteocyte-osteoblast communication acts as an important mediator in the coordination of bone formation and turnover. In this study, we developed a novel 3D trabecular bone explant co-culture model that allows live osteocytes situated in their native extracellular matrix environment to be interconnected with seeded osteoblasts on the bone surface. Using a low-level medium perfusion system, the viability of in situ osteocytes in bone explants was maintained for up to 4 weeks, and functional gap junction intercellular communication (GJIC) was successfully established between osteocytes and seeded primary osteoblasts. Using this novel co-culture model, the effects of dynamic deformational loading, GJIC, and prostaglandin E-2 (PGE(2)) release on functional bone adaptation were further investigated. The results showed that dynamical deformational loading can significantly increase the PGE(2) release by bone cells, bone formation, and the apparent elastic modulus of bone explants. However, the inhibition of gap junctions or the PGE(2) pathway dramatically attenuated the effects of mechanical loading. This 3D trabecular bone explant co-culture model has great potential to fill in the critical gap in knowledge regarding the role of osteocytes as a mechano-sensor and how osteocytes transmit signals to regulate osteoblasts function and skeletal integrity as reflected in its mechanical properties.